Datasets:

G-reen
/

fastdetector-test-stat-test

original string	prompt dict	topic string	format string	response_0 string	response_1 string	generator_model string	generation_params string	final_response string	jaccard_1 float64	jaccard_2 float64	levenshtein float64	softngram float64	cosdist float64	bertscore_precision float64	bertscore_recall float64	bertscore float64	moverscore float64	reranker float64	original_perplexity_llama_instruct float64	original_entropy_llama_instruct float64	original_topp_outlier_llama_instruct float64	original_topk_outlier_llama_instruct float64	original_fastdetectgpt_llama_instruct float64	original_perplexity_llama_base float64	original_entropy_llama_base float64	original_topp_outlier_llama_base float64	original_topk_outlier_llama_base float64	original_fastdetectgpt_llama_base float64	final_response_perplexity_llama_instruct float64	final_response_entropy_llama_instruct float64	final_response_topp_outlier_llama_instruct float64	final_response_topk_outlier_llama_instruct float64	final_response_fastdetectgpt_llama_instruct float64	final_response_perplexity_llama_base float64	final_response_entropy_llama_base float64	final_response_topp_outlier_llama_base float64	final_response_topk_outlier_llama_base float64	final_response_fastdetectgpt_llama_base float64	original_binoculars float64	final_response_binoculars float64	original_editlens_score_roberta_large float64	final_response_editlens_bucket_roberta_large int64	final_response_editlens_score_roberta_large float64	original_editlens_score_llama_3_2_3b float64	final_response_editlens_bucket_llama_3_2_3b int64	final_response_editlens_score_llama_3_2_3b float64	original_editlens_score_giga_roberta float64	final_response_editlens_bucket_giga_roberta int64	final_response_editlens_score_giga_roberta float64	original_editlens_score_giga_llama_3_2_3b float64	final_response_editlens_bucket_giga_llama_3_2_3b int64	final_response_editlens_score_giga_llama_3_2_3b float64	original_detector_score_qwen35_35b_a3b float64	final_response_detector_score_qwen35_35b_a3b float64	original_detector_score_llama_3b float64	final_response_detector_score_llama_3b float64	original_detector_score_gemma3_27b float64	final_response_detector_score_gemma3_27b float64	original_detector_score_qwen35_35b_a3b_soft float64	final_response_detector_score_qwen35_35b_a3b_soft float64	original_detector_score_llama_3b_soft float64	final_response_detector_score_llama_3b_soft float64	original_detector_score_llama_3b_reg float64	final_response_detector_score_llama_3b_reg float64	original_detector_score_model_family float64	final_response_detector_score_model_family float64
Welcome to the online home of the National Center for Arts Research, where data-driven insights help advance the arts in the U.S. It all began on February 13, 2012, when SMU’s Meadows School of the Arts and Cox School of Business hosted a public forum with the Cultural Data Project (CDP) and the national arts marketing...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nCreate a prompt that would allow an LLM to generate this text as accurately/verbatim as possible, without the prompt containing any of the actual text. That is, you must use descriptors to accurately describe the text to generate.\nDo not output anything besi...	Art & Design	About (Org.)	Write a polished seven-paragraph welcome statement for the newly launched online home of a U.S. national arts research center based at Southern Methodist University. Reconstruct the source wording as closely as possible. Include these details in order: 1. Open by welcoming readers and state that data-driven insights ...	Welcome to the newly launched online home of the National Center for Arts Research at Southern Methodist University. We believe data-driven insights will help advance the arts nationwide. Our initiative began on February 13, 2012, with a public forum jointly hosted by SMU’s Meadows School of the Arts and Cox School of...	gpt-5.6-sol	{"reasoning_effort": "none"}	Welcome to the newly launched online home of the National Center for Arts Research at Southern Methodist University. We believe data-driven insights will help advance the arts nationwide. Our initiative began on February 13, 2012, with a public forum jointly hosted by SMU’s Meadows School of the Arts and Cox School of ...	0.54023	0.779956	824	0.271637	0.028756	0.095189	0.089647	0.092427	0.43229	-7.75	12.583448	2.566656	0.050938	0.067024	0.309212	9.924471	2.389306	0.042895	0.053619	0.900116	14.498236	2.789064	0.052174	0.084058	0.977449	10.330088	2.470964	0.043478	0.063768	1.25073	0.796701	0.742512	0.014681	0	0.200267	0.005375	0	0.035778	0.003304	0	0.006373	0.001178	0	0.005134	0.039343	0.971992	0.065605	0.969556	0.033086	0.682785	0.012054	0.31237	0.058346	0.474509	0.011414	0.06543	0.000131	0.000231
“Doing what you love is a fight – an ongoing battle against every influence that wants you to conform. There is no solace in the conformity, only a blurred view of what you set out to be. Spend your life creating something. Betray The Institution.” This manifesto of mine adorns every package that ships out of Dark Coll...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nEdit this text by upgrading the verbs to be more dynamic and action-oriented, leaving the subjects, objects, and overall length completely unchanged.\nOutput the full new text with no extra statements or commentations.\nBegin directly with the text itself. Do...	Art & Design	About (Pers.)	“Pursuing what you love demands a fight – an ongoing battle against every influence that pressures you to conform. Conformity offers no solace, only a blurred view of what you set out to become. Devote your life to creating something. Defy The Institution.” This manifesto of mine graces every package that ships out of...	Every piece I create carries a story of independence, identity, and an unrelenting commitment to making art on my own terms. “Pursuing what you love demands a fight – an ongoing battle against every influence that pressures you to conform. Conformity offers no solace, only a blurred view of what you set out to become....	gpt-5.6-sol	{"reasoning_effort": "none"}	Every piece I create carries a story of independence, identity, and an unrelenting commitment to making art on my own terms. “Pursuing what you love demands a fight – an ongoing battle against every influence that pressures you to conform. Conformity offers no solace, only a blurred view of what you set out to become. ...	0.351429	0.466338	546	0.184441	0.026116	0.061355	0.042402	0.051973	0.313163	-9.125	15.248926	2.685937	0.048544	0.092233	-0.448498	12.712022	2.562055	0.042071	0.079288	0.22844	19.724715	3.052818	0.041074	0.126382	0.798662	17.344069	2.692009	0.047393	0.118483	-1.872013	0.861043	0.851551	0.010218	0	0.022724	0.008564	0	0.18764	0.003678	0	0.004162	0.001198	0	0.015595	0.034359	0.999979	0.023689	0.999938	0.042088	0.999989	0.008316	0.558809	0.022977	0.534127	0.005768	0.096191	0.000075	0.000116
ATO RIBEIRO: Home Coming November 3 - December 10, 2017 Anastasia Tinari Projects is thrilled to present Home Coming, Ato Ribeiro's first U.S. solo exhibition. With extreme precision and craftsmanship, the artist creates intricately detailed, complex patterned “wooden kente cloths” using humble pieces of pine, ebony, a...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nExpand this text by adding real-world applications and specific details.\nOutput the full new text with no extra statements or commentations.\nBegin directly with the text itself. Do not add a title, a heading, or a label naming what you have written." ], ...	Art & Design	About (Pers.)	ATO RIBEIRO: Home Coming November 3–December 10, 2017 Anastasia Tinari Projects is thrilled to present Home Coming, Ato Ribeiro’s first U.S. solo exhibition. With extraordinary precision and craftsmanship, Ribeiro creates intricately detailed, complexly patterned “wooden kente cloths” from humble pieces of pine, ebo...		gpt-5.6-sol	{"reasoning_effort": "none"}	ATO RIBEIRO: Home Coming November 3–December 10, 2017 Anastasia Tinari Projects is thrilled to present Home Coming, Ato Ribeiro’s first U.S. solo exhibition. With extraordinary precision and craftsmanship, Ribeiro creates intricately detailed, complexly patterned “wooden kente cloths” from humble pieces of pine, ebony,...	0.692246	0.794273	10,073	0.656403	0.056657	0.14538	0.071443	0.109944	0.567095	-8.3125	13.870806	2.800173	0.039451	0.114065	2.525136	10.529681	2.370141	0.037736	0.102058	0.251073	14.274308	2.661651	0.040691	0.113072	0.078524	11.017943	2.464129	0.028088	0.092906	1.647565	0.752953	0.804858	0.048313	1	0.333973	0.060326	1	0.405797	0.001567	0	0.016601	0.002011	0	0.003339	0.09739	0.999942	0.026964	0.99995	0.026355	0.99991	0.010987	0.634317	0.035679	0.565539	0.003159	0.213867	0.000245	0.999974
Philip Jay O’Meilia was born in 1927 in Tulsa, Oklahoma, where he resides today. Between military stints during WWII and Korea (serving as a Navy artist), he attended the Art Students League in New York (1945-46, 1947-49) and the Chicago Academy of Fine Art (1946-47). He also attended the Cape School of Art in Province...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nWrite a new piece of text that uses this text as a citation or reference.\nOutput the full new text with no extra statements or commentations.\nBegin directly with the text itself. Do not add a title, a heading, or a label naming what you have written.\nThe f...	Art & Design	About (Pers.)	Philip Jay O’Meilia’s career illustrates how artistic versatility can unite fine art, public design, and popular culture. Born in Tulsa, Oklahoma, in 1927, O’Meilia developed his skills through study at institutions including the Art Students League in New York, the Chicago Academy of Fine Art, and the Cape School of A...		gpt-5.6-sol	{"reasoning_effort": "none"}	Philip Jay O’Meilia’s career illustrates how artistic versatility can unite fine art, public design, and popular culture. Born in Tulsa, Oklahoma, in 1927, O’Meilia developed his skills through study at institutions including the Art Students League in New York, the Chicago Academy of Fine Art, and the Cape School of A...	0.748503	0.884112	1,603	0.601151	0.118197	0.130861	0.135159	0.133015	0.556739	-5.0625	7.833371	2.060283	0.05423	0.05423	0.0201	6.296499	1.961206	0.036876	0.05423	1.361784	12.064904	2.43965	0.036408	0.09466	-0.4929	10.380865	2.38512	0.021845	0.097087	0.447314	0.794076	0.845316	0.116582	3	0.996982	0.058427	3	0.815417	0.004645	3	0.968174	0.011922	3	0.916431	0.066569	0.999995	0.041153	0.999999	0.044019	1	0.022977	0.717012	0.039048	0.619241	0.015015	0.353516	0.000085	0.999981
I’m a Hobby/Amateur Photographer, and started taking photo’s when I was around 11 years old and a member of the school Camera Club. Due to a major change in lifestyle and responsibilities I was forced to sideline my hobby. Now things have changed, I am taking the time to reignite my passion as I find photography very r...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nExpand this text by adding backstory and context to enrich understanding.\nOutput the full new text with no extra statements or commentations.\nBegin directly with the text itself. Do not add a title, a heading, or a label naming what you have written." ], ...	Art & Design	About (Pers.)	I’m a hobby and amateur photographer, and my interest in photography began when I was around 11 years old. At the time, I joined my school Camera Club, where I first experienced the excitement of capturing a moment and turning an everyday scene into something worth remembering. Those early years gave me an appreciation...		gpt-5.6-sol	{"reasoning_effort": "none"}	I’m a hobby and amateur photographer, and my interest in photography began when I was around 11 years old. At the time, I joined my school Camera Club, where I first experienced the excitement of capturing a moment and turning an everyday scene into something worth remembering. Those early years gave me an appreciation...	0.788177	0.917424	5,272	0.855498	0.136121	0.187424	0.107837	0.149488	0.625453	-5.25	21.514899	2.599877	0.079412	0.094118	-3.925525	15.482302	2.680543	0.058824	0.064706	-0.503429	7.619148	1.842843	0.053435	0.031298	-3.922254	6.372927	1.918736	0.024427	0.026718	1.384051	0.929534	0.880026	0.011468	1	0.377059	0.038945	1	0.406773	0.026761	2	0.580025	0.025231	1	0.433484	0.105948	0.99999	0.043691	0.999994	0.037047	0.999999	0.046725	0.68342	0.044681	0.610453	0.007935	0.283203	0.000444	0.999997
My name is Kanika Marshall and my goal is to share the strength and majesty of ancient cultures via beautiful, spiritual sculptures and clay pottery that I hand-sculpt in my private art studio near Sacramento, California. I am most inspired by artwork that uses the interplay of various media to show the connection betw...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nWrite a detailed rebuttal or critique challenging the core ideas, themes, or facts presented in this text, using the same language and roughly matching the intellectual level of the text.\nOutput the full new text with no extra statements or commentations.\nB...	Art & Design	About (Pers.)	My name is Kanika Marshall, and while my sculptures and pottery are deeply personal, I recognize that presenting them as expressions of ancient cultures raises difficult questions about identity, history, and artistic authority. Beauty and spiritual feeling alone do not necessarily communicate the strength or majesty o...		gpt-5.6-sol	{"reasoning_effort": "none"}	My name is Kanika Marshall, and while my sculptures and pottery are deeply personal, I recognize that presenting them as expressions of ancient cultures raises difficult questions about identity, history, and artistic authority. Beauty and spiritual feeling alone do not necessarily communicate the strength or majesty o...	0.829233	0.9348	5,039	0.913208	0.21951	0.193199	0.148277	0.171347	0.649664	-4.8125	22.833186	2.962651	0.057569	0.138593	-1.578282	17.223014	2.825121	0.03838	0.100213	-0.206743	18.629657	2.689601	0.057803	0.113955	-3.870424	14.873047	2.716568	0.033031	0.100743	0.281721	0.873084	0.905527	0.016615	3	0.990635	0.046057	3	0.792906	0.004805	2	0.711718	0.006278	1	0.609294	0.101589	0.999999	0.064653	0.999997	0.087564	1	0.027169	0.720958	0.058346	0.64512	0.021484	0.445313	0.000169	1
Hello and welcome to my website. My name's Lee Bofkin and I'm passionate about many things. On this page you'll find some clips of me bboying (breakdancing), interviewing bboys and running around London in a gorilla suit. I also photograph graffiti, write for Sense Magazine in LA and produce original designs and artwor...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nImprove this text by rewriting it. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose.\nOutput the full new text with no extra statements or commentations.\nBegin directly with the t...	Art & Design	About (Pers.)	Hello and welcome to my website. My name's Lee Bofkin and I'm passionate about many things. Here you can explore clips of me bboying (breakdancing), interviewing fellow bboys, and running around London in a gorilla suit. I also photograph graffiti, create original artwork and designs, and write for Sense Magazine in LA...	Hello and welcome to my art and design website. My name's Lee Bofkin and I'm passionate about many things. Here you can explore clips of me bboying (breakdancing), interviewing fellow bboys, and running around London in a gorilla suit. I also photograph graffiti, create original artwork, illustrations, and graphic desi...	gpt-5.6-sol	{"reasoning_effort": "none"}	Hello and welcome to my art and design website. My name's Lee Bofkin and I'm passionate about many things. Here you can explore clips of me bboying (breakdancing), interviewing fellow bboys, and running around London in a gorilla suit. I also photograph graffiti, create original artwork, illustrations, and graphic desi...	0.459259	0.592326	738	0.500893	0.112676	0.080704	0.060946	0.07093	0.375629	-6.6875	19.75513	2.776061	0.068421	0.115789	-1.784081	14.874791	2.665096	0.044737	0.105263	-0.30562	19.207579	2.917459	0.056075	0.121495	-0.341515	15.378051	2.660919	0.03972	0.11215	-0.680146	0.865098	0.843831	0.030599	0	0.132924	0.042537	0	0.112466	0.020576	0	0.074182	0.010713	0	0.037751	0.720958	0.999981	0.072637	0.815233	0.335351	0.998273	0.484868	0.61555	0.067547	0.395601	0.006989	0.139648	0.000368	0.002551
New research by WIPO shows that a wide range of factors, and not income alone, motivates creators to engage in creative activity and explores what this means for copyright policy and public funding of the arts. Legal scholars worry about the increasing disconnect between the formal exercise of rights conferred by copyr...	{ "chat_turns": [ "<document>\n{{DOC}}\n</document>\n\nTranslate the given text to French.\nDo not output anything besides what you were requested to write, and do not output any extra commentary.", "{{RESP_0}}\nTranslate this text to English. Output only the English translation with no extra commentary.\nBeg...	Art & Design	Academic Writing	De nouvelles recherches menées par l’OMPI montrent qu’un large éventail de facteurs, et non le seul revenu, incitent les créateurs à se livrer à une activité créative, et examinent ce que cela implique pour la politique du droit d’auteur et le financement public des arts. Les juristes s’inquiètent du décalage croissan...	New research conducted by WIPO shows that a wide range of factors—not income alone—motivate creators to engage in creative activity, and examines what this means for copyright policy and public funding of the arts. Legal scholars are concerned about the growing disconnect between the formal exercise of rights granted ...	gpt-5.6-sol	{"reasoning_effort": "none"}	New research conducted by WIPO shows that a wide range of factors—not income alone—motivate creators to engage in creative activity, and examines what this means for copyright policy and public funding of the arts. Legal scholars are concerned about the growing disconnect between the formal exercise of rights granted b...	0.499405	0.702902	2,888	0.187752	0.008984	0.082811	0.080013	0.081414	0.401011	-8.4375	15.575469	2.77249	0.042668	0.103966	0.496506	12.484473	2.519773	0.041466	0.079327	-0.094079	13.147042	2.427918	0.052858	0.081131	-2.87393	10.373319	2.359669	0.038107	0.067609	0.421861	0.838169	0.879506	0.039043	1	0.367545	0.059691	1	0.345466	0.003731	0	0.006362	0.001983	0	0.031909	0.028871	0.574335	0.021287	0.994698	0.024054	0.851459	0.008062	0.100879	0.028436	0.485295	0.00415	0.126953	0.000431	0.002473
"Today we share with you an interview with Scott Riley a user experience designer who has a focus on(...TRUNCATED)	{"chat_turns":["<document>\n{{DOC}}\n</document>\n\nWrite a 'few-shot' prompt using minimal, fragmen(...TRUNCATED)	Art & Design	Audio Transcript	"Continue the interview article from the sparse examples below. Match its informal, candid tone, fir(...TRUNCATED)	"Scott Riley is a user experience designer with a particular fondness for interaction design, though(...TRUNCATED)	gpt-5.6-sol	{"reasoning_effort": "none"}	"Scott Riley is a user experience designer with a particular fondness for interaction design, though(...TRUNCATED)	0.798921	0.934211	8,000	0.923588	0.078816	0.173281	0.148423	0.161036	0.616532	-7.25	17.236032	2.835266	0.038228	0.109223	-0.212346	13.120511	2.564277	0.041262	0.087985	-0.190752	17.033708	2.683017	0.057182	0.114822	-3.267473	13.414453	2.594954	0.039799	0.095608	-0.030293	0.813381	0.854064	0.0135	1	0.261806	0.04149	0	0.433327	0.00395	0	0.016069	0.002582	0	0.034501	0.023871	0.999987	0.021781	0.999918	0.035411	0.999993	0.008062	0.712232	0.028871	0.656658	0.00589	0.425781	0.000123	0.999999
"Godfather - you are brilliant! I put the original foot back on and it had slightly shorter screws a(...TRUNCATED)	{"chat_turns":["<document>\n{{DOC}}\n</document>\n\nWrite a piece of text in the same style and lang(...TRUNCATED)	Art & Design	Comment Section	"Godfather - you are impossible! I swapped the original frame back in and the shorter clips made no (...TRUNCATED)		gpt-5.6-sol	{"reasoning_effort": "none"}	"Godfather - you are impossible! I swapped the original frame back in and the shorter clips made no (...TRUNCATED)	0.70696	0.835391	888	0.933019	0.455946	0.133816	0.153615	0.14383	0.554169	-4.375	32.651003	3.357692	0.044041	0.145078	-0.984921	21.471686	3.019011	0.03886	0.121762	-0.403733	41.428235	3.304927	0.065831	0.172414	-3.123976	33.335624	3.20077	0.065831	0.169279	-2.372993	0.789944	0.951208	0.013444	0	0.0347	0.014646	0	0.237893	0.006358	0	0.019248	0.004645	0	0.089704	0.021781	0.999993	0.015664	0.99998	0.018125	1	0.006488	0.799442	0.018833	0.727201	0.005249	0.5625	0.000158	0.002051

End of preview. Expand in Data Studio

Auto-generated FastDetector dataset

Fastdetector Test Stat

Best detector

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base)

0.8371 TPR @ 1% FPR

Hardest prompt subset

rewrite

0.5950 max detector TPR @ 1% FPR

Hardest generator config

claude-opus-5 (Temp: Unknown)

0.7594 max detector TPR @ 1% FPR

17,291

rows

generator configs

prompt subsets

detectors

01Leaderboard 02Distances 03Appendix ↗Model analytics

Detector leaderboard

Score-based detectors ranked by overall AUROC. Thresholds are placed exactly on every human score, so TPR is reported at a true 1% and 0.1% false positive rate. Open a row for its per-subset table, threshold sweep, its TPR as AI rows with low cosdist are dropped or relabelled human, and its score distributions.

01Trained detector: Gemma-3-27B (QLoRA, hard labels)N34,582AUROC0.9615TPR @ 1% FPR0.8270TPR @ 0.1% FPR0.7761⌄

1% FPR threshold 0.68970.1% FPR threshold 0.9504

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9615	0.8270	0.7761
prompt ● direct_reference	8,874	0.9894	0.9355	0.9035
prompt ● indirect_reference	8,798	0.9593	0.7904	0.7213
prompt ● revise ▲ best	8,960	0.9957	0.9614	0.9357
prompt ● rewrite ▼ worst	7,950	0.8938	0.5950	0.5145
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9770	0.8859	0.8519
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9350	0.6997	0.6269
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9529	0.7487	0.6729
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9625	0.8626	0.8193
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9752	0.9023	0.8724
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9421	0.7691	0.6945
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9635	0.8080	0.7300
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9653	0.8437	0.8071
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9613	0.8372	0.7948
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9675	0.8505	0.7989
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9739	0.8907	0.8706

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Gemma-3-27B (QLoRA, hard labels)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Gemma-3-27B (QLoRA, hard labels): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Gemma-3-27B (QLoRA, hard labels): performance over different AI class thresholds

Performance over different AI class thresholds · AI rows with cosdist below the cutoff are relabelled human, the 1% FPR threshold is refitted on that larger human pool, and TPR is read on the rows still labelled AI · AUC is the mean TPR over the cutoff rangeOpen full size ↗

Score distributions

Trained detector: Gemma-3-27B (QLoRA, hard labels): overall human

Overall · human onlyOpen full size ↗

Trained detector: Gemma-3-27B (QLoRA, hard labels): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Gemma-3-27B (QLoRA, hard labels): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Gemma-3-27B (QLoRA, hard labels): generator subsets

AI texts · one envelope per generator configOpen full size ↗

02Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base)N34,582AUROC0.9599TPR @ 1% FPR0.8371TPR @ 0.1% FPR0.6662⌄

1% FPR threshold 0.84290.1% FPR threshold 0.9997

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9599	0.8371	0.6662
prompt ● direct_reference	8,874	0.9877	0.9457	0.8616
prompt ● indirect_reference	8,798	0.9611	0.8170	0.5613
prompt ● revise ▲ best	8,960	0.9965	0.9748	0.8721
prompt ● rewrite ▼ worst	7,950	0.8858	0.5831	0.3323
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9773	0.8927	0.7471
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9466	0.7594	0.5188
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9607	0.7920	0.5595
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9569	0.8533	0.7129
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9696	0.8892	0.7642
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9387	0.7867	0.5772
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9598	0.8161	0.5599
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9594	0.8375	0.7033
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9596	0.8451	0.6929
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9606	0.8479	0.6991
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9687	0.8882	0.7984

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): performance over different AI class thresholds

Score distributions

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): overall human

Overall · human onlyOpen full size ↗

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base): generator subsets

AI texts · one envelope per generator configOpen full size ↗

03Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai)N34,582AUROC0.9590TPR @ 1% FPR0.8325TPR @ 0.1% FPR0.5812⌄

1% FPR threshold 0.42200.1% FPR threshold 0.6053

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9590	0.8325	0.5812
prompt ● direct_reference	8,874	0.9884	0.9479	0.9024
prompt ● indirect_reference	8,798	0.9609	0.8138	0.5188
prompt ● revise ▲ best	8,960	0.9958	0.9723	0.7078
prompt ● rewrite ▼ worst	7,950	0.8826	0.5665	0.1492
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9774	0.8964	0.6379
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9431	0.7399	0.4447
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9567	0.7751	0.4787
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9552	0.8490	0.6368
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9694	0.8880	0.6845
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9392	0.7867	0.5276
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9599	0.8061	0.4931
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9579	0.8375	0.6321
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9601	0.8432	0.5983
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9623	0.8472	0.6005
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9675	0.8882	0.6639

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): performance over different AI class thresholds

Score distributions

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): overall human

Overall · human onlyOpen full size ↗

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai): generator subsets

AI texts · one envelope per generator configOpen full size ↗

04Trained detector: Llama-3.2-3B soft labels (soft_ai)N34,582AUROC0.9523TPR @ 1% FPR0.7794TPR @ 0.1% FPR0.6840⌄

1% FPR threshold 0.37110.1% FPR threshold 0.5226

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9523	0.7794	0.6840
prompt ● direct_reference	8,874	0.9852	0.9103	0.8772
prompt ● indirect_reference	8,798	0.9493	0.7340	0.6176
prompt ● revise ▲ best	8,960	0.9937	0.9368	0.8692
prompt ● rewrite ▼ worst	7,950	0.8716	0.5059	0.3331
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9711	0.8538	0.7736
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9291	0.6394	0.5044
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9431	0.6748	0.5576
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9517	0.8212	0.7401
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9687	0.8662	0.7897
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9267	0.6989	0.5803
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9559	0.7456	0.6278
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9563	0.8119	0.7241
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9502	0.7935	0.7095
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9559	0.8062	0.7063
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9660	0.8643	0.8141

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Llama-3.2-3B soft labels (soft_ai)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Llama-3.2-3B soft labels (soft_ai): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Llama-3.2-3B soft labels (soft_ai): performance over different AI class thresholds

Score distributions

Trained detector: Llama-3.2-3B soft labels (soft_ai): overall human

Overall · human onlyOpen full size ↗

Trained detector: Llama-3.2-3B soft labels (soft_ai): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Llama-3.2-3B soft labels (soft_ai): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Llama-3.2-3B soft labels (soft_ai): generator subsets

AI texts · one envelope per generator configOpen full size ↗

05Trained detector: Llama-3.2-3B (QLoRA)N34,582AUROC0.9514TPR @ 1% FPR0.7858TPR @ 0.1% FPR0.5778⌄

1% FPR threshold 0.72250.1% FPR threshold 0.9996

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9514	0.7858	0.5778
prompt ● direct_reference	8,874	0.9816	0.8961	0.7361
prompt ● indirect_reference	8,798	0.9494	0.7429	0.4749
prompt ● revise ▲ best	8,960	0.9935	0.9424	0.7614
prompt ● rewrite ▼ worst	7,950	0.8717	0.5336	0.3082
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9737	0.8593	0.7082
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9262	0.6489	0.3587
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9423	0.6861	0.4029
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9511	0.8267	0.6429
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9673	0.8743	0.7274
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9241	0.7051	0.4435
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9549	0.7525	0.4445
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9550	0.8154	0.6425
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9494	0.7995	0.6023
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9564	0.8122	0.6171
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9645	0.8662	0.7720

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Llama-3.2-3B (QLoRA)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Llama-3.2-3B (QLoRA): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Llama-3.2-3B (QLoRA): performance over different AI class thresholds

Score distributions

Trained detector: Llama-3.2-3B (QLoRA): overall human

Overall · human onlyOpen full size ↗

Trained detector: Llama-3.2-3B (QLoRA): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Llama-3.2-3B (QLoRA): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Llama-3.2-3B (QLoRA): generator subsets

AI texts · one envelope per generator configOpen full size ↗

06Trained detector: Llama-3.2-3B regression (reg_sqrt)N34,582AUROC0.9408TPR @ 1% FPR0.7637TPR @ 0.1% FPR0.6345⌄

1% FPR threshold 0.09030.1% FPR threshold 0.1729

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9408	0.7637	0.6345
prompt ● direct_reference ▲ best	8,874	0.9883	0.9367	0.9078
prompt ● indirect_reference	8,798	0.9433	0.7343	0.5592
prompt ● revise	8,960	0.9914	0.9355	0.8185
prompt ● rewrite ▼ worst	7,950	0.8283	0.4098	0.2055
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9611	0.8285	0.7045
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.9064	0.6256	0.4893
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.9278	0.6830	0.5238
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9455	0.8032	0.6844
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9604	0.8457	0.7343
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.9105	0.6989	0.5721
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9448	0.7394	0.5773
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9480	0.7773	0.6694
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9421	0.7869	0.6539
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9476	0.7698	0.6468
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9553	0.8436	0.7268

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Trained detector: Llama-3.2-3B regression (reg_sqrt)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Trained detector: Llama-3.2-3B regression (reg_sqrt): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Trained detector: Llama-3.2-3B regression (reg_sqrt): performance over different AI class thresholds

Score distributions

Trained detector: Llama-3.2-3B regression (reg_sqrt): overall human

Overall · human onlyOpen full size ↗

Trained detector: Llama-3.2-3B regression (reg_sqrt): overall AI

Overall · AI onlyOpen full size ↗

Trained detector: Llama-3.2-3B regression (reg_sqrt): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Trained detector: Llama-3.2-3B regression (reg_sqrt): generator subsets

AI texts · one envelope per generator configOpen full size ↗

07Giga EditLens Llama-3.2-3B ScoreN34,582AUROC0.9004TPR @ 1% FPR0.6214TPR @ 0.1% FPR0.4765⌄

1% FPR threshold 0.04330.1% FPR threshold 0.1831

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.9004	0.6214	0.4765
prompt ● direct_reference	8,874	0.9503	0.7805	0.6689
prompt ● indirect_reference	8,798	0.9007	0.5699	0.4055
prompt ● revise ▲ best	8,960	0.9789	0.8583	0.6866
prompt ● rewrite ▼ worst	7,950	0.7551	0.2340	0.1034
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.9279	0.7138	0.5762
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.8350	0.4089	0.2531
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.8648	0.4787	0.2945
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.9101	0.6498	0.5111
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9368	0.7380	0.6167
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.8667	0.5226	0.3601
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.9094	0.5829	0.3741
gen ● grok-4.3 (Temp: Unknown)	2,892	0.9045	0.6687	0.5622
gen ● kimi-k3 (Temp: Unknown)	3,022	0.9139	0.6843	0.5321
gen ● minimax-m3 (Temp: Unknown)	3,024	0.9123	0.6653	0.5443
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9232	0.7299	0.6288

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Giga EditLens Llama-3.2-3B Score

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Giga EditLens Llama-3.2-3B Score: TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Giga EditLens Llama-3.2-3B Score: performance over different AI class thresholds

Score distributions

Giga EditLens Llama-3.2-3B Score: overall human

Overall · human onlyOpen full size ↗

Giga EditLens Llama-3.2-3B Score: overall AI

Overall · AI onlyOpen full size ↗

Giga EditLens Llama-3.2-3B Score: prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Giga EditLens Llama-3.2-3B Score: generator subsets

AI texts · one envelope per generator configOpen full size ↗

08EditLens Llama-3.2-3B ScoreN34,582AUROC0.8744TPR @ 1% FPR0.5535TPR @ 0.1% FPR0.3795⌄

1% FPR threshold 0.20160.1% FPR threshold 0.3730

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.8744	0.5535	0.3795
prompt ● direct_reference	8,874	0.9272	0.6969	0.5486
prompt ● indirect_reference	8,798	0.8687	0.5067	0.3217
prompt ● revise ▲ best	8,960	0.9714	0.8065	0.5527
prompt ● rewrite ▼ worst	7,950	0.7134	0.1600	0.0596
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.8949	0.6373	0.4800
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.7919	0.3116	0.1288
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.8141	0.3810	0.2011
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.8897	0.5922	0.4270
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9171	0.6914	0.5302
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.8541	0.4699	0.2905
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.8945	0.5237	0.2893
gen ● grok-4.3 (Temp: Unknown)	2,892	0.8885	0.6113	0.4689
gen ● kimi-k3 (Temp: Unknown)	3,022	0.8870	0.6016	0.4077
gen ● minimax-m3 (Temp: Unknown)	3,024	0.8810	0.5985	0.4325
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9060	0.6771	0.5283

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: EditLens Llama-3.2-3B Score

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

EditLens Llama-3.2-3B Score: TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

EditLens Llama-3.2-3B Score: performance over different AI class thresholds

Score distributions

EditLens Llama-3.2-3B Score: overall human

Overall · human onlyOpen full size ↗

Overall · AI onlyOpen full size ↗

EditLens Llama-3.2-3B Score: prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

EditLens Llama-3.2-3B Score: generator subsets

AI texts · one envelope per generator configOpen full size ↗

09Giga EditLens Roberta ScoreN34,582AUROC0.8627TPR @ 1% FPR0.5370TPR @ 0.1% FPR0.4695⌄

1% FPR threshold 0.12390.1% FPR threshold 0.2631

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.8627	0.5370	0.4695
prompt ● direct_reference	8,874	0.9195	0.6716	0.6189
prompt ● indirect_reference	8,798	0.8561	0.4869	0.4130
prompt ● revise ▲ best	8,960	0.9663	0.7913	0.7058
prompt ● rewrite ▼ worst	7,950	0.6880	0.1555	0.0989
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.8997	0.6459	0.5743
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.7755	0.2783	0.2117
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.8136	0.3665	0.2957
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.8784	0.5984	0.5439
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.9104	0.6901	0.6248
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.8170	0.4053	0.3325
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.8663	0.4507	0.3473
gen ● grok-4.3 (Temp: Unknown)	2,892	0.8632	0.6024	0.5505
gen ● kimi-k3 (Temp: Unknown)	3,022	0.8852	0.6142	0.5367
gen ● minimax-m3 (Temp: Unknown)	3,024	0.8788	0.5880	0.5397
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.9016	0.6765	0.6187

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Giga EditLens Roberta Score

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Giga EditLens Roberta Score: TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Giga EditLens Roberta Score: performance over different AI class thresholds

Score distributions

Giga EditLens Roberta Score: overall human

Overall · human onlyOpen full size ↗

Overall · AI onlyOpen full size ↗

Giga EditLens Roberta Score: prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Giga EditLens Roberta Score: generator subsets

AI texts · one envelope per generator configOpen full size ↗

10Model-family detector: AI probability (Gemma-3-4B-it LoRA)N34,582AUROC0.8512TPR @ 1% FPR0.5864TPR @ 0.1% FPR0.5288⌄

1% FPR threshold 0.00170.1% FPR threshold 0.0081

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.8512	0.5864	0.5288
prompt ● direct_reference	8,874	0.9320	0.7785	0.7350
prompt ● indirect_reference	8,798	0.8406	0.5308	0.4669
prompt ● revise ▲ best	8,960	0.9532	0.8116	0.7364
prompt ● rewrite ▼ worst	7,950	0.6620	0.1796	0.1331
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.8699	0.6527	0.6114
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.7835	0.4133	0.3279
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.7889	0.4173	0.3114
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.8543	0.5996	0.5594
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.8909	0.7013	0.6571
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.8321	0.5314	0.4655
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.8605	0.5673	0.4913
gen ● grok-4.3 (Temp: Unknown)	2,892	0.8688	0.6252	0.5775
gen ● kimi-k3 (Temp: Unknown)	3,022	0.8647	0.6241	0.5711
gen ● minimax-m3 (Temp: Unknown)	3,024	0.8596	0.6144	0.5741
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.8925	0.7085	0.6765

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Model-family detector: AI probability (Gemma-3-4B-it LoRA)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Model-family detector: AI probability (Gemma-3-4B-it LoRA): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Model-family detector: AI probability (Gemma-3-4B-it LoRA): performance over different AI class thresholds

Score distributions

Model-family detector: AI probability (Gemma-3-4B-it LoRA): overall human

Overall · human onlyOpen full size ↗

Model-family detector: AI probability (Gemma-3-4B-it LoRA): overall AI

Overall · AI onlyOpen full size ↗

Model-family detector: AI probability (Gemma-3-4B-it LoRA): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Model-family detector: AI probability (Gemma-3-4B-it LoRA): generator subsets

AI texts · one envelope per generator configOpen full size ↗

11EditLens Roberta-Large ScoreN34,582AUROC0.8478TPR @ 1% FPR0.4169TPR @ 0.1% FPR0.1705⌄

1% FPR threshold 0.30800.1% FPR threshold 0.7948

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.8478	0.4169	0.1705
prompt ● direct_reference	8,874	0.8821	0.5229	0.2991
prompt ● indirect_reference	8,798	0.8667	0.3858	0.1600
prompt ● revise ▲ best	8,960	0.9588	0.6411	0.1886
prompt ● rewrite ▼ worst	7,950	0.6658	0.0803	0.0181
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.8846	0.5262	0.2634
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.7889	0.1602	0.0352
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.7773	0.1591	0.0451
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.8566	0.5093	0.1813
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.8966	0.6285	0.2794
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.8108	0.2666	0.0816
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.8532	0.3286	0.0935
gen ● grok-4.3 (Temp: Unknown)	2,892	0.8536	0.5062	0.2434
gen ● kimi-k3 (Temp: Unknown)	3,022	0.8724	0.4911	0.1800
gen ● minimax-m3 (Temp: Unknown)	3,024	0.8540	0.4392	0.1786
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.8794	0.5798	0.2996

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: EditLens Roberta-Large Score

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

EditLens Roberta-Large Score: TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

EditLens Roberta-Large Score: performance over different AI class thresholds

Score distributions

EditLens Roberta-Large Score: overall human

Overall · human onlyOpen full size ↗

EditLens Roberta-Large Score: overall AI

Overall · AI onlyOpen full size ↗

EditLens Roberta-Large Score: prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

EditLens Roberta-Large Score: generator subsets

AI texts · one envelope per generator configOpen full size ↗

12BinocularsN34,582AUROC0.5993TPR @ 1% FPR0.0623TPR @ 0.1% FPR0.0166⌄

1% FPR threshold 0.95030.1% FPR threshold 0.9907

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5993	0.0623	0.0166
prompt ● direct_reference	8,874	0.6595	0.0868	0.0207
prompt ● indirect_reference	8,798	0.5906	0.0559	0.0175
prompt ● revise	8,960	0.5990	0.0656	0.0147
prompt ● rewrite	7,950	0.5417	0.0385	0.0131
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.6465	0.0734	0.0185
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.5455	0.0251	0.0031
gen ● claude-opus-5-5 (Temp: Unknown) ▼ worst	3,192	0.5295	0.0169	0.0013
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.6168	0.0736	0.0223
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.6346	0.1133	0.0367
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5932	0.0295	0.0050
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.6401	0.0574	0.0125
gen ● grok-4.3 (Temp: Unknown)	2,892	0.6292	0.0920	0.0249
gen ● kimi-k3 (Temp: Unknown)	3,022	0.5536	0.0377	0.0099
gen ● minimax-m3 (Temp: Unknown)	3,024	0.5330	0.0337	0.0079
gen ● qwen3.7-max (Temp: Unknown) ▲ best	3,184	0.6658	0.1325	0.0402

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Binoculars: performance over different AI class thresholds

Score distributions

Overall · human onlyOpen full size ↗

Overall · AI onlyOpen full size ↗

AI texts · one envelope per prompt subsetOpen full size ↗

AI texts · one envelope per generator configOpen full size ↗

13Top-p Outliers (Llama-3.2-3B)N34,582AUROC0.5804TPR @ 1% FPR0.0525TPR @ 0.1% FPR0.0020⌄

1% FPR threshold 0.02190.1% FPR threshold 0.0071

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5804	0.0525	0.0020
prompt ● direct_reference	8,874	0.6362	0.0913	0.0029
prompt ● indirect_reference	8,798	0.6218	0.0707	0.0036
prompt ● revise	8,960	0.5840	0.0310	0.0007
prompt ● rewrite	7,950	0.4692	0.0133	0.0008
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.5806	0.0537	0.0019
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.5652	0.0276	0.0000
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.6852	0.0746	0.0000
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.5144	0.0210	0.0019
gen ● gemini-3.8-flash (Temp: Unknown) ▼ worst	3,214	0.4656	0.0118	0.0012
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.6271	0.0445	0.0006
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.6033	0.0499	0.0037
gen ● grok-4.3 (Temp: Unknown)	2,892	0.5683	0.0733	0.0041
gen ● kimi-k3 (Temp: Unknown)	3,022	0.6175	0.0847	0.0040
gen ● minimax-m3 (Temp: Unknown) ▲ best	3,024	0.6603	0.1177	0.0040
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.5025	0.0264	0.0013

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Top-p Outliers (Llama-3.2-3B)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Top-p Outliers (Llama-3.2-3B): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Top-p Outliers (Llama-3.2-3B): performance over different AI class thresholds

Score distributions

Top-p Outliers (Llama-3.2-3B): overall human

Overall · human onlyOpen full size ↗

Top-p Outliers (Llama-3.2-3B): overall AI

Overall · AI onlyOpen full size ↗

Top-p Outliers (Llama-3.2-3B): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Top-p Outliers (Llama-3.2-3B): generator subsets

AI texts · one envelope per generator configOpen full size ↗

14Entropy (Llama-3.2-3B-Instruct)N34,582AUROC0.5428TPR @ 1% FPR0.0189TPR @ 0.1% FPR0.0040⌄

1% FPR threshold 1.46500.1% FPR threshold 1.1720

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5428	0.0189	0.0040
prompt ● direct_reference	8,874	0.5364	0.0255	0.0052
prompt ● indirect_reference	8,798	0.6080	0.0309	0.0075
prompt ● revise	8,960	0.5563	0.0116	0.0020
prompt ● rewrite ▼ worst	7,950	0.4631	0.0063	0.0010
gen ● claude-opus-4-5-20251101 (Temp: Unknown) ▲ best	3,242	0.6176	0.0382	0.0074
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.5200	0.0113	0.0013
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.6242	0.0201	0.0025
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.4644	0.0087	0.0025
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.4456	0.0081	0.0006
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5742	0.0163	0.0031
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5951	0.0181	0.0044
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4754	0.0083	0.0048
gen ● kimi-k3 (Temp: Unknown)	3,022	0.5661	0.0351	0.0099
gen ● minimax-m3 (Temp: Unknown)	3,024	0.6048	0.0337	0.0066
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4813	0.0101	0.0013

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Entropy (Llama-3.2-3B-Instruct)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Entropy (Llama-3.2-3B-Instruct): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Entropy (Llama-3.2-3B-Instruct): performance over different AI class thresholds

Score distributions

Entropy (Llama-3.2-3B-Instruct): overall human

Overall · human onlyOpen full size ↗

Entropy (Llama-3.2-3B-Instruct): overall AI

Overall · AI onlyOpen full size ↗

Entropy (Llama-3.2-3B-Instruct): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Entropy (Llama-3.2-3B-Instruct): generator subsets

AI texts · one envelope per generator configOpen full size ↗

15Perplexity (Llama-3.2-3B-Instruct)N34,582AUROC0.5334TPR @ 1% FPR0.0190TPR @ 0.1% FPR0.0038⌄

1% FPR threshold 5.25180.1% FPR threshold 3.6443

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5334	0.0190	0.0038
prompt ● direct_reference	8,874	0.5175	0.0221	0.0052
prompt ● indirect_reference	8,798	0.5985	0.0334	0.0073
prompt ● revise	8,960	0.5498	0.0125	0.0013
prompt ● rewrite ▼ worst	7,950	0.4611	0.0068	0.0010
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.6076	0.0315	0.0056
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.5124	0.0144	0.0006
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.6311	0.0188	0.0038
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.4607	0.0087	0.0025
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.4227	0.0087	0.0000
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5723	0.0176	0.0013
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5591	0.0143	0.0037
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4458	0.0097	0.0041
gen ● kimi-k3 (Temp: Unknown) ▲ best	3,022	0.5755	0.0404	0.0099
gen ● minimax-m3 (Temp: Unknown)	3,024	0.6251	0.0351	0.0093
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4538	0.0107	0.0013

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Perplexity (Llama-3.2-3B-Instruct)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Perplexity (Llama-3.2-3B-Instruct): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Perplexity (Llama-3.2-3B-Instruct): performance over different AI class thresholds

Score distributions

Perplexity (Llama-3.2-3B-Instruct): overall human

Overall · human onlyOpen full size ↗

Perplexity (Llama-3.2-3B-Instruct): overall AI

Overall · AI onlyOpen full size ↗

Perplexity (Llama-3.2-3B-Instruct): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Perplexity (Llama-3.2-3B-Instruct): generator subsets

AI texts · one envelope per generator configOpen full size ↗

16Top-k Outliers (Llama-3.2-3B-Instruct)N34,582AUROC0.5208TPR @ 1% FPR0.0249TPR @ 0.1% FPR0.0034⌄

1% FPR threshold 0.03120.1% FPR threshold 0.0161

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5208	0.0249	0.0034
prompt ● direct_reference	8,874	0.5219	0.0433	0.0068
prompt ● indirect_reference	8,798	0.5807	0.0361	0.0048
prompt ● revise	8,960	0.5208	0.0127	0.0007
prompt ● rewrite ▼ worst	7,950	0.4542	0.0058	0.0010
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.6038	0.0487	0.0074
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.5011	0.0144	0.0013
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.6176	0.0276	0.0013
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.4381	0.0093	0.0000
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.3881	0.0081	0.0006
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5777	0.0251	0.0019
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5465	0.0175	0.0006
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4453	0.0159	0.0028
gen ● kimi-k3 (Temp: Unknown)	3,022	0.5608	0.0457	0.0152
gen ● minimax-m3 (Temp: Unknown) ▲ best	3,024	0.6078	0.0536	0.0060
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4412	0.0101	0.0006

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Top-k Outliers (Llama-3.2-3B-Instruct)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Top-k Outliers (Llama-3.2-3B-Instruct): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Top-k Outliers (Llama-3.2-3B-Instruct): performance over different AI class thresholds

Score distributions

Top-k Outliers (Llama-3.2-3B-Instruct): overall human

Overall · human onlyOpen full size ↗

Top-k Outliers (Llama-3.2-3B-Instruct): overall AI

Overall · AI onlyOpen full size ↗

Top-k Outliers (Llama-3.2-3B-Instruct): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Top-k Outliers (Llama-3.2-3B-Instruct): generator subsets

AI texts · one envelope per generator configOpen full size ↗

17Top-p Outliers (Llama-3.2-3B-Instruct)N34,582AUROC0.5061TPR @ 1% FPR0.0266TPR @ 0.1% FPR0.0035⌄

1% FPR threshold 0.03280.1% FPR threshold 0.0234

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.5061	0.0266	0.0035
prompt ● direct_reference	8,874	0.4959	0.0329	0.0038
prompt ● indirect_reference	8,798	0.4936	0.0298	0.0066
prompt ● revise	8,960	0.5192	0.0263	0.0016
prompt ● rewrite	7,950	0.5163	0.0164	0.0018
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.4697	0.0191	0.0019
gen ● claude-opus-5 (Temp: Unknown) ▼ worst	3,184	0.4897	0.0132	0.0006
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.5418	0.0194	0.0006
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.5192	0.0303	0.0056
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.5113	0.0398	0.0025
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5188	0.0188	0.0019
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.4622	0.0181	0.0031
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4722	0.0290	0.0048
gen ● kimi-k3 (Temp: Unknown)	3,022	0.5463	0.0410	0.0060
gen ● minimax-m3 (Temp: Unknown) ▲ best	3,024	0.5658	0.0430	0.0086
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4727	0.0226	0.0031

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Top-p Outliers (Llama-3.2-3B-Instruct)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Top-p Outliers (Llama-3.2-3B-Instruct): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Top-p Outliers (Llama-3.2-3B-Instruct): performance over different AI class thresholds

Score distributions

Top-p Outliers (Llama-3.2-3B-Instruct): overall human

Overall · human onlyOpen full size ↗

Top-p Outliers (Llama-3.2-3B-Instruct): overall AI

Overall · AI onlyOpen full size ↗

Top-p Outliers (Llama-3.2-3B-Instruct): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Top-p Outliers (Llama-3.2-3B-Instruct): generator subsets

AI texts · one envelope per generator configOpen full size ↗

18Perplexity (Llama-3.2-3B)N34,582AUROC0.4986TPR @ 1% FPR0.0093TPR @ 0.1% FPR0.0004⌄

1% FPR threshold 3.80030.1% FPR threshold 1.3646

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.4986	0.0093	0.0004
prompt ● direct_reference	8,874	0.4665	0.0092	0.0002
prompt ● indirect_reference	8,798	0.5646	0.0159	0.0007
prompt ● revise	8,960	0.5107	0.0069	0.0004
prompt ● rewrite	7,950	0.4482	0.0045	0.0003
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.5539	0.0136	0.0012
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.4993	0.0082	0.0000
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.6090	0.0107	0.0000
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.4113	0.0056	0.0000
gen ● gemini-3.8-flash (Temp: Unknown) ▼ worst	3,214	0.3803	0.0031	0.0000
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5467	0.0063	0.0000
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5465	0.0087	0.0000
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4152	0.0069	0.0007
gen ● kimi-k3 (Temp: Unknown) ▲ best	3,022	0.5348	0.0192	0.0013
gen ● minimax-m3 (Temp: Unknown)	3,024	0.5826	0.0146	0.0013
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4037	0.0057	0.0000

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Perplexity (Llama-3.2-3B)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Perplexity (Llama-3.2-3B): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Perplexity (Llama-3.2-3B): performance over different AI class thresholds

Score distributions

Perplexity (Llama-3.2-3B): overall human

Overall · human onlyOpen full size ↗

Overall · AI onlyOpen full size ↗

Perplexity (Llama-3.2-3B): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Perplexity (Llama-3.2-3B): generator subsets

AI texts · one envelope per generator configOpen full size ↗

19Top-k Outliers (Llama-3.2-3B)N34,582AUROC0.4913TPR @ 1% FPR0.0130TPR @ 0.1% FPR0.0006⌄

1% FPR threshold 0.01960.1% FPR threshold 0.0043

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.4913	0.0130	0.0006
prompt ● direct_reference	8,874	0.4797	0.0207	0.0002
prompt ● indirect_reference	8,798	0.5486	0.0195	0.0011
prompt ● revise	8,960	0.4894	0.0067	0.0004
prompt ● rewrite	7,950	0.4438	0.0040	0.0005
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.5577	0.0179	0.0019
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.4893	0.0101	0.0000
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.5959	0.0138	0.0000
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.3952	0.0074	0.0000
gen ● gemini-3.8-flash (Temp: Unknown)	3,214	0.3528	0.0056	0.0000
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5602	0.0144	0.0000
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5356	0.0081	0.0006
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4216	0.0111	0.0007
gen ● kimi-k3 (Temp: Unknown) ▲ best	3,022	0.5265	0.0298	0.0013
gen ● minimax-m3 (Temp: Unknown)	3,024	0.5719	0.0218	0.0013
gen ● qwen3.7-max (Temp: Unknown) ▼ worst	3,184	0.3976	0.0038	0.0006

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: Top-k Outliers (Llama-3.2-3B)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Top-k Outliers (Llama-3.2-3B): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Top-k Outliers (Llama-3.2-3B): performance over different AI class thresholds

Score distributions

Top-k Outliers (Llama-3.2-3B): overall human

Overall · human onlyOpen full size ↗

Top-k Outliers (Llama-3.2-3B): overall AI

Overall · AI onlyOpen full size ↗

Top-k Outliers (Llama-3.2-3B): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

Top-k Outliers (Llama-3.2-3B): generator subsets

AI texts · one envelope per generator configOpen full size ↗

20Entropy (Llama-3.2-3B)N34,582AUROC0.4832TPR @ 1% FPR0.0072TPR @ 0.1% FPR0.0005⌄

1% FPR threshold 1.35260.1% FPR threshold 0.3718

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.4832	0.0072	0.0005
prompt ● direct_reference	8,874	0.4268	0.0054	0.0002
prompt ● indirect_reference	8,798	0.5334	0.0123	0.0009
prompt ● revise	8,960	0.5070	0.0065	0.0004
prompt ● rewrite	7,950	0.4641	0.0045	0.0003
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.5370	0.0111	0.0012
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.4806	0.0063	0.0000
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.5624	0.0075	0.0000
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.4164	0.0043	0.0000
gen ● gemini-3.8-flash (Temp: Unknown) ▼ worst	3,214	0.4229	0.0025	0.0000
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.5192	0.0063	0.0000
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.5365	0.0081	0.0000
gen ● grok-4.3 (Temp: Unknown)	2,892	0.3902	0.0055	0.0007
gen ● kimi-k3 (Temp: Unknown) ▲ best	3,022	0.4962	0.0126	0.0013
gen ● minimax-m3 (Temp: Unknown)	3,024	0.5310	0.0126	0.0013
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4170	0.0031	0.0006

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

Entropy (Llama-3.2-3B): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

Entropy (Llama-3.2-3B): performance over different AI class thresholds

Score distributions

Overall · human onlyOpen full size ↗

Overall · AI onlyOpen full size ↗

AI texts · one envelope per prompt subsetOpen full size ↗

Entropy (Llama-3.2-3B): generator subsets

AI texts · one envelope per generator configOpen full size ↗

21FastDetectGPT (Llama-3.2-3B-Instruct)N34,582AUROC0.4804TPR @ 1% FPR0.0145TPR @ 0.1% FPR0.0013⌄

1% FPR threshold 2.20840.1% FPR threshold 5.4559

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.4804	0.0145	0.0013
prompt ● direct_reference	8,874	0.4685	0.0131	0.0014
prompt ● indirect_reference	8,798	0.4727	0.0080	0.0007
prompt ● revise	8,960	0.4708	0.0199	0.0022
prompt ● rewrite	7,950	0.5132	0.0171	0.0008
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.4546	0.0099	0.0012
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.4599	0.0113	0.0006
gen ● claude-opus-5-5 (Temp: Unknown)	3,192	0.4578	0.0132	0.0000
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.5458	0.0223	0.0012
gen ● gemini-3.8-flash (Temp: Unknown) ▲ best	3,214	0.4959	0.0255	0.0037
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.4524	0.0107	0.0006
gen ● gpt-5.6-sol (Temp: Unknown) ▼ worst	3,208	0.3676	0.0037	0.0000
gen ● grok-4.3 (Temp: Unknown)	2,892	0.5034	0.0097	0.0007
gen ● kimi-k3 (Temp: Unknown)	3,022	0.5274	0.0225	0.0046
gen ● minimax-m3 (Temp: Unknown)	3,024	0.5422	0.0198	0.0007
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.4853	0.0107	0.0006

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: FastDetectGPT (Llama-3.2-3B-Instruct)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

FastDetectGPT (Llama-3.2-3B-Instruct): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

FastDetectGPT (Llama-3.2-3B-Instruct): performance over different AI class thresholds

Score distributions

FastDetectGPT (Llama-3.2-3B-Instruct): overall human

Overall · human onlyOpen full size ↗

FastDetectGPT (Llama-3.2-3B-Instruct): overall AI

Overall · AI onlyOpen full size ↗

FastDetectGPT (Llama-3.2-3B-Instruct): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

FastDetectGPT (Llama-3.2-3B-Instruct): generator subsets

AI texts · one envelope per generator configOpen full size ↗

22FastDetectGPT (Llama-3.2-3B)N34,582AUROC0.4436TPR @ 1% FPR0.0230TPR @ 0.1% FPR0.0014⌄

1% FPR threshold -2.68350.1% FPR threshold -4.6137

Subset	N	AUROC	TPR @ 1% FPR	TPR @ 0.1% FPR
Overall	34,582	0.4436	0.0230	0.0014
prompt ● direct_reference	8,874	0.3911	0.0252	0.0011
prompt ● indirect_reference	8,798	0.3638	0.0150	0.0009
prompt ● revise	8,960	0.4619	0.0250	0.0016
prompt ● rewrite	7,950	0.5694	0.0272	0.0023
gen ● claude-opus-4-5-20251101 (Temp: Unknown)	3,242	0.4305	0.0247	0.0006
gen ● claude-opus-5 (Temp: Unknown)	3,184	0.4188	0.0144	0.0025
gen ● claude-opus-5-5 (Temp: Unknown) ▼ worst	3,192	0.2876	0.0031	0.0006
gen ● gemini-3.1-pro-preview (Temp: Unknown)	3,232	0.5269	0.0328	0.0025
gen ● gemini-3.8-flash (Temp: Unknown) ▲ best	3,214	0.6834	0.0853	0.0019
gen ● gpt-5.4 (Temp: Unknown)	3,188	0.3657	0.0069	0.0013
gen ● gpt-5.6-sol (Temp: Unknown)	3,208	0.4445	0.0156	0.0019
gen ● grok-4.3 (Temp: Unknown)	2,892	0.4473	0.0138	0.0007
gen ● kimi-k3 (Temp: Unknown)	3,022	0.3637	0.0146	0.0020
gen ● minimax-m3 (Temp: Unknown)	3,024	0.3353	0.0099	0.0007
gen ● qwen3.7-max (Temp: Unknown)	3,184	0.5641	0.0295	0.0013

Every subset is scored at the corpus-wide thresholds above, so its own false positive rate can differ from the budget. ▲ and ▼ mark the subsets with the highest and lowest TPR @ 1% FPR.

Threshold sweep: FastDetectGPT (Llama-3.2-3B)

TPR and FPR at every threshold · both FPR operating points markedOpen full size ↗

FastDetectGPT (Llama-3.2-3B): TPR against minimum cosdist

AI rows with cosdist below the cutoff are dropped · human rows and thresholds are unchangedOpen full size ↗

FastDetectGPT (Llama-3.2-3B): performance over different AI class thresholds

Score distributions

FastDetectGPT (Llama-3.2-3B): overall human

Overall · human onlyOpen full size ↗

FastDetectGPT (Llama-3.2-3B): overall AI

Overall · AI onlyOpen full size ↗

FastDetectGPT (Llama-3.2-3B): prompt subsets

AI texts · one envelope per prompt subsetOpen full size ↗

FastDetectGPT (Llama-3.2-3B): generator subsets

AI texts · one envelope per generator configOpen full size ↗

How far do the rewrites move?

Every distance measure between each source and its rewrite: overall, then with every prompt subset and every generator config overlaid on one chart each.

jaccard_1overall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

jaccard_2overall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

levenshteinoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

softngramoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

cosdistoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

bertscoreoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

bertscore_precisionoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

bertscore_recalloverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

moverscoreoverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

rerankeroverall · prompt subsets · generator configs⌄

OverallOpen full size ↗

One envelope per prompt subsetOpen full size ↗

One envelope per generator configOpen full size ↗

Appendix

Reference material: how the dataset was built and evaluated, its schema and loading code, univariate statistics, cross-statistic correlations, and the link to the per-model analytics page.

How the dataset was built11 generator configs · 4 prompt subsets · evaluation protocol⌄

Every row pairs original (Human) and final_response (AI): a human-written source and the text a generator model produced from it. Inclusion filtering happens while the dataset is built, so this report evaluates the published corpus exactly as it is.

Generator configs

Generator model	Temp	Sampling	Rows	Share
claude-opus-4-5-20251101	—	type disabled	1,621	9.4%
claude-opus-5	—	type disabled	1,592	9.2%
claude-opus-5-5	—	effort low	1,596	9.2%
gemini-3.1-pro-preview	—	thinking_level low	1,616	9.3%
gemini-3.8-flash	—	thinking_level low	1,607	9.3%
gpt-5.4	—	reasoning_effort none	1,594	9.2%
gpt-5.6-sol	—	reasoning_effort none	1,604	9.3%
x-ai/grok-4.3	—	enabled False · max_completion_tokens 16000	1,446	8.4%
moonshotai/kimi-k3	—	enabled False · max_completion_tokens 16000	1,511	8.7%
minimax/minimax-m3	—	enabled False · max_completion_tokens 16000	1,512	8.7%
qwen/qwen3.7-max	—	enabled False · max_completion_tokens 16000	1,592	9.2%
Total			17,291	100%

Prompt subsets

Instructions are grouped by their first message. Counts cover all 17,291 rows.

direct_reference4,437 rows

single-turn · 60 instructions

Show the 60 instructions ▾

89× Write a new piece of text that uses this text as a citation or reference.
88× Write a detailed rebuttal or critique challenging the core ideas, themes, or facts presented in this text, using the same language and roughly matching the intellectual level of the text. The final text must have the topic <<topic>>.
87× Write a prequel to this text, describing the events, thoughts, or historical context that immediately led up to this piece being written, making sure to use similar language and wording as the author, as if it was written by the author. The final text must have the topic <<topic>>.
85× Write a different piece of text on the same topic. The final text must have the format <<format>> and topic <<topic>>.
84× Write a followup to this piece of text, in the same style and language as the text. The final text must have the format <<format>> and topic <<topic>>.
84× Write an alternate version of this text, it should follow the style and format of the original text.
84× Write another piece of text that resembles the style and language of this text.
83× Write a followup to this piece of text, in the same style and language as the text. The final text must have the topic <<topic>>.
83× Write another piece of text that is of a similar theme or background as this text. The final text must have the format <<format>>.
82× Take a subpoint within this text and write a full new text about it, using the language and wording from the text as appropriate. The final text must have the format <<format>>.
81× Take a subpoint within this text and write a full new text about it, using the language and wording from the text as appropriate.
81× Write a detailed rebuttal or critique challenging the core ideas, themes, or facts presented in this text, using the same language and roughly matching the intellectual level of the text. The final text must have the format <<format>>.
81× Write a piece of text in the same style and language as this text, but is directly contrasting to the themes, ideas, and/or details in this text. The final text must have the topic <<topic>>.
80× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should write the new text as if it was a rough draft being typed out on a phone.
80× Write a followup to this piece of text, in the same style and language as the text.
80× Write a new piece of text that is on a completely different topic, but reuses verbatim as many phrases and sentences from this text as possible.
79× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should make sure there are slight, subtle grammatical mistakes, to make it sound like it was written by a human. The final text must have the topic <<topic>>.
78× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should use varied sentence lengths, and remove all common AI buzzwords such as 'tapestry', 'delve', or 'testament', to ensure that it sounds like it was written by a human. The final text must have the topic <<topic>>.
78× Write a different piece of text on the same topic. The final text must have the format <<format>>.
78× Write a piece of text with a different topic that the author of this text might write about. The final text must have the format <<format>>.
77× Take a subpoint within this text and write a full new text about it, using the language and wording from the text as appropriate. The final text must have the format <<format>> and topic <<topic>>.
77× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should write the new text as if it was a rough draft being typed out on a phone. The final text must have the format <<format>> and topic <<topic>>.
77× Write a new piece of text that uses this text as a citation or reference. The final text must have the topic <<topic>>.
77× Write a piece of text with a different topic that the author of this text might write about. The final text must have the topic <<topic>>.
76× Write a new piece of text that uses this text as a citation or reference. The final text must have the format <<format>>.
76× Write a prequel to this text, describing the events, thoughts, or historical context that immediately led up to this piece being written, making sure to use similar language and wording as the author, as if it was written by the author.
76× Write another piece of text that resembles the style and language of this text. The final text must have the format <<format>>.
76× Write another piece of text that resembles the style and language of this text. The final text must have the topic <<topic>>.
75× Write another piece of text that is of a similar theme or background as this text.
75× Write another piece of text that is of a similar theme or background as this text. The final text must have the format <<format>> and topic <<topic>>.
74× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should use varied sentence lengths, and remove all common AI buzzwords such as 'tapestry', 'delve', or 'testament', to ensure that it sounds like it was written by a human. The final text must have the format <<format>>.
73× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should make sure there are slight, subtle grammatical mistakes, to make it sound like it was written by a human.
73× Write a new piece of text that uses this text as a citation or reference. The final text must have the format <<format>> and topic <<topic>>.
73× Write a piece of text in the same style and language as this text, but is directly contrasting to the themes, ideas, and/or details in this text. The final text must have the format <<format>> and topic <<topic>>.
73× Write a piece of text with a different topic that the author of this text might write about.
72× Take a subpoint within this text and write a full new text about it, using the language and wording from the text as appropriate. The final text must have the topic <<topic>>.
72× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should use varied sentence lengths, and remove all common AI buzzwords such as 'tapestry', 'delve', or 'testament', to ensure that it sounds like it was written by a human. The final text must have the format <<format>> and topic <<topic>>.
72× Write a piece of text in the same style and language as this text, but is directly contrasting to the themes, ideas, and/or details in this text.
70× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should make sure there are slight, subtle grammatical mistakes, to make it sound like it was written by a human. The final text must have the format <<format>> and topic <<topic>>.
70× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should write the new text as if it was a rough draft being typed out on a phone. The final text must have the format <<format>>.
70× Write a piece of text with a different topic that the author of this text might write about. The final text must have the format <<format>> and topic <<topic>>.
70× Write a prequel to this text, describing the events, thoughts, or historical context that immediately led up to this piece being written, making sure to use similar language and wording as the author, as if it was written by the author. The final text must have the format <<format>> and topic <<topic>>.
70× Write a prequel to this text, describing the events, thoughts, or historical context that immediately led up to this piece being written, making sure to use similar language and wording as the author, as if it was written by the author. The final text must have the format <<format>>.
69× Write a different piece of text on the same topic. The final text must have the topic <<topic>>.
68× Write an alternate version of this text, it should follow the style and format of the original text. The final text must have the format <<format>>.
67× Write a detailed rebuttal or critique challenging the core ideas, themes, or facts presented in this text, using the same language and roughly matching the intellectual level of the text. The final text must have the format <<format>> and topic <<topic>>.
67× Write a new piece of text that is on a completely different topic, but reuses verbatim as many phrases and sentences from this text as possible. The final text must have the topic <<topic>>.
66× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should make sure there are slight, subtle grammatical mistakes, to make it sound like it was written by a human. The final text must have the format <<format>>.
66× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should use varied sentence lengths, and remove all common AI buzzwords such as 'tapestry', 'delve', or 'testament', to ensure that it sounds like it was written by a human.
66× Write a detailed rebuttal or critique challenging the core ideas, themes, or facts presented in this text, using the same language and roughly matching the intellectual level of the text.
66× Write a different piece of text on the same topic.
66× Write an alternate version of this text, it should follow the style and format of the original text. The final text must have the format <<format>> and topic <<topic>>.
65× Write a new piece of text that is on a completely different topic, but reuses verbatim as many phrases and sentences from this text as possible. The final text must have the format <<format>>.
65× Write another piece of text that resembles the style and language of this text. The final text must have the format <<format>> and topic <<topic>>.
63× Write a new piece of text that is on a completely different topic, but reuses verbatim as many phrases and sentences from this text as possible. The final text must have the format <<format>> and topic <<topic>>.
63× Write an alternate version of this text, it should follow the style and format of the original text. The final text must have the topic <<topic>>.
62× Write another piece of text that is of a similar theme or background as this text. The final text must have the topic <<topic>>.
60× Write a followup to this piece of text, in the same style and language as the text. The final text must have the format <<format>>.
60× Write a piece of text in the same style and language as this text, but is directly contrasting to the themes, ideas, and/or details in this text. The final text must have the format <<format>>.
59× Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should write the new text as if it was a rough draft being typed out on a phone. The final text must have the topic <<topic>>.

indirect_reference4,399 rows

two-turn · 25 instructions

Show the 25 instructions ▾

243× Translate the given text to Chinese. Do not output anything besides what you were requested to write, and do not output any extra commentary.
236× Translate the given text to Spanish. Do not output anything besides what you were requested to write, and do not output any extra commentary.
232× Translate the given text to French. Do not output anything besides what you were requested to write, and do not output any extra commentary.
230× Translate the given text to German. Do not output anything besides what you were requested to write, and do not output any extra commentary.
220× Formulate a set of strict, absolute constraints regarding format, word count, vocabulary restrictions, and syntax rules that, if followed perfectly by an AI, would result in generating a text nearly identical to this one. Note that your response should be formatted as a prompt that would be sent to an AI requesting it to generate something, not just a list of rules. Do not output anything besides what you were requested to write, and do not output any extra commentary.
208× Translate the given text to Hindi. Do not output anything besides what you were requested to write, and do not output any extra commentary.
196× Create a prompt that might cause an LLM to generate an output resembling this text. Do not output anything besides what you were requested to write, and do not output any extra commentary.
195× This piece of text was cleverely generated by an LLM with a human supervising it so that the text seems authentic. Reverse engineer the process and come up with a prompt that is most probable to be the one used to generate the text. Do not output anything besides what you were requested to write, and do not output any extra commentary.
194× Summarize this text, attempting to preserve as much of the original language of the text as possible. Do not output anything besides what you were requested to write, and do not output any extra commentary.
188× Convert this text into a series of logical propositions or syllogisms that represent the core argument or narrative progression, stripped of all rhetorical flair. Do not output anything besides what you were requested to write, and do not output any extra commentary.
183× Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. The prompt should be designed to teach an LLM the pattern needed to generate the rest of the text from scratch. Do not output anything besides what you were requested to write, and do not output any extra commentary.
180× Take this text, but extract only the most meaningful sentences out of it to create a new text that is a series of disjoint sentences representing the original text. Do not output anything besides what you were requested to write, and do not output any extra commentary.
168× Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articles and punctuation exactly as they appear. Reproduce the document in full, in its original order and paragraph structure. Do not output a legend or a list of the tags you used. Do not output anything besides what you were requested to write, and do not output any extra commentary.
168× Take this text and replace one word in every four with a single underscore (use one underscore for each removed word, never several in a row). This way, it will be appropriate for a FIM task. You may decide which words to replace. Reproduce the rest of the document in full, in its original order and paragraph structure. Do not output anything besides what you were requested to write, and do not output any extra commentary.
163× Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The prompt may include verbatim pieces of the original text, however, it must be short and accurate, working as an effective compression of the original text. Do not output anything besides what you were requested to write, and do not output any extra commentary.
162× Create a prompt that would allow an LLM to generate this text as accurately/verbatim as possible, without the prompt containing any of the actual text. That is, you must use descriptors to accurately describe the text to generate. Do not output anything besides what you were requested to write, and do not output any extra commentary.
152× Provide only the first and last word of every sentence in this text, replacing the middle of each sentence with an ellipsis '...'. Maintain the original paragraph structure. Do not output anything besides what you were requested to write, and do not output any extra commentary.
151× Write a detailed descriptor for how this text is stylistically differentiated from other texts, including grammatical choices and word choices not present in other texts of this kind. Do not output anything besides what you were requested to write, and do not output any extra commentary.
145× Write a detailed descriptor for the style of writing of this text, including its qualitative and quantitative properties. Do not output anything besides what you were requested to write, and do not output any extra commentary.
143× Create a detailed dictionary that fully describes all the meaningful phrases and words this author utilizes to create this text, and how the author utilizes those words and phrases, and why. Do not output anything besides what you were requested to write, and do not output any extra commentary.
142× Envision the scenario in which the author wrote this text. Describe that scenario in exhaustive detail. Do not output anything besides what you were requested to write, and do not output any extra commentary.
141× Write a detailed descriptor for the manner in which the author of this uses language, and things such as their word choice. Do not output anything besides what you were requested to write, and do not output any extra commentary.
132× Reformat this text into a sensible, structured JSON object. Do not output anything besides what you were requested to write, and do not output any extra commentary.
121× Write a detailed descriptor for the (imaginary) personality of the author who wrote this text, including their livelihood and day to day activities, as well as how that influences their writing. Do not output anything besides what you were requested to write, and do not output any extra commentary.
106× Translate the entirety of this text into a sequence of emojis that captures the literal meaning, tone, and narrative progression of the writing. Do not use any letters or words. Do not output anything besides what you were requested to write, and do not output any extra commentary.

revise4,480 rows

2,195 single · 2,285 two-turn · 172 instructions

Show the 172 instructions ▾

88× Change the tone of this text to be more objective and unbiased.
81× Restructure this text to eliminate awkward phrasing and improve the overall rhythm.
80× Clarify this text by making the cause-and-effect relationships more obvious.
79× Restructure this text to arrange the points in order of importance.
78× Clarify this text by presenting the same information from a fresh angle.
75× Change the tone of this text to be more academic and scholarly.
75× Change the tone of this text to be more lighthearted and humorous.
75× Change the tone of this text to be more suited for a business audience.
75× Clarify this text by eliminating any ambiguous or vague language.
73× Restructure this text to improve the transitions and connections between ideas.
72× Expand this text by adding vivid imagery and sensory details to bring it to life.
71× Clarify this text by paraphrasing it with different vocabulary and varied sentence structures.
71× Expand this text by adding backstory and context to enrich understanding.
71× Expand this text by adding storytelling elements and dialogue to increase engagement.
70× Change the tone of this text to be more persuasive and convincing.
70× Change the tone of this text to be more urgent and compelling.
69× Clarify this text by making the argument and main points stronger and easier to understand.
69× Restructure this text to improve the logical flow and argument progression.
68× Adapt this text for a teenage audience.
68× Change the tone of this text to be more empathetic and understanding.
67× Adapt this text for an elderly adult with limited knowledge of modern day affairs.
67× Change the tone of this text to be more inspiring and motivational.
66× Restructure this text to read more naturally, as if written by a native English speaker.
64× Edit this text to trim the length without losing important information or tone.
64× Restructure this text to group related ideas and include clear section breaks.
63× Edit this text to eliminate redundant and filler words.
62× Adapt this text for a lay audience with no technical background.
62× Change the tone of this text to be more direct and bold.
62× Change the tone of this text to be more enthusiastic and energetic.
61× Change the tone of this text to be more analytical and logical.
61× Change the tone of this text to be more professional and authoritative.
61× Expand this text by adding real-world applications and specific details.
60× Adapt this text for a 5th-grade reading level.
60× Expand this text by adding concrete examples to illustrate the main points.
59× Edit this text to be more concise and powerful.
58× Adapt this text for social media sharing.
58× Clarify this text by improving the phrasing while preserving its original intent.
56× Change the tone of this text to be more conversational and approachable.
56× Clarify this text by defining any terms that might be unclear to the reader.
55× Change the tone of this text to be more relaxed and friendly.
53× Expand this text by adding supporting information to elaborate on the key points.
52× Restructure this text to ensure a clear introduction, body, and compelling conclusion.
50× Change the tone of this text to be more polite and diplomatic.
21× Clarify this text by presenting the same information from a fresh angle. The final text must have the format <<format>> and topic <<topic>>.
20× Change the tone of this text to be more objective and unbiased. The final text must have the format <<format>>.
20× Change the tone of this text to be more suited for a business audience. The final text must have the topic <<topic>>.
20× Edit this text to trim the length without losing important information or tone. The final text must have the format <<format>> and topic <<topic>>.
20× Expand this text by adding storytelling elements and dialogue to increase engagement. The final text must have the format <<format>>.
19× Change the tone of this text to be more direct and bold. The final text must have the format <<format>> and topic <<topic>>.
19× Change the tone of this text to be more enthusiastic and energetic. The final text must have the format <<format>>.
19× Change the tone of this text to be more relaxed and friendly. The final text must have the format <<format>>.
18× Change the tone of this text to be more empathetic and understanding. The final text must have the format <<format>> and topic <<topic>>.
18× Change the tone of this text to be more inspiring and motivational. The final text must have the format <<format>>.
18× Change the tone of this text to be more polite and diplomatic. The final text must have the format <<format>>.
18× Clarify this text by making the argument and main points stronger and easier to understand. The final text must have the format <<format>>.
18× Expand this text by adding concrete examples to illustrate the main points. The final text must have the format <<format>>.
18× Restructure this text to read more naturally, as if written by a native English speaker. The final text must have the topic <<topic>>.
17× Adapt this text for a 5th-grade reading level. The final text must have the format <<format>>.
17× Adapt this text for a teenage audience. The final text must have the topic <<topic>>.
17× Change the tone of this text to be more enthusiastic and energetic. The final text must have the format <<format>> and topic <<topic>>.
17× Restructure this text to group related ideas and include clear section breaks. The final text must have the topic <<topic>>.
16× Adapt this text for a lay audience with no technical background. The final text must have the format <<format>> and topic <<topic>>.
16× Adapt this text for a lay audience with no technical background. The final text must have the topic <<topic>>.
16× Adapt this text for social media sharing. The final text must have the format <<format>> and topic <<topic>>.
16× Change the tone of this text to be more conversational and approachable. The final text must have the format <<format>> and topic <<topic>>.
16× Change the tone of this text to be more conversational and approachable. The final text must have the format <<format>>.
16× Change the tone of this text to be more direct and bold. The final text must have the topic <<topic>>.
16× Change the tone of this text to be more empathetic and understanding. The final text must have the format <<format>>.
16× Change the tone of this text to be more suited for a business audience. The final text must have the format <<format>>.
16× Clarify this text by paraphrasing it with different vocabulary and varied sentence structures. The final text must have the format <<format>> and topic <<topic>>.
16× Edit this text to be more concise and powerful. The final text must have the topic <<topic>>.
16× Expand this text by adding backstory and context to enrich understanding. The final text must have the format <<format>> and topic <<topic>>.
16× Expand this text by adding supporting information to elaborate on the key points. The final text must have the format <<format>> and topic <<topic>>.
16× Restructure this text to arrange the points in order of importance. The final text must have the topic <<topic>>.
16× Restructure this text to group related ideas and include clear section breaks. The final text must have the format <<format>> and topic <<topic>>.
15× Adapt this text for social media sharing. The final text must have the topic <<topic>>.
15× Change the tone of this text to be more academic and scholarly. The final text must have the format <<format>> and topic <<topic>>.
15× Change the tone of this text to be more lighthearted and humorous. The final text must have the format <<format>> and topic <<topic>>.
15× Change the tone of this text to be more lighthearted and humorous. The final text must have the format <<format>>.
15× Clarify this text by eliminating any ambiguous or vague language. The final text must have the format <<format>> and topic <<topic>>.
15× Restructure this text to eliminate awkward phrasing and improve the overall rhythm. The final text must have the format <<format>> and topic <<topic>>.
15× Restructure this text to improve the logical flow and argument progression. The final text must have the topic <<topic>>.
14× Adapt this text for a lay audience with no technical background. The final text must have the format <<format>>.
14× Adapt this text for an elderly adult with limited knowledge of modern day affairs. The final text must have the topic <<topic>>.
14× Change the tone of this text to be more persuasive and convincing. The final text must have the format <<format>> and topic <<topic>>.
14× Clarify this text by defining any terms that might be unclear to the reader. The final text must have the format <<format>>.
14× Clarify this text by presenting the same information from a fresh angle. The final text must have the format <<format>>.
14× Clarify this text by presenting the same information from a fresh angle. The final text must have the topic <<topic>>.
14× Expand this text by adding real-world applications and specific details. The final text must have the format <<format>>.
14× Expand this text by adding real-world applications and specific details. The final text must have the topic <<topic>>.
14× Expand this text by adding supporting information to elaborate on the key points. The final text must have the topic <<topic>>.
14× Restructure this text to read more naturally, as if written by a native English speaker. The final text must have the format <<format>> and topic <<topic>>.
13× Adapt this text for a 5th-grade reading level. The final text must have the topic <<topic>>.
13× Adapt this text for a teenage audience. The final text must have the format <<format>> and topic <<topic>>.
13× Adapt this text for an elderly adult with limited knowledge of modern day affairs. The final text must have the format <<format>> and topic <<topic>>.
13× Change the tone of this text to be more direct and bold. The final text must have the format <<format>>.
13× Change the tone of this text to be more professional and authoritative. The final text must have the topic <<topic>>.
13× Clarify this text by defining any terms that might be unclear to the reader. The final text must have the topic <<topic>>.
13× Clarify this text by improving the phrasing while preserving its original intent. The final text must have the format <<format>>.
13× Clarify this text by making the cause-and-effect relationships more obvious. The final text must have the format <<format>> and topic <<topic>>.
13× Edit this text to eliminate redundant and filler words. The final text must have the format <<format>> and topic <<topic>>.
13× Restructure this text to read more naturally, as if written by a native English speaker. The final text must have the format <<format>>.
12× Change the tone of this text to be more academic and scholarly. The final text must have the format <<format>>.
12× Change the tone of this text to be more inspiring and motivational. The final text must have the topic <<topic>>.
12× Change the tone of this text to be more objective and unbiased. The final text must have the format <<format>> and topic <<topic>>.
12× Change the tone of this text to be more persuasive and convincing. The final text must have the topic <<topic>>.
12× Change the tone of this text to be more polite and diplomatic. The final text must have the topic <<topic>>.
12× Change the tone of this text to be more relaxed and friendly. The final text must have the format <<format>> and topic <<topic>>.
12× Change the tone of this text to be more urgent and compelling. The final text must have the topic <<topic>>.
12× Clarify this text by making the argument and main points stronger and easier to understand. The final text must have the format <<format>> and topic <<topic>>.
12× Expand this text by adding concrete examples to illustrate the main points. The final text must have the topic <<topic>>.
12× Expand this text by adding real-world applications and specific details. The final text must have the format <<format>> and topic <<topic>>.
12× Expand this text by adding vivid imagery and sensory details to bring it to life. The final text must have the format <<format>>.
12× Expand this text by adding vivid imagery and sensory details to bring it to life. The final text must have the topic <<topic>>.
12× Restructure this text to arrange the points in order of importance. The final text must have the format <<format>> and topic <<topic>>.
12× Restructure this text to eliminate awkward phrasing and improve the overall rhythm. The final text must have the format <<format>>.
12× Restructure this text to improve the logical flow and argument progression. The final text must have the format <<format>> and topic <<topic>>.
12× Restructure this text to improve the transitions and connections between ideas. The final text must have the format <<format>>.
11× Adapt this text for social media sharing. The final text must have the format <<format>>.
11× Change the tone of this text to be more academic and scholarly. The final text must have the topic <<topic>>.
11× Change the tone of this text to be more analytical and logical. The final text must have the format <<format>> and topic <<topic>>.
11× Change the tone of this text to be more analytical and logical. The final text must have the topic <<topic>>.
11× Change the tone of this text to be more conversational and approachable. The final text must have the topic <<topic>>.
11× Change the tone of this text to be more persuasive and convincing. The final text must have the format <<format>>.
11× Change the tone of this text to be more polite and diplomatic. The final text must have the format <<format>> and topic <<topic>>.
11× Clarify this text by eliminating any ambiguous or vague language. The final text must have the format <<format>>.
11× Clarify this text by making the cause-and-effect relationships more obvious. The final text must have the format <<format>>.
11× Clarify this text by paraphrasing it with different vocabulary and varied sentence structures. The final text must have the topic <<topic>>.
11× Edit this text to eliminate redundant and filler words. The final text must have the format <<format>>.
11× Edit this text to eliminate redundant and filler words. The final text must have the topic <<topic>>.
11× Edit this text to trim the length without losing important information or tone. The final text must have the topic <<topic>>.
11× Expand this text by adding backstory and context to enrich understanding. The final text must have the topic <<topic>>.
11× Expand this text by adding vivid imagery and sensory details to bring it to life. The final text must have the format <<format>> and topic <<topic>>.
11× Restructure this text to ensure a clear introduction, body, and compelling conclusion. The final text must have the format <<format>> and topic <<topic>>.
11× Restructure this text to ensure a clear introduction, body, and compelling conclusion. The final text must have the topic <<topic>>.
11× Restructure this text to group related ideas and include clear section breaks. The final text must have the format <<format>>.
10× Adapt this text for an elderly adult with limited knowledge of modern day affairs. The final text must have the format <<format>>.
10× Change the tone of this text to be more inspiring and motivational. The final text must have the format <<format>> and topic <<topic>>.
10× Change the tone of this text to be more lighthearted and humorous. The final text must have the topic <<topic>>.
10× Clarify this text by eliminating any ambiguous or vague language. The final text must have the topic <<topic>>.
10× Clarify this text by making the argument and main points stronger and easier to understand. The final text must have the topic <<topic>>.
10× Clarify this text by paraphrasing it with different vocabulary and varied sentence structures. The final text must have the format <<format>>.
10× Edit this text to be more concise and powerful. The final text must have the format <<format>> and topic <<topic>>.
10× Restructure this text to improve the logical flow and argument progression. The final text must have the format <<format>>.
9× Adapt this text for a 5th-grade reading level. The final text must have the format <<format>> and topic <<topic>>.
9× Adapt this text for a teenage audience. The final text must have the format <<format>>.
9× Change the tone of this text to be more analytical and logical. The final text must have the format <<format>>.
9× Change the tone of this text to be more objective and unbiased. The final text must have the topic <<topic>>.
9× Change the tone of this text to be more professional and authoritative. The final text must have the format <<format>> and topic <<topic>>.
9× Change the tone of this text to be more relaxed and friendly. The final text must have the topic <<topic>>.
9× Change the tone of this text to be more urgent and compelling. The final text must have the format <<format>> and topic <<topic>>.
9× Change the tone of this text to be more urgent and compelling. The final text must have the format <<format>>.
9× Clarify this text by improving the phrasing while preserving its original intent. The final text must have the format <<format>> and topic <<topic>>.
9× Clarify this text by improving the phrasing while preserving its original intent. The final text must have the topic <<topic>>.
9× Expand this text by adding backstory and context to enrich understanding. The final text must have the format <<format>>.
9× Expand this text by adding storytelling elements and dialogue to increase engagement. The final text must have the format <<format>> and topic <<topic>>.
9× Expand this text by adding storytelling elements and dialogue to increase engagement. The final text must have the topic <<topic>>.
9× Restructure this text to arrange the points in order of importance. The final text must have the format <<format>>.
9× Restructure this text to ensure a clear introduction, body, and compelling conclusion. The final text must have the format <<format>>.
9× Restructure this text to improve the transitions and connections between ideas. The final text must have the format <<format>> and topic <<topic>>.
9× Restructure this text to improve the transitions and connections between ideas. The final text must have the topic <<topic>>.
8× Change the tone of this text to be more suited for a business audience. The final text must have the format <<format>> and topic <<topic>>.
8× Expand this text by adding supporting information to elaborate on the key points. The final text must have the format <<format>>.
8× Restructure this text to eliminate awkward phrasing and improve the overall rhythm. The final text must have the topic <<topic>>.
7× Clarify this text by making the cause-and-effect relationships more obvious. The final text must have the topic <<topic>>.
7× Edit this text to be more concise and powerful. The final text must have the format <<format>>.
6× Change the tone of this text to be more empathetic and understanding. The final text must have the topic <<topic>>.
6× Clarify this text by defining any terms that might be unclear to the reader. The final text must have the format <<format>> and topic <<topic>>.
6× Expand this text by adding concrete examples to illustrate the main points. The final text must have the format <<format>> and topic <<topic>>.
5× Change the tone of this text to be more enthusiastic and energetic. The final text must have the topic <<topic>>.
5× Change the tone of this text to be more professional and authoritative. The final text must have the format <<format>>.
3× Edit this text to trim the length without losing important information or tone. The final text must have the format <<format>>.

rewrite3,975 rows

1,880 single · 2,095 two-turn · 104 instructions

Show the 104 instructions ▾

118× Edit this text by upgrading the verbs to be more dynamic and action-oriented, leaving the subjects, objects, and overall length completely unchanged.
112× Rewrite the weakest section in this text, but do not make any modifications to any other section.
107× Edit this text by completely rewriting the topic sentence of every paragraph, while preserving the supporting sentences in those paragraphs verbatim.
107× Modify this text so that it is as short as possible. At least a quarter of the original text must be repeated verbatim, and at most half of it; the remainder must be new prose.
106× Edit this text by removing all adverbs, making only minimal adjustments to the surrounding grammar to ensure the sentences remain structurally sound.
105× Edit this text by replacing all existing transition words and phrases (e.g., 'however', 'therefore', 'in addition') with different, synonymous transitions, making no other changes.
104× Edit this text by inserting a single, highly engaging introductory sentence at the very beginning, without modifying any of the original text that follows.
104× Edit this text by replacing all instances of passive voice with active voice, without altering any sentences that are already in the active voice.
104× Edit this text by replacing exactly one to three sentences, without changing anything else.
104× Rewrite only the opening hook and the final concluding sentence of this text, leaving the entire middle portion exactly as it was originally written.
102× Edit this text by removing exactly one to three sentences from it (the least necessary), you may only make minimal modifications to adjacent sentences for better flow.
102× Edit this text by removing one section from it (the least necessary), you may only make minimal modifications to adjacent sections for better flow.
101× Edit this text by removing five to ten percent of the words from it (the least necessary), without changing anything else.
101× Modify this text to be as long as possible. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose.
100× Edit this text by replacing generic adjectives with highly specific, evocative ones, modifying no more than ten percent of the total words.
96× Edit this text by combining at least four pairs of short, adjacent sentences into complex sentences using appropriate conjunctions, leaving all other sentences untouched.
94× Rewrite exactly one section of this text so that its wording is entirely new, and leave every other section untouched. The rewritten section must share no sentence with the original.
92× Edit this text by replacing the final sentence with a more open-ended, thought-provoking statement, without altering any preceding text.
91× Edit this text by inserting exactly one to three sentences (in appropriate locations), you may only make minimal modifications to transitions in adjacent sentences for better flow.
89× Edit this text by inserting a singular new section (in an appropriate locations), you may only make minimal modifications to transitions in adjacent sections for better flow.
89× Edit this text by replacing five to ten percent of words in it, without changing anything else.
88× Improve this text by rewriting it. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose.
87× Edit this text by inserting twenty to thirty words (in appropriate locations), without changing anything else.
86× Edit this text by splitting the three longest sentences into shorter, distinct sentences, without changing the vocabulary or structure of the rest of the text.
82× Edit this text by replacing, removing or inserting exactly one to three sentences, without changing anything else.
71× Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else.
31× Edit this text by completely rewriting the topic sentence of every paragraph, while preserving the supporting sentences in those paragraphs verbatim. The final text must have the format <<format>>.
28× Improve this text by rewriting it. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the format <<format>> and topic <<topic>>.
28× Improve this text by rewriting it. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the topic <<topic>>.
26× Edit this text by completely rewriting the topic sentence of every paragraph, while preserving the supporting sentences in those paragraphs verbatim. The final text must have the format <<format>> and topic <<topic>>.
26× Edit this text by removing one section from it (the least necessary), you may only make minimal modifications to adjacent sections for better flow. The final text must have the format <<format>>.
25× Edit this text by inserting a single, highly engaging introductory sentence at the very beginning, without modifying any of the original text that follows. The final text must have the topic <<topic>>.
25× Edit this text by replacing all instances of passive voice with active voice, without altering any sentences that are already in the active voice. The final text must have the topic <<topic>>.
24× Edit this text by upgrading the verbs to be more dynamic and action-oriented, leaving the subjects, objects, and overall length completely unchanged. The final text must have the topic <<topic>>.
24× Modify this text so that it is as short as possible. At least a quarter of the original text must be repeated verbatim, and at most half of it; the remainder must be new prose. The final text must have the format <<format>> and topic <<topic>>.
24× Modify this text so that it is as short as possible. At least a quarter of the original text must be repeated verbatim, and at most half of it; the remainder must be new prose. The final text must have the topic <<topic>>.
23× Edit this text by inserting exactly one to three sentences (in appropriate locations), you may only make minimal modifications to transitions in adjacent sentences for better flow. The final text must have the format <<format>> and topic <<topic>>.
23× Edit this text by inserting exactly one to three sentences (in appropriate locations), you may only make minimal modifications to transitions in adjacent sentences for better flow. The final text must have the format <<format>>.
23× Edit this text by removing all adverbs, making only minimal adjustments to the surrounding grammar to ensure the sentences remain structurally sound. The final text must have the format <<format>> and topic <<topic>>.
22× Edit this text by inserting a single, highly engaging introductory sentence at the very beginning, without modifying any of the original text that follows. The final text must have the format <<format>>.
22× Edit this text by inserting a singular new section (in an appropriate locations), you may only make minimal modifications to transitions in adjacent sections for better flow. The final text must have the format <<format>> and topic <<topic>>.
22× Edit this text by removing one section from it (the least necessary), you may only make minimal modifications to adjacent sections for better flow. The final text must have the format <<format>> and topic <<topic>>.
22× Edit this text by removing one section from it (the least necessary), you may only make minimal modifications to adjacent sections for better flow. The final text must have the topic <<topic>>.
22× Edit this text by replacing the final sentence with a more open-ended, thought-provoking statement, without altering any preceding text. The final text must have the topic <<topic>>.
22× Modify this text to be as long as possible. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the format <<format>> and topic <<topic>>.
22× Rewrite only the opening hook and the final concluding sentence of this text, leaving the entire middle portion exactly as it was originally written. The final text must have the format <<format>> and topic <<topic>>.
21× Edit this text by inserting a single, highly engaging introductory sentence at the very beginning, without modifying any of the original text that follows. The final text must have the format <<format>> and topic <<topic>>.
21× Edit this text by inserting twenty to thirty words (in appropriate locations), without changing anything else. The final text must have the format <<format>>.
21× Edit this text by removing five to ten percent of the words from it (the least necessary), without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
21× Edit this text by replacing all instances of passive voice with active voice, without altering any sentences that are already in the active voice. The final text must have the format <<format>>.
21× Edit this text by replacing five to ten percent of words in it, without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
21× Edit this text by replacing, removing or inserting exactly one to three sentences, without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
21× Rewrite the weakest section in this text, but do not make any modifications to any other section. The final text must have the format <<format>> and topic <<topic>>.
21× Rewrite the weakest section in this text, but do not make any modifications to any other section. The final text must have the topic <<topic>>.
20× Edit this text by removing all adverbs, making only minimal adjustments to the surrounding grammar to ensure the sentences remain structurally sound. The final text must have the format <<format>>.
20× Edit this text by replacing all existing transition words and phrases (e.g., 'however', 'therefore', 'in addition') with different, synonymous transitions, making no other changes. The final text must have the format <<format>> and topic <<topic>>.
20× Edit this text by replacing exactly one to three sentences, without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
20× Edit this text by splitting the three longest sentences into shorter, distinct sentences, without changing the vocabulary or structure of the rest of the text. The final text must have the format <<format>>.
20× Modify this text so that it is as short as possible. At least a quarter of the original text must be repeated verbatim, and at most half of it; the remainder must be new prose. The final text must have the format <<format>>.
20× Modify this text to be as long as possible. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the topic <<topic>>.
20× Rewrite exactly one section of this text so that its wording is entirely new, and leave every other section untouched. The rewritten section must share no sentence with the original. The final text must have the format <<format>> and topic <<topic>>.
19× Edit this text by combining at least four pairs of short, adjacent sentences into complex sentences using appropriate conjunctions, leaving all other sentences untouched. The final text must have the topic <<topic>>.
19× Edit this text by inserting a singular new section (in an appropriate locations), you may only make minimal modifications to transitions in adjacent sections for better flow. The final text must have the topic <<topic>>.
19× Edit this text by replacing the final sentence with a more open-ended, thought-provoking statement, without altering any preceding text. The final text must have the format <<format>>.
19× Rewrite exactly one section of this text so that its wording is entirely new, and leave every other section untouched. The rewritten section must share no sentence with the original. The final text must have the topic <<topic>>.
18× Edit this text by inserting twenty to thirty words (in appropriate locations), without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
18× Edit this text by replacing exactly one to three sentences, without changing anything else. The final text must have the format <<format>>.
18× Edit this text by replacing five to ten percent of words in it, without changing anything else. The final text must have the topic <<topic>>.
18× Edit this text by upgrading the verbs to be more dynamic and action-oriented, leaving the subjects, objects, and overall length completely unchanged. The final text must have the format <<format>>.
18× Rewrite only the opening hook and the final concluding sentence of this text, leaving the entire middle portion exactly as it was originally written. The final text must have the format <<format>>.
17× Edit this text by completely rewriting the topic sentence of every paragraph, while preserving the supporting sentences in those paragraphs verbatim. The final text must have the topic <<topic>>.
17× Edit this text by inserting exactly one to three sentences (in appropriate locations), you may only make minimal modifications to transitions in adjacent sentences for better flow. The final text must have the topic <<topic>>.
17× Edit this text by inserting twenty to thirty words (in appropriate locations), without changing anything else. The final text must have the topic <<topic>>.
17× Edit this text by removing exactly one to three sentences from it (the least necessary), you may only make minimal modifications to adjacent sentences for better flow. The final text must have the format <<format>> and topic <<topic>>.
17× Edit this text by splitting the three longest sentences into shorter, distinct sentences, without changing the vocabulary or structure of the rest of the text. The final text must have the topic <<topic>>.
17× Modify this text to be as long as possible. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the format <<format>>.
16× Edit this text by removing exactly one to three sentences from it (the least necessary), you may only make minimal modifications to adjacent sentences for better flow. The final text must have the format <<format>>.
16× Edit this text by replacing all instances of passive voice with active voice, without altering any sentences that are already in the active voice. The final text must have the format <<format>> and topic <<topic>>.
16× Edit this text by replacing five to ten percent of words in it, without changing anything else. The final text must have the format <<format>>.
16× Edit this text by replacing generic adjectives with highly specific, evocative ones, modifying no more than ten percent of the total words. The final text must have the format <<format>>.
16× Edit this text by splitting the three longest sentences into shorter, distinct sentences, without changing the vocabulary or structure of the rest of the text. The final text must have the format <<format>> and topic <<topic>>.
15× Edit this text by inserting a singular new section (in an appropriate locations), you may only make minimal modifications to transitions in adjacent sections for better flow. The final text must have the format <<format>>.
15× Edit this text by replacing generic adjectives with highly specific, evocative ones, modifying no more than ten percent of the total words. The final text must have the format <<format>> and topic <<topic>>.
15× Edit this text by replacing generic adjectives with highly specific, evocative ones, modifying no more than ten percent of the total words. The final text must have the topic <<topic>>.
15× Edit this text by replacing the final sentence with a more open-ended, thought-provoking statement, without altering any preceding text. The final text must have the format <<format>> and topic <<topic>>.
15× Rewrite exactly one section of this text so that its wording is entirely new, and leave every other section untouched. The rewritten section must share no sentence with the original. The final text must have the format <<format>>.
15× Rewrite only the opening hook and the final concluding sentence of this text, leaving the entire middle portion exactly as it was originally written. The final text must have the topic <<topic>>.
15× Rewrite the weakest section in this text, but do not make any modifications to any other section. The final text must have the format <<format>>.
14× Edit this text by removing five to ten percent of the words from it (the least necessary), without changing anything else. The final text must have the format <<format>>.
14× Edit this text by replacing, removing or inserting exactly one to three sentences, without changing anything else. The final text must have the format <<format>>.
14× Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else. The final text must have the format <<format>> and topic <<topic>>.
14× Edit this text by upgrading the verbs to be more dynamic and action-oriented, leaving the subjects, objects, and overall length completely unchanged. The final text must have the format <<format>> and topic <<topic>>.
13× Edit this text by combining at least four pairs of short, adjacent sentences into complex sentences using appropriate conjunctions, leaving all other sentences untouched. The final text must have the format <<format>>.
13× Edit this text by removing five to ten percent of the words from it (the least necessary), without changing anything else. The final text must have the topic <<topic>>.
13× Edit this text by replacing exactly one to three sentences, without changing anything else. The final text must have the topic <<topic>>.
12× Edit this text by combining at least four pairs of short, adjacent sentences into complex sentences using appropriate conjunctions, leaving all other sentences untouched. The final text must have the format <<format>> and topic <<topic>>.
12× Edit this text by replacing all existing transition words and phrases (e.g., 'however', 'therefore', 'in addition') with different, synonymous transitions, making no other changes. The final text must have the format <<format>>.
12× Edit this text by replacing all existing transition words and phrases (e.g., 'however', 'therefore', 'in addition') with different, synonymous transitions, making no other changes. The final text must have the topic <<topic>>.
11× Edit this text by removing exactly one to three sentences from it (the least necessary), you may only make minimal modifications to adjacent sentences for better flow. The final text must have the topic <<topic>>.
11× Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else. The final text must have the topic <<topic>>.
11× Improve this text by rewriting it. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose. The final text must have the format <<format>>.
9× Edit this text by removing all adverbs, making only minimal adjustments to the surrounding grammar to ensure the sentences remain structurally sound. The final text must have the topic <<topic>>.
9× Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else. The final text must have the format <<format>>.
6× Edit this text by replacing, removing or inserting exactly one to three sentences, without changing anything else. The final text must have the topic <<topic>>.

Topics

Topic	Rows	Share
Finance & Business	1,238	7.2%
Education & Jobs	992	5.7%
Health	940	5.4%
Software Dev.	868	5.0%
Sports & Fitness	853	4.9%
Literature	850	4.9%
Entertainment	829	4.8%
Crime & Law	798	4.6%
Science & Tech.	786	4.5%
Social Life	786	4.5%
Politics	729	4.2%
Software	712	4.1%
Religion	685	4.0%
Hardware	671	3.9%
Transportation	656	3.8%
Home & Hobbies	642	3.7%
Industrial	642	3.7%
Food & Dining	634	3.7%
Games	627	3.6%
Art & Design	612	3.5%
Travel	597	3.5%
History	589	3.4%
Fashion & Beauty	555	3.2%

Formats

Format	Rows	Share
Personal Blog	3,284	19.0%
News Article	2,773	16.0%
Nonfiction Writing	1,479	8.6%
News (Org.)	704	4.1%
About (Pers.)	531	3.1%
Tutorial	531	3.1%
User Review	530	3.1%
Knowledge Article	529	3.1%
Product Page	526	3.0%
Customer Support	525	3.0%
Documentation	524	3.0%
Creative Writing	519	3.0%
FAQ	519	3.0%
Legal Notices	519	3.0%
Academic Writing	514	3.0%
Listicle	512	3.0%
About (Org.)	511	3.0%
Comment Section	497	2.9%
Q&A Forum	485	2.8%
Audio Transcript	458	2.6%
Structured Data	432	2.5%
Content Listing	389	2.2%

Evaluation protocol

Classes

original is Human, final_response is AI. Each detector scores both texts of every row.

Thresholds

All 17,291 rows are used; there is no sweep/evaluation split. Each detector's thresholds are placed exactly on its human scores so that 1% and 0.1% of human texts are called AI. 9 detectors call lower scores AI (for example, perplexity).

Metrics

N (texts scored), AUROC, and TPR at the 1% and 0.1% FPR thresholds. Subsets are scored at the corpus-wide thresholds rather than refitted.

Filters

None. No cosine, soft-ngram, Jaccard, or other analysis-time filtering is applied.

Skipped

None

Config

config/analysis_test_detectors.toml

Schema & usage71 columns · loading snippet⌄

71 columns per row. Detector columns come in pairs, one per text of the row.

Text

original

Human-written source (class: Human)

final_response

Generated text (class: AI)

response_0

Generator output for the first turn

response_1

Generator output for the second turn (empty for single-turn prompts)

Generation

prompt

struct: chat_turns (list of strings), use_multiturn, examples, metadata.PROMPT_TYPE

generator_model

Hugging Face id of the generator model

generation_params

JSON string of sampling parameters

Metadata

topic

Topic of the human source

format

Format of the human source

Distances

jaccard_1, jaccard_2, levenshtein, softngram, cosdist, bertscore, bertscore_precision, bertscore_recall, moverscore, reranker

10 float columns measuring how far the rewrite moved from the original

Other columns

original_editlens_bucket_roberta_large, final_response_editlens_bucket_roberta_large, original_editlens_bucket_llama_3_2_3b, final_response_editlens_bucket_llama_3_2_3b, original_editlens_bucket_giga_roberta, final_response_editlens_bucket_giga_roberta, original_editlens_bucket_giga_llama_3_2_3b, final_response_editlens_bucket_giga_llama_3_2_3b

Not used by this report

Detector scores

One float column per text: {original|final_response}<suffix>

_editlens_score_roberta_large

EditLens Roberta-Large Score

_editlens_score_llama_3_2_3b

EditLens Llama-3.2-3B Score

_editlens_score_giga_roberta

Giga EditLens Roberta Score

_editlens_score_giga_llama_3_2_3b

Giga EditLens Llama-3.2-3B Score

_perplexity_llama_instruct

Perplexity (Llama-3.2-3B-Instruct)

_perplexity_llama_base

Perplexity (Llama-3.2-3B)

_entropy_llama_instruct

Entropy (Llama-3.2-3B-Instruct)

_entropy_llama_base

Entropy (Llama-3.2-3B)

_topp_outlier_llama_instruct

Top-p Outliers (Llama-3.2-3B-Instruct)

_topp_outlier_llama_base

Top-p Outliers (Llama-3.2-3B)

_topk_outlier_llama_instruct

Top-k Outliers (Llama-3.2-3B-Instruct)

_topk_outlier_llama_base

Top-k Outliers (Llama-3.2-3B)

_fastdetectgpt_llama_instruct

FastDetectGPT (Llama-3.2-3B-Instruct)

_fastdetectgpt_llama_base

FastDetectGPT (Llama-3.2-3B)

_binoculars

Binoculars

_detector_score_qwen35_35b_a3b

Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base)

_detector_score_llama_3b

Trained detector: Llama-3.2-3B (QLoRA)

_detector_score_gemma3_27b

Trained detector: Gemma-3-27B (QLoRA, hard labels)

_detector_score_qwen35_35b_a3b_soft

Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai)

_detector_score_llama_3b_soft

Trained detector: Llama-3.2-3B soft labels (soft_ai)

_detector_score_llama_3b_reg

Trained detector: Llama-3.2-3B regression (reg_sqrt)

_detector_score_model_family

Model-family detector: AI probability (Gemma-3-4B-it LoRA)

Load it

from datasets import concatenate_datasets, get_dataset_config_names, load_dataset

repo = "G-reen/fastdetector-test-stat-test"
# every generator config is its own dataset config
ds = concatenate_datasets([
    load_dataset(repo, config, split="train")
    for config in get_dataset_config_names(repo)
])

Univariate statistics54 statistics · all 17,291 rows valid⌄

Every statistic the report does arithmetic on, over the complete corpus. Missing or non-finite values are counted as invalid.

Statistic	Mean	Median	Std	Min	Max
`jaccard_1`	0.5727	0.6652	0.2792	0.0000	1.0000
`jaccard_2`	0.6917	0.8310	0.3067	0.0180	1.0000
`levenshtein`	2129.0771	1543.0000	2507.2039	3.0000	79984.0000
`softngram`	0.5281	0.5812	0.3579	0.0000	1.0000
`cosdist`	0.1562	0.0830	0.1797	-0.0072	0.9648
`bertscore`	0.1170	0.1222	0.0725	0.0012	0.7966
`bertscore_precision`	0.1165	0.1221	0.0714	0.0012	0.7497
`bertscore_recall`	0.1167	0.1187	0.0775	0.0006	0.8287
`moverscore`	0.4836	0.5351	0.1982	0.0595	1.3348
`reranker`	-5.9861	-7.3750	4.0143	-10.9375	17.1250
`EditLens Roberta-Large Score (Human)`	0.0463	0.0204	0.0716	0.0063	0.9844
`EditLens Roberta-Large Score (AI)`	0.3453	0.2358	0.3354	0.0064	0.9996
`EditLens Llama-3.2-3B Score (Human)`	0.0299	0.0159	0.0406	0.0003	0.4905
`EditLens Llama-3.2-3B Score (AI)`	0.3311	0.2596	0.3059	0.0004	1.0000
`Giga EditLens Roberta Score (Human)`	0.0086	0.0038	0.0259	0.0013	0.6720
`Giga EditLens Roberta Score (AI)`	0.3674	0.1953	0.3957	0.0013	1.0000
`Giga EditLens Llama-3.2-3B Score (Human)`	0.0050	0.0026	0.0114	0.0005	0.1835
`Giga EditLens Llama-3.2-3B Score (AI)`	0.3151	0.1526	0.3590	0.0006	1.0000
`Perplexity (Llama-3.2-3B-Instruct) (Human)`	15.5706	13.9608	6.9955	2.9954	65.3062
`Perplexity (Llama-3.2-3B-Instruct) (AI)`	15.2729	13.4184	10.6979	1.0071	537.9031
`Perplexity (Llama-3.2-3B) (Human)`	11.8272	10.8108	4.9435	1.2068	47.0632
`Perplexity (Llama-3.2-3B) (AI)`	12.2759	10.9087	8.6827	1.0034	612.3403
`Entropy (Llama-3.2-3B-Instruct) (Human)`	2.4981	2.4871	0.4349	1.0731	4.2297
`Entropy (Llama-3.2-3B-Instruct) (AI)`	2.4331	2.4211	0.4850	0.0365	7.1878
`Entropy (Llama-3.2-3B) (Human)`	2.3783	2.3792	0.3803	0.2294	3.7963
`Entropy (Llama-3.2-3B) (AI)`	2.4044	2.4040	0.3956	0.0100	5.5116
`Top-p Outliers (Llama-3.2-3B-Instruct) (Human)`	0.0550	0.0537	0.0119	0.0215	0.1231
`Top-p Outliers (Llama-3.2-3B-Instruct) (AI)`	0.0550	0.0535	0.0140	0.0000	0.2083
`Top-p Outliers (Llama-3.2-3B) (Human)`	0.0421	0.0417	0.0094	0.0064	0.0796
`Top-p Outliers (Llama-3.2-3B) (AI)`	0.0395	0.0389	0.0117	0.0000	0.1818
`Top-k Outliers (Llama-3.2-3B-Instruct) (Human)`	0.0960	0.0917	0.0346	0.0064	0.2630
`Top-k Outliers (Llama-3.2-3B-Instruct) (AI)`	0.0944	0.0900	0.0402	0.0001	0.5000
`Top-k Outliers (Llama-3.2-3B) (Human)`	0.0778	0.0739	0.0306	0.0039	0.2192
`Top-k Outliers (Llama-3.2-3B) (AI)`	0.0801	0.0753	0.0367	0.0001	0.5000
`FastDetectGPT (Llama-3.2-3B-Instruct) (Human)`	-1.8356	-1.7960	1.6608	-10.5708	5.9897
`FastDetectGPT (Llama-3.2-3B-Instruct) (AI)`	-1.9421	-1.9255	1.8819	-14.4647	56.7824
`FastDetectGPT (Llama-3.2-3B) (Human)`	-0.1383	-0.0941	1.0285	-5.8795	2.7665
`FastDetectGPT (Llama-3.2-3B) (AI)`	0.1789	0.1079	1.5603	-10.4853	16.7697
`Binoculars (Human)`	0.8469	0.8544	0.0681	0.0379	1.0063
`Binoculars (AI)`	0.8677	0.8679	0.0613	0.0355	1.2666
`Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base) (Human)`	0.1095	0.0518	0.1538	0.0051	0.9999
`Trained detector: Qwen3.5-35B-A3B (QLoRA-style LoRA, bf16 base) (AI)`	0.8800	1.0000	0.2815	0.0084	1.0000
`Trained detector: Llama-3.2-3B (QLoRA) (Human)`	0.0590	0.0328	0.1104	0.0026	0.9998
`Trained detector: Llama-3.2-3B (QLoRA) (AI)`	0.8153	0.9999	0.3475	0.0034	1.0000
`Trained detector: Gemma-3-27B (QLoRA, hard labels) (Human)`	0.0684	0.0354	0.1105	0.0031	0.9553
`Trained detector: Gemma-3-27B (QLoRA, hard labels) (AI)`	0.8525	1.0000	0.3109	0.0042	1.0000
`Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai) (Human)`	0.0379	0.0145	0.0722	0.0014	0.6910
`Trained detector: Qwen3.5-35B-A3B soft labels (soft_ai) (AI)`	0.5805	0.6325	0.2364	0.0019	0.9533
`Trained detector: Llama-3.2-3B soft labels (soft_ai) (Human)`	0.0468	0.0284	0.0620	0.0049	0.5578
`Trained detector: Llama-3.2-3B soft labels (soft_ai) (AI)`	0.5088	0.5699	0.2272	0.0056	0.9137
`Trained detector: Llama-3.2-3B regression (reg_sqrt) (Human)`	0.0117	0.0061	0.0163	-0.0011	0.1865
`Trained detector: Llama-3.2-3B regression (reg_sqrt) (AI)`	0.2820	0.2480	0.2157	-0.0015	0.9102
`Model-family detector: AI probability (Gemma-3-4B-it LoRA) (Human)`	0.0003	0.0002	0.0015	0.0000	0.0592
`Model-family detector: AI probability (Gemma-3-4B-it LoRA) (AI)`	0.4514	0.0326	0.4825	0.0000	1.0000

Correlation heatmapspairwise Pearson r · Spearman ρ⌄

Each pair is correlated over the rows where both statistics are finite. Spearman compares ranks, so it shows whether a detector's score rises with edit distance even when the relationship is not linear.

Pearson r

Open full size ↗

Spearman ρ

Open full size ↗

Model-specific analyticsseparate page · 22 detectors⌄

Per-detector score grids by topic × format and by generator × prompt category, and the ranked specific-prompt tables, are on their own page to keep this card within the Hub's size limit: open ANALYTICS.md ↗.

Generated by the FastDetector analysis pipeline from the complete published corpus.config/analysis_test_detectors.toml

Downloads last month: 1,679