Datasets:
Access AI Slop Reports
This dataset contains public posts from X that volunteers flagged as AI-generated "slop", including the authors' handles. Access is granted automatically once you accept the terms.
By requesting access you agree to: (1) use this dataset only for research, education or building detection tools; (2) not use it to harass, target, shame or run campaigns against any author in it; (3) not try to identify people beyond what the data already shows; (4) remember that every label is one volunteer's opinion, not proof that a post was written by AI.
Log in or Sign Up to review the conditions and access this dataset content.
AI Slop Reports
Posts on X that people flagged as AI slop: text that reads as obviously machine-written, such as generic replies, "it's not X, it's Y", the em-dash pile-up and the hollow "honestly?" opener.
Reports come from the Slop Reporter browser extension. Volunteers click a button on a post, can add tags explaining why, and the post lands here. The dataset updates every 10 minutes.
Configs
| config | one row per | use it for |
|---|---|---|
posts (default) |
reported post | training and evaluating detectors |
reports |
(reporter, post) pair | agreement between reporters, reporter quality, tag analysis |
accounts |
reported account | studying bot accounts; joining to posts by author_handle |
from datasets import load_dataset
posts = load_dataset("slop-detection/ai-slop-reports", "posts", split="train")
reports = load_dataset("slop-detection/ai-slop-reports", "reports", split="train")
accounts = load_dataset("slop-detection/ai-slop-reports", "accounts", split="train")
Fields
Post fields (both configs)
| field | description |
|---|---|
platform |
Always x for now. |
post_id, url |
The post's ID and link. |
author_handle, author_display_name, author_id, author_is_blue_verified |
The post's author. |
posted_at |
When the post was published (UTC). |
text |
The post text. Reply @mentions are removed and links expanded. |
text_verified |
full: the text was fetched from X by our server. prefix: a long post whose start matched X's version and whose remainder was read from the page. none: X returned nothing, so the text is what the reporter's browser saw. |
text_complete |
false when only a truncated version of a long post was available. |
lang |
Language of the post. |
is_reply, parent_post_id, parent_author_handle, parent_text |
For replies, the post being replied to. Many slop posts are replies, so the parent gives useful context. |
is_quote, quoted_post_id, quoted_text |
For quote posts, the quoted post. |
has_media |
Whether the post has images, video or a link card. Media files are not included. |
posts only: report_count (split into post_report_count and account_report_count), tags, tag_counts (a JSON object), first_reported_at, last_reported_at.
reports only: report_id, schema_version, label_source, tags, note, submitter (HF username), submitter_hf_id, reported_at, extension_version.
accounts: platform, handle, author_id, report_count, posts_collected, tags, tag_counts, first_reported_at, last_reported_at.
label_source: post means a volunteer judged that specific post. account means the volunteer reported the whole account as a slop bot, and the post was collected from its profile. Account-level labels are weaker: bot accounts sometimes post normal text. Filter on post_report_count > 0 for the cleanest labels.
Tags: em-dash, not-x-but-y, rule-of-three, hype-opener, emoji-bullets, generic-reply, engagement-bait, delve-vocab, fake-personal-anecdote, other.
Caveats
- Positive examples only. There are no "human-written" negatives, so pair this with a sample of unreported posts to train a classifier.
- Labels are opinions. A report means one volunteer thought the post was AI slop. Use
report_countand agreement between reporters to filter. - Selection bias. Volunteers report what they come across, so the data skews towards English, tech and AI accounts and viral threads.
- Reports can change. Reporters can withdraw their reports, and banned reporters' reports are removed, so the latest version can drop rows.
License
- The annotations (labels, tags, notes, aggregates) are licensed CC-BY-4.0.
- The text of each post belongs to its author and is included for research only, under the terms above.
Removal
If you're an author in this dataset and want your posts removed, open a Discussion on this repo. We'll remove the posts from the dataset and block them from being reported again.
- Downloads last month
- -