Wikipedia:WikiProject AI Cleanup/Guide

This is a guide to finding and fixing AI-generated content on Wikipedia.

Spotting AI

This guide is primarily intended for finding articles with AI tells, not finding AI tells in articles you already know about.

What to look for when looking for AI-generated content and how to interpret what you are seeing when getting a feeling that you might have encountered it is not such a simple topic. Do not put too much trust in your own subjective interpretation. Research shows that heavy users of LLMs can correctly determine whether an article was generated by AI about 90% of the time, which means that if you are an expert user of LLMs and you tag 10 pages as being AI-generated, you've probably falsely accused one editor.^[1] People who don't use LLMs much do only slightly better than random chance (in both directions).^[1]

Knowing what to look for

Your first step is Wikipedia:Signs of AI writing, a field guide to help find and detect undisclosed AI-generated content on Wikipedia, backed by real examples (and in some cases by reliable sources). It's strongly encouraged to read the whole thing, but if you want to read only the highest-signal entries, read the two "undue emphasis" and "superficial analyses" section under Content, the "AI vocabulary" section under Language, the first two "communication intended for the user" sections, the "Markdown" section, and the edit summaries section.

Finding the articles

A straightforward way to find AI text in articles is to search for the stuff above. There are two basic ways to use Wikipedia search for this:

Searching in quotation marks, e.g., "stands as a testament" or "fostering" "pivotal". To decrease obvious false positives, search for multiple words or phrases at the same time, preferably those from similar "AI eras."
Searching in regex. Regex should not be used indiscriminately, but it can preserve punctuation, capitalization, etc. typical of AI, and search for multiple kinds of words at once. For instance, the search insource:/[a-z]\, [a-z]+ing its/ "underscoring" picks up one very common "superficial analysis" pattern from 2024-25.

Finding the diff

Wikipedia has been around for over 25 years; before you identify text as possibly AI-generated, find the diff (which may be the first edit) and make sure it was added in 2023 or later. If it wasn't, it almost definitely was not AI. When searching based on the WP:AISIGNS list, this will be the case startlingly often. You must do this for any suspected text. These diffs should be referenced and/or linked to in the reason= field of the template, the edit summary, and/or the talk page.

Some signs are more indicative of newer LLMs than older ones, which can help speed up your search. For instance, text that overuses delve and stands as a testament was likely to be added in 2023 or early 2024, so check there first. The subsequent edit history may also give you some clues. It's common, especially for older AI edits when the problem was not known to as many people, for editors to repeatedly fix hallucinations or source-to-text integrity issues in someone else's AI text, without knowing why they came about. The original diffs might also reveal even stronger signs that text is AI that were since reviewed, such as unreviewed markdown formatting since fixed, or even commentary by the chatbot.

LLM detection software

Shortcuts

Many AI detectors, such as GPTZero, are currently on the market. However, they often suffer unacceptably high false positive rates, particularly free ones,^[2] and should never be used as the sole evidence when accusing someone of AI usage.

These tools can suffer from both false positives and false negatives, and have a lack of transparency in how an "AI" score is calculated. Many sites now offer ways to "humanize" LLM text to attempt to fool these tools. Some human writers may have styles that naturally trigger the tool, and others may use writing plugins that improve their writing but also trigger the tool. The algorithms for such tools also may change, and are usually only free for a small amount of text, or will hide the results behind a paywall modal. However, such tools have improved over time and the best ones (usually non-free) have >95% accuracy rates,^[1] and they can help check dozens of suspected edits for further inspection. A consistently high AI score could warrant additional investigation. Mixed scores are lower signal, since other editors may have changed the content substantially, and wiki formatting/markup can throw some tools off.

Cleaning up

Articles with AI-generated content

The category Category:Articles containing suspected AI-generated texts contains all articles tagged with the {{AI-generated}} template.

When an article cites a reference that does not exist, remove the citation and the content that it is cited for.
Remove citations of questionable sources, and either replace them with reliable sources or remove the content they are cited for.
Ensure that the article text accurately summarizes the remaining cited sources in an encyclopedic tone.
If an entire article (or draft) is obviously LLM-generated with no plausible human review, and the page is not worth keeping (e.g. if the topic is not notable), nominate it for speedy deletion under the G15 criterion. If the page is worth keeping, rewrite it or stubify it to remove LLM-generated content.

See the Signs of AI writing page for tips on identifying LLM-generated text. The Unreliable/Predatory Source Detector (UPSD) user script highlights citation links that contain a specific piece of text added by some AI chatbots.

Talk page discussions with LLM-generated messages

It is inappropriate to post LLM-generated messages in talk page discussions, especially without disclosing that they are LLM-generated.

Use {{cait}} and {{caib}} to collapse discussions that are disruptive due to the use of LLM-generated text.
Ask the editor who posted the LLM-generated message to express their argument in their own words without using an LLM.
If the editor continues to post LLM-generated comments after being asked to stop using an LLM, report them to the incidents noticeboard.

Discussion comments also show signs of AI writing. Do not solely rely on AI content detection tools (such as GPTZero) to determine whether a message is LLM-generated, as these tools have high error rates.

Sources with AI-generated material

Sources produced by machine learning are considered unreliable, and should not be cited as sources in articles. The category All articles containing suspected AI-generated sources contains all articles tagged with the {{AI-generated source}} template.

If a non-notable website that only contains AI-generated content is repeatedly being cited or linked to in articles, request for the website to be placed on the spam blacklist.
Remove citations of AI-generated sources (and the content they are cited for) or tag them with the {{rs}} inline cleanup template.
To inquire about the reliability of a source that incorporates AI-generated content, start a discussion on the reliable sources noticeboard.

Prominent media outlets that have adopted AI-generated content, such as multiple Red Ventures websites, have had their reliability reassessed after noticeboard discussions.

Warning editors

{{Uw-ai1}}, {{Uw-ai2}}, {{Uw-ai3}}, {{Uw-ai4}} – for warning users.
{{Uw-create-ai1}}, {{Uw-create-ai2}} – for new article creations. Use {{Uw-ai3}} for a higher level warning.
{{Db-llm-notice}} – for G15 speedy deletion.

^ ^a ^b ^c "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text". arxiv.org. Retrieved 2025-11-28.
^ Fowler, Geoffrey (August 14, 2023). "What to do when you're accused of AI cheating". The Washington Post. Retrieved October 20, 2023.

[:0-1] "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text". arxiv.org. Retrieved 2025-11-28.

[2] Fowler, Geoffrey (August 14, 2023). "What to do when you're accused of AI cheating". The Washington Post. Retrieved October 20, 2023.

[1]

[2]