Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
3
3
type
stringclasses
3 values
question
stringlengths
46
242
gold_units
listlengths
0
3
gold_refs
listlengths
0
3
reference
stringlengths
13
470
probe
stringlengths
15
121
stale
dict
s01
single_hop
What is the maximum administrative fine for non-compliance with the prohibited AI practices in Article 5?
[ "art_99" ]
[ "art_99(3)" ]
Up to EUR 35 000 000 or, if the offender is an undertaking, up to 7 % of its total worldwide annual turnover for the preceding financial year, whichever is higher.
exact figures; model must not confuse the 7% tier with the 3% tier
null
s02
single_hop
From what date does the EU AI Act generally apply?
[ "art_113" ]
[ "art_113" ]
It applies from 2 August 2026, subject to the exceptions listed in Article 113.
date recall; distractor dates exist in the same article
null
s03
single_hop
What amount of training compute creates a presumption that a general-purpose AI model has high impact capabilities?
[ "art_51" ]
[ "art_51(2)" ]
A cumulative amount of computation used for training measured in floating point operations greater than 10^25.
numeric threshold that is marked up as a superscript in the source
null
s04
single_hop
How does the Regulation define an 'AI system'?
[ "art_3" ]
[ "art_3(1)" ]
A machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or v...
definition retrieval from a very long definitions article
null
s05
single_hop
Does the Regulation apply to AI systems used exclusively for military, defence or national security purposes?
[ "art_2" ]
[ "art_2(3)" ]
No. Article 2(3) excludes AI systems placed on the market, put into service or used exclusively for military, defence or national security purposes, regardless of the type of entity carrying out those activities.
scope exclusion
null
s06
single_hop
In the general case, within how many days must a provider report a serious incident to the market surveillance authorities?
[ "art_73" ]
[ "art_73(2)" ]
Immediately after establishing a causal link, and in any event not later than 15 days after becoming aware of the serious incident.
one of three competing deadlines in the same article (15 / 10 / 2 days)
null
s07
single_hop
What is the reporting deadline for a serious incident that resulted in the death of a person?
[ "art_73" ]
[ "art_73(4)" ]
Immediately after establishing or suspecting a causal relationship, and not later than 10 days after becoming aware of the incident.
must pick the 10-day rule, not the 15-day default
null
s08
single_hop
What is the reporting deadline in the event of a widespread infringement?
[ "art_73" ]
[ "art_73(3)" ]
Immediately, and not later than 2 days after the provider or deployer becomes aware of the incident.
third competing deadline in the same article
null
s09
single_hop
Is using AI to infer the emotions of employees in the workplace prohibited?
[ "art_5" ]
[ "art_5(1)(f)" ]
Yes. Article 5(1)(f) prohibits AI systems used to infer emotions of a natural person in the areas of workplace and education institutions, except where the system is intended to be put in place or on the market for medical or safety reasons.
prohibition plus its exception; an answer omitting the medical/safety carve-out is incomplete
null
s10
single_hop
What must providers of AI systems that generate synthetic audio, image, video or text ensure about the outputs?
[ "art_50" ]
[ "art_50(2)" ]
That the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust and reliable as far as technically feasible.
operative wording of the marking obligation
null
s11
single_hop
Does the Regulation apply to AI systems developed and put into service for the sole purpose of scientific research and development?
[ "art_2" ]
[ "art_2(6)" ]
No. Article 2(6) excludes AI systems and models, including their output, specifically developed and put into service for the sole purpose of scientific research and development.
scope exclusion
null
s12
single_hop
Is testing in real world conditions covered by the exclusion for research, testing and development activities?
[ "art_2" ]
[ "art_2(8)" ]
No. Article 2(8) excludes research, testing and development activity prior to placing on the market or putting into service, but expressly states that testing in real world conditions is not covered by that exclusion.
exception-to-the-exception; 'close enough' answers get this backwards
null
s13
single_hop
What must providers of general-purpose AI models do in relation to copyright?
[ "art_53" ]
[ "art_53(1)(c)" ]
Put in place a policy to comply with Union law on copyright and related rights, in particular to identify and comply with a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790, including through state-of-the-art technologies.
specific cross-referenced obligation
null
s14
single_hop
Are AI systems that create or expand facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage prohibited?
[ "art_5" ]
[ "art_5(1)(e)" ]
Yes, this is prohibited under Article 5(1)(e).
prohibition recall
null
s15
single_hop
From what date do Chapters I and II of the Regulation apply?
[ "art_113" ]
[ "art_113(a)" ]
From 2 February 2025.
must not return the general 2 August 2026 date
{ "version": "2026-07-27T00:00:00", "status": "incomplete", "note": "Article 113, third paragraph, point (a) was replaced by Regulation (EU) 2026/1744, Article 1, point (40)(a). 2 February 2025 still holds, except Article 5(1) points (ba) and (bb) and Article 5(1a) and (1b), which apply from 2 December 2026." }
s16
single_hop
What is the maximum fine for supplying incorrect, incomplete or misleading information to notified bodies or national competent authorities?
[ "art_99" ]
[ "art_99(5)" ]
Up to EUR 7 500 000 or, if the offender is an undertaking, up to 1 % of total worldwide annual turnover for the preceding financial year, whichever is higher.
lowest penalty tier; easily confused with the 15M/3% tier
null
s17
single_hop
For an SME, is the administrative fine capped at the percentage or at the fixed amount?
[ "art_99" ]
[ "art_99(6)" ]
For SMEs, including start-ups, each fine is up to the percentages or the amount referred to in Article 99(3), (4) and (5), whichever thereof is lower.
trap: the general rule is 'whichever is higher', the SME rule inverts it to 'lower'
null
s18
single_hop
By when must providers of general-purpose AI models placed on the market before 2 August 2025 comply with the Regulation?
[ "art_111" ]
[ "art_111(3)" ]
They must take the necessary steps to comply by 2 August 2027.
transitional deadline
null
s19
single_hop
Is a biometric verification system whose sole purpose is to confirm a person is who they claim to be classified as high-risk under Annex III?
[ "anx_III" ]
[ "anx_III(1)(a)" ]
No. Annex III point 1(a) covers remote biometric identification systems but expressly excludes AI systems intended to be used for biometric verification whose sole purpose is to confirm that a specific natural person is the person he or she claims to be.
explicit carve-out inside an Annex list item
null
s20
single_hop
Which penalty tier applies to a deployer that breaches its obligations under Article 26?
[ "art_99" ]
[ "art_99(4)(e)" ]
Up to EUR 15 000 000 or, if the offender is an undertaking, up to 3 % of total worldwide annual turnover, whichever is higher (Article 99(4)(e)).
mapping a referenced article to the correct fine tier
null
m01
multi_hop
A hospital deploys an emotion recognition system on patients for medical reasons. Is that prohibited, and does it have to tell the patients anything?
[ "art_5", "art_50" ]
[ "art_5(1)(f)", "art_50(3)" ]
It is not prohibited: Article 5(1)(f) exempts emotion inference put in place for medical or safety reasons. However, under Article 50(3) the deployer must inform the natural persons exposed to the system of its operation and process personal data in accordance with the applicable data protection law.
requires combining a prohibition's exception with a separate transparency duty
null
m02
multi_hop
An AI system listed in Annex III performs profiling of natural persons. Can its provider rely on the derogation in Article 6(3) to treat it as not high-risk?
[ "art_6", "anx_III" ]
[ "art_6(3)" ]
No. Article 6(3) allows an Annex III system not to be considered high-risk if it meets one of the listed conditions, but expressly states that notwithstanding that, a system is always high-risk where it performs profiling of natural persons.
the derogation and its override are in different subparagraphs; needs Annex III context
null
m03
multi_hop
A general-purpose AI model was trained with 10^26 floating point operations and is released under a free and open-source licence. Does the open-source exemption in Article 53(2) relieve its provider of the technical documentation obligations?
[ "art_53", "art_51" ]
[ "art_51(2)", "art_53(2)" ]
No. Under Article 51(2) a model trained with more than 10^25 FLOP is presumed to have high impact capabilities and is therefore classified as a general-purpose AI model with systemic risk. Article 53(2) states that the free and open-source exception does not apply to general-purpose AI models with systemic risk, so the...
chains a numeric threshold to a classification to an exemption carve-out
null
m04
multi_hop
For a high-risk AI system that is a safety component of a product covered by the Union harmonisation legislation in Section B of Annex I, which provisions of the Regulation actually apply?
[ "art_2", "art_6" ]
[ "art_2(2)", "art_6(1)" ]
Under Article 2(2), for such systems only Article 6(1), Articles 102 to 109 and Article 112 apply, and Article 57 applies only in so far as the high-risk requirements have been integrated into that Union harmonisation legislation.
scope-limiting rule that overrides the general high-risk regime
{ "version": "2026-07-27T00:00:00", "status": "superseded", "note": "Article 2(2) was replaced by Regulation (EU) 2026/1744, Article 1, point (2)(a). It now reads: only Article 6(1), Article 60a and Articles 102 to 112 apply, and Articles 57, 58 and 59 apply only in so far as the high-risk requirements have been ...
m05
multi_hop
Real-time remote biometric identification in public spaces for law enforcement is restricted by Article 5. Are remote biometric identification systems also classified as high-risk?
[ "art_5", "anx_III" ]
[ "art_5(1)(h)", "anx_III(1)(a)" ]
Yes, both apply. Article 5(1)(h) prohibits the use of real-time remote biometric identification in publicly accessible spaces for law enforcement unless strictly necessary for one of the listed objectives, while Annex III point 1(a) separately classifies remote biometric identification systems as high-risk, in so far a...
the same technology appears in both the prohibition and the high-risk list
null
m06
multi_hop
An AI system is used to filter job applications. Is it high-risk, and what fine would its provider face for breaching Article 16?
[ "anx_III", "art_6", "art_99" ]
[ "anx_III(4)(a)", "art_6(2)", "art_99(4)(a)" ]
Yes. Annex III point 4(a) covers AI systems intended for the recruitment or selection of natural persons, including analysing and filtering job applications, and Article 6(2) makes Annex III systems high-risk. A breach of the provider obligations in Article 16 falls under Article 99(4)(a): up to EUR 15 000 000 or 3 % o...
three-unit chain: Annex classification, high-risk rule, penalty tier
null
m07
multi_hop
When do the general-purpose AI obligations in Chapter V start to apply, and by when must models already on the market comply?
[ "art_113", "art_111" ]
[ "art_113(b)", "art_111(3)" ]
Chapter V applies from 2 August 2025 under Article 113(b). Under Article 111(3), providers of general-purpose AI models placed on the market before that date must take the necessary steps to comply by 2 August 2027.
two different dates in two different articles
null
m08
multi_hop
Is emotion recognition in educational institutions prohibited, and if a permitted emotion recognition system is used elsewhere, is it high-risk?
[ "art_5", "anx_III" ]
[ "art_5(1)(f)", "anx_III(1)(c)" ]
Article 5(1)(f) prohibits emotion inference in the areas of workplace and education institutions except for medical or safety reasons. Where emotion recognition is otherwise permitted, Annex III point 1(c) classifies AI systems intended to be used for emotion recognition as high-risk.
prohibited in one context, high-risk in another
null
s21
single_hop
A provider concludes that its Annex III AI system is not high-risk. What must it do before placing the system on the market, and is it still required to register?
[ "art_6" ]
[ "art_6(4)" ]
Under Article 6(4) the provider must document its assessment before the system is placed on the market or put into service, remains subject to the registration obligation in Article 49(2), and must provide the documentation to national competent authorities on request.
obligation that survives a negative classification
null
m10
multi_hop
A general-purpose AI model with systemic risk is involved in a serious incident. What must its provider do, and under which article?
[ "art_55", "art_73" ]
[ "art_55(1)(c)", "art_73(1)" ]
Under Article 55(1)(c) the provider must keep track of, document and report without undue delay to the AI Office and, as appropriate, national competent authorities, relevant information about serious incidents and possible corrective measures. Article 73 governs serious incident reporting for high-risk AI systems to m...
two parallel incident regimes that are easy to conflate
null
m11
multi_hop
What must a chatbot provider disclose to users, and what fine applies if it fails to?
[ "art_50", "art_99" ]
[ "art_50(1)", "art_99(4)(g)" ]
Under Article 50(1) providers must ensure that AI systems intended to interact directly with natural persons are designed so that those persons are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person. A breach of Article 50 falls under ...
obligation plus its enforcement tier
null
m12
multi_hop
The prohibitions in Article 5 became applicable before the rest of the Regulation. From when, and what is the penalty for breaching them?
[ "art_113", "art_99", "art_5" ]
[ "art_113(a)", "art_99(3)", "art_5(1)" ]
Article 5 sits in Chapter II, which applies from 2 February 2025 under Article 113(a). Breaching the Article 5 prohibitions is subject to administrative fines of up to EUR 35 000 000 or 7 % of total worldwide annual turnover, whichever is higher (Article 99(3)).
requires knowing Article 5 is in Chapter II
{ "version": "2026-07-27T00:00:00", "status": "incomplete", "note": "The two Article 5 prohibitions added by Regulation (EU) 2026/1744 (points (ba) and (bb)) apply from 2 December 2026, not 2 February 2025 (amended Article 113, third paragraph, point (a)). The penalty is unchanged." }
m13
multi_hop
A provider established in the United States places a general-purpose AI model on the Union market. Does the Regulation apply to it, and what documentation must it maintain?
[ "art_2", "art_53" ]
[ "art_2(1)(a)", "art_53(1)(a)", "art_53(1)(b)" ]
Yes. Article 2(1)(a) applies the Regulation to providers placing general-purpose AI models on the market in the Union irrespective of whether they are established in the Union or in a third country. Under Article 53(1)(a) and (b) the provider must draw up and keep up to date technical documentation containing at least ...
extraterritorial scope plus substantive obligation
null
u01
unanswerable
What are the six lawful bases for processing personal data under Article 6 of the GDPR?
[]
[]
NOT_IN_CORPUS
adjacent EU regulation the model almost certainly knows from pretraining
null
u02
unanswerable
How many staff members does the European AI Office employ?
[]
[]
NOT_IN_CORPUS
plausible-sounding institutional fact that the Regulation never states
null
u03
unanswerable
What fee does a notified body charge for a conformity assessment under the AI Act?
[]
[]
NOT_IN_CORPUS
the Act regulates notified bodies but sets no fees
null
u04
unanswerable
Which Member States voted against the AI Act in the Council?
[]
[]
NOT_IN_CORPUS
legislative history, not the legal text
null
u05
unanswerable
What is the maximum fine under the Digital Services Act?
[]
[]
NOT_IN_CORPUS
different instrument; the corpus contains a similar-looking penalties article
null
u06
unanswerable
What minimum accuracy percentage must a high-risk AI system achieve?
[]
[]
NOT_IN_CORPUS
the Act requires an 'appropriate level' of accuracy but sets no number; a hallucinated percentage is the expected failure
null
u07
unanswerable
How many floating point operations were used to train GPT-4?
[]
[]
NOT_IN_CORPUS
the corpus contains a FLOP threshold, inviting a confident but unsupported answer
null
u08
unanswerable
Under the US Executive Order on artificial intelligence, what compute threshold triggers a reporting requirement?
[]
[]
NOT_IN_CORPUS
non-EU instrument with a superficially similar threshold
null
u09
unanswerable
What is the street address of the European AI Office in Brussels?
[]
[]
NOT_IN_CORPUS
specific fact absent from the corpus
null
u10
unanswerable
How many high-risk AI systems have been registered in the EU database so far?
[]
[]
NOT_IN_CORPUS
the Act creates the database but contains no statistics
null
u11
unanswerable
Who was appointed as the first Chair of the European Artificial Intelligence Board?
[]
[]
NOT_IN_CORPUS
the Act establishes the Board but names no individuals
null
u12
unanswerable
Which ISO standard number does the AI Act designate as the harmonised standard for AI risk management?
[]
[]
NOT_IN_CORPUS
the Act refers to harmonised standards generically; no ISO number appears
null

An open law book with question cards rising from its pages

EU AI Act QA

123 questions about the EU AI Act (Regulation (EU) 2024/1689), each with the articles and paragraphs that answer it and a short gold answer. I wrote all of them by hand to test a retrieval-augmented system, eu-ai-act-rag. These are the same three files that repository scores itself on, byte for byte.

It is small on purpose. Every question was written against the text, and a test checks that every gold reference exists in the corpus and every quoted passage sits inside a gold paragraph.

What's in it

config questions single-hop multi-hop unanswerable written against
original 45 21 12 12 the 2024 text as published
amended 11 9 2 0 provisions changed by Regulation (EU) 2026/1744
extended 67 61 6 0 the consolidated text of 27 July 2026

Each config has one split, test. There is no training split, and these questions should not be trained on.

  • Multi-hop questions need two or three articles at once, for example what a provider must do and what the fine is for not doing it.
  • Unanswerable questions are about things the Act doesn't contain (the GDPR's lawful bases, for example). Their gold answer is NOT_IN_CORPUS and a system should refuse them.
  • Amended questions have a different answer after the Digital Omnibus on AI. A system reading the 2024 text gets them wrong while still finding an article with the right number.
from datasets import load_dataset

original = load_dataset("aghasalim/eu-ai-act-qa", "original", split="test")
extended = load_dataset("aghasalim/eu-ai-act-qa", "extended", split="test")

Fields

field in meaning
id all s/m/u single-hop, multi-hop, unanswerable (original); a amended; x extended
type all single_hop, multi_hop or unanswerable
question all the question
gold_units all articles, annexes or recitals that answer it, e.g. art_99, anx_III. Empty for unanswerable
gold_refs all the same at paragraph level, e.g. art_5(1)(f)
reference all a short gold answer, or NOT_IN_CORPUS
probe original, amended what the question is meant to catch
evidence amended, extended the exact wording in the Act the answer rests on
stale original (3 rows) how the 2026 amendments affect this question: incomplete or superseded, with a note
amended_by amended the amending provision
topic extended risk_tiers, prohibited, gpai, timelines, penalties, transparency, amended
new_in_2026 extended true if the answer quotes wording the 2024 text doesn't have (13 of 67)

The ids in gold_units and gold_refs match the chunk files in the repository (data/processed/chunks_<version>.jsonl for whole articles, chunks_<version>_paragraph.jsonl for paragraphs, with versions 2024-07-12 and 2026-07-27). They follow the Act's own numbering, so they can be mapped onto any other copy of the text.

How I use it

Retrieval is scored on the questions that have an answer: hit rate, recall, full recall (every gold article in the top k), MRR and nDCG. One question in original (m04) is superseded by the amendments, so on the 2026 text I use 122 of the 123, of which 110 have an answer.

On those 110, with paragraph chunks of the 2026 text and the top 6 passages:

search full recall multi-hop full recall
hybrid (bge-small + BM25, reciprocal rank fusion) 84.5% 26.3%
+ question splitting 90.0% 57.9%
+ question splitting, fine-tuned embedder 92.7% 63.2%

The scripts that produce these numbers, and the answer-quality evaluation with an LLM judge, are in the repository. make experiments reproduces this table with no API key.

What to keep in mind

  • One person wrote and checked every question. Nobody else has reviewed them.
  • It is small. On 110 questions a gap of two questions is not a result, so compare systems question by question (the repository uses an exact paired sign test).
  • 19 multi-hop questions on the 2026 text is very few.
  • The questions are public, here and on GitHub, so a model may have seen them.
  • English only, and only this one regulation.
  • Not legal advice.

Source and license

The questions, gold answers and labels are mine, under the MIT license. Quoted wording comes from the Act as published on EUR-Lex (CELEX 32024R1689, and 02024R1689-20260727 as amended), © European Union, which may be reused with attribution.

The banner image is AI-generated.

Citation

@software{mustafazada_eu_ai_act_rag,
  author = {Mustafazada, Aghasalim},
  title  = {EU AI Act RAG, with an evaluation I actually ran},
  doi    = {10.5281/zenodo.23003624},
  url    = {https://github.com/aghasalim/eu-ai-act-rag}
}
Downloads last month
-