Wikipedia:Help, I've been accused of using AI!
 | This is an essay on the WP:NOLLM guideline. It contains the advice or opinions of one or more Wikipedia contributors. This page is not an encyclopedia article or a Wikipedia policy, as it has not been reviewed by the community. |
 | This page in a nutshell: Remain calm, respond without arguing, and be as forthcoming as possible. |
Suppose someone posts a comment on your talk page accusing you of using AI, or tags an article you edited as likely containing AI-generated text.
Take a moment to look over the comment or tags and think from the accuser's perspective. They are pointing to things you may be uniquely equipped to explain. There are many ways to respond, some of which are helpful and some of which will make you look bad.
What to say
[edit]Yes
[edit]If you used AI tools to write or edit the content in question, then say so, preferably with details about the specific AI tool(s) you used (e.g. ChatGPT, Gemini, DeepL, Grammarly, etc), the prompt(s) you used, and the chatbot session if you have it. Most people will be much more willing to engage constructively with you if you tell the truth.
No
[edit]If you didn't use AI, you can deny the accusation. There is literally no reason not to tell the truth.
Consider, however, whether you used AI without being aware of it. Grammarly, for instance, makes calls to OpenAI under the hood for most of its rewriting functionality, as well as its "writing suggestions." Microsoft has integrated Copilot into many of its applications, including Microsoft Word and Notepad. Translating with an AI tool may lead to embellishments in the translated text.
If you didn't use AI, you tell the truth, and the person accuses you of lying about it, then see Wikipedia's guides to resolving disputes and responding to personal attacks.
What not to say
[edit]Dismissals
[edit]"You have no proof!"
[edit]Although an accuser cannot procure definitive evidence that you have used a particular large language model or other writing tool unless you leave things such as oaicite in your text, they can still point to other signs which are rather uncommon in text that humans have generated prior to 2023, including markdown. Of course, those may merely be stylistic indicators, but that doesn't mean you can just dismiss their concerns as being purely speculative.
"You just think it's AI because of the tone and the writing style!"
[edit]Generally, people do not claim that text is AI-generated without a reason—which, ideally, they will have mentioned—and AI-generated Wikipedia text has specific and identiable markers that go beyond simple matters of "style" and rarely appear in older, human-written Wikipedia text.
If you are one of those rare exceptions, you can say so, but note that there's a good chance the reason someone found your edit in the first place was likely because of promotional text, original research or synthesis, misrepresentations of sources, and other issues that are both common in LLM output and against Wikipedia's policies.
Rebuttals
[edit]"Humans can write like this!"
[edit]This doesn't answer the question of whether you, specifically, did use AI.
It's theoretically possible for humans to write like AI; anything is possible. A monkey randomly mashing keys could, statistically speaking, produce the complete works of William Shakespeare. However, in the entire evolutionary history of monkeys, none of them actually did. Similarly, the linguistic characteristics of AI-generated text are things that simply did not appear very often in the many centuries' worth of text that humans actually wrote, and then, almost immediately, started to appear everywhere after 2023.[1]
This holds true even when comparing the same kinds of text, e.g., formal academic writing by humans versus formal academic writing by AI. It even holds true when comparing AI-generated text from the base language models—i.e., based only on the training data humans provided, with no changes after the fact—to text generated by chatbots available to the public. These things especially did not show up en masse, repeatedly and formulaically, in the same piece of writing. There are more than 1 billion edits to Wikipedia, going back more than 25 years, and yet out of those 1 billion edits, very few of them show the linguistic characteristics of AI-generated Wikipedia text that became ubiquitous after 2023. Think of the Fermi paradox: if there are so many people out there writing like AI, then where are they?
Of course, correlation is not causation, and some people are indeed the rare exceptions here. If that exception is you, you can just say so.
"But this is just formal writing!" / "I was trained to write like this!"
[edit]AI-generated "formal writing" also sounds different from human-generated "formal writing." In fact, most of the studies comparing AI- and human-written text specifically compared human academic writing to AI-generated versions of the same kind of writing. There are differences, and those specific differences are not taught in any known curriculum.
Furthermore, although "professional" or "formal" writing can be (and has been done) by humans without any AI assistance, not all such styles are suitable for Wikipedia articles. Wikipedia is an encyclopedia, so articles should be written in an encyclopedic tone and in a way that ordinary readers can understand, without promotional language such as puffery or marketing buzzspeak. You don't have to "refine" your writing with fancy words. As George Orwell wrote, never use a long word where a short word will do.
Using an AI tool such as Grammarly to rewrite your text to sound formal isn't a good idea either; these tools should only be used for basic copyediting, such as fixing spelling and grammatical errors. As a rule of thumb, if someone can tell that you used an AI tool to rewrite your original draft, then that isn't basic copyediting.
"But every comment is based on my own understanding!"
[edit]If your comments look like they were pasted from an LLM, we can't be sure that your comments truly reflect your thoughts accurately and aren't just a machine's guess as to what we would like to hear from you.
Excuses
[edit]"But AI helps me get across what I want to say!"
[edit]Oftentimes, when an AI chatbot is asked to help "organize" its user's thoughts or fix text to ensure that it is grammatically correct, adheres to Wikipedia's style guidelines, and is understandable to readers, it might insert new text that says things that differ from what you may have meant to say when you wrote your original text.
Even if your unassisted comments aren't perfect or well-organized, we still prefer for you to produce content and comments manually so that you can spot errors and other things to improve as you go. If the content you introduce to an article happens to contain a few imperfections, other editors might come along and correct those later. That doesn't give you an excuse to keep introducing errors though, but using AI to fix or "refine" them may cause other problems as well.
If you can't manually produce coherent text without the assistance of a text generation or proofreading tool such as Grammarly, you may lack the competence to contribute to Wikipedia effectively.
If you're considering using an AI tool to help you communicate more clearly with us, please try to communicate with us without AI assistance. If we can understand your unassisted comments, you don't need AI assistance, not even if we don't agree with what we think you're trying to say. AI can't help you win people over, as it would often lace your text with stylistic indicators that other users can spot and criticize you for.
"But English isn't my first language!"
[edit]The carveout for translation, which is further elaborated at WP:LLMTRANSLATE, is designed to permit editors already fluent in both languages to more quickly translate text from one language to the other. The bilingual fluency is key, because an editor using an LLM to translate text is expected to be able to verify that the translation is authentic. It is not designed to act as an aid for editors who are not proficient in English; because if you lack the proficiency to write like that in the first place, you also lack the proficiency to ensure the AI's output is accurate and appropriate.
Perfect English fluency is not required to contribute to Wikipedia. All that is required is "the ability to read and write English well enough to avoid introducing incomprehensible text into articles and to communicate effectively." We would rather communicate with you directly, making your best effort with your own ability, rather than communicate through the proxy of an LLM.
"I'm only using it for basic copyediting, which is allowed!"
[edit]As of June 2026[update], the guideline says the following about basic copyediting, emphasis mine; Editors are permitted to use LLMs to suggest basic copyedits to their own writing, and to incorporate some of them after human review, provided the LLM does not introduce content of its own... Examples of basic copyedits include spelling, punctuation, and capitalization.
Note that the examples given are of cosmetic corrections to text written by the human editor, which is supported by the statement that the LLM must not introduce content of its own. Using standard AI chatbots, or AI-powered correction tools like Grammarly to reword entire sentences to make them more neutral, more encyclopedic, to have better grammar etc is not covered by this exception. Any instance in which the AI introduces new words that were not written by the human editor is not covered by this exception.
"Whether it's AI-generated doesn't matter as long as every claim is verified!"
[edit]On Wikipedia, it does matter. Using AI to write content violates Wikipedia's guidelines, no matter how much the AI chatbot tells you otherwise, or how good the text may appear to you personally. You may read the relevant guideline at WP:Writing articles with large language models, but a brief summary follows.
Except for basic copyediting[a] or translation, you should not use an LLM to write content; doing so is considered disruptive. Articles whose first revisions contain strong indicators of having been entirely AI-generated may be speedily deleted under criterion G15 or converted into a subpage of Wikipedia:Signs of AI writing/Examples, and comments containing strong indicators of being entirely AI-generated may be struck or collapsed via {{collapse AI top}} and {{collapse AI bottom}} per WP:AITALK.
If you try to use a generative AI tool for basic copyediting or translation, you have to make sure that the information in the text you want to add matches what the cited sources say, that all of the citation details are accurate, and that you wouldn't be introducing statements that have been made up by the tool itself if you published your edits. Prompting an AI chatbot to check or modify (i.e. "refine") its own output so that it "sounds human" (at least to AI detectors) or adheres to Wikipedia's policies and guidelines is generally not effective. It is much better to manually write content yourself in your own words.
Depending on which AI chatbot you may use, consulting it for Wikipedia-related advice may be a bad idea, especially if you're using ChatGPT, which is known to frequently misrepresent Wikipedia's policies and guidelines.[b]
Diversions
[edit]"This AI detector says this isn't AI!"
[edit]Even the best AI detection software cannot divine with perfect accuracy whether your writing was generated by AI. However, you can. Saying this just raises the question of why you need to use an app to tell you what you, yourself, did, and why you are hiding behind this secondhand information.
"Revised the article to address concerns about "large language model" tone"
[edit]The "concerns" are about whether you used AI. Therefore, the actual way to "address the concerns" is to say whether you used AI. If you did use AI, then the article requires much more than a superficial edit for "tone."
Comments like these frequently show up to accompany revisions that are themselves done by AI, which just exacerbates the problem. Using AI to "fix" AI does not fix anything.
"Your accusations are uncivil!"
[edit]If someone accuses you of using AI, you should assume good faith and give an honest explanation about your editing. They aren't accusing you of AI use out of malice; they're doing so based on their observations of your contributions. If they have noticed you making mistakes that are unusual for a human to make as often as you have (e.g. using markdown, making up citation details, creating a draft with a decline notice on the very first revision, citing nonexistent shortcuts), then they may be right to suspect that you've been pasting content generated by an LLM.
Accusing those who criticize you of making personal attacks against you is not a good idea. Per Wikipedia's civility policy, doing so is in itself potentially disruptive, and may result in warnings or even blocks if repeated.
"Let's focus on improving these articles instead of making accusations."
[edit]Dismissing an accuser's concerns as merely speculative and telling them to focus solely on improving articles doesn't address the issues you may have introduced with your editing.
Although both humans and LLMs make mistakes, we have a right to be concerned about AI use in particular because LLMs can generate multiple paragraphs in seconds, so if you paste lots of AI-generated text onto Wikipedia, you might leave lots of issues in your wake that other editors would have to spend lots of their time cleaning up, and pasting even more AI-generated text while they're trying to clean up after your earlier messes would make them feel like Sisyphus.
Feigning ignorance
[edit]"What parts of this sound like AI?"
[edit]This doesn't answer the question of whether you used AI; whether or not this is your intention, it makes you sound like you are mostly concerned with covering your tracks.
If someone suspects that some of your contributions are AI-generated, asking them which specific passages gave them that idea so you can happily fix them doesn't properly address their concerns. AI-generated content on Wikipedia tends to contain other issues besides tone or style, such as hallucinated citations or other mistakes we wouldn't expect a human to unintentionally make. It's better to write content manually so you can better notice and fix problems as you go. That way, you can save us the trouble of cleaning up after you or telling you what needs to be fixed.
Other
[edit]"If there is any AI ..."
[edit]Making statements that start with something along the lines of "If I used AI..." makes you sound like OJ Simpson. Either AI was used or it wasn't. If you created a draft or added content that appears AI-generated, you should be able to recall whether or not the content you introduced actually came from an AI language model (and if so, which one) or was "refined" by a grammar checker, word processor, or other tool you used that has such a model integrated into it.
"AI detectors don't work!"
[edit]While AI detection software is not perfect, it is much more reliable than laypeople think; the best ones are accurate more than 99% of the time.[2] When they do get it wrong, it's more likely to be a false negative—[3]i.e., claiming that a piece of writing was written by a human instead of AI, not vice versa.
More to the point, though, this is only relevant if someone actually used one of those detectors. Many people, including the creator of this essay, don't.
"But can you remove the tag?"
[edit]If you used AI in producing an article, then the template stating that the article contains AI-generated content is a factual statement, and one that readers—many of whom donate money to Wikipedia specifically because they want to read writing without AI—deserve to know about. If you don't like readers knowing that you used AI, then don't use AI, and then ask yourself why you are using a tool that you are ashamed of.
Any AI-generated response
[edit]There are established guidelines on Wikipedia that strongly discourage AI-generated comments on talk pages. This is because people don't want to hear from an AI chatbot. They want to hear from you.
It's also generally a bad idea to try to hide your AI use by using more AI. Most chatbots have a strong and recognizable "speaking pattern"; just as people can recognize the voice of Gilbert Gottfried, Fran Drescher, or Miss Piggy, people can recognize the voice of ChatGPT, Claude, Gemini, and other large language models. Chatbots also tend to produce many of the poor responses detailed above, along with other things.
If you are accusing someone of AI usage
[edit]If you are accusing someone of AI usage or asking them if they used AI to write the content you came across, make sure to provide evidence for a specific diff or section; try to link specific edits to known signs of AI-use where possible, so that other editors can understand your concerns. Be aware that asking if an edit is AI may inherently be seen as an insult or attack, even when it is not, and provide relevant guidelines and information around AI usage.
If you point out specific edits and they dismiss your accusations as "speculation" based on "stylistic indications" and your "subjective interpretation" of their writing style, and they demand even stronger evidence that they have used AI, asking them to disclose evidence that they used AI may not get you far with them. Instead, consider asking them if the content they added was produced with the aid of any writing assistance technology, and if so, which specific tools they have used and how they have used them. If they admit to using AI in a way that violates Wikipedia's LLM guidelines, you can remind them that what they have been doing violates Wikipedia's LLM guidelines.
Even in clear-cut cases of AI usage, try to provide empathy especially if an editor is inexperienced. Many editors may feel insecure about their writing skills. Assure the editor that their own words are far more valuable than AI-polished words.
If an article you've been editing is flagged for AI
[edit]First, don't panic. Wikipedia has lots of editors, and the person flagging the article is not necessarily talking about you. Ideally, the person flagging an article will note whose edit or diff they found suspicious, and then you can ask yourself:
- If that edit is by you, or if no specific diff was mentioned but you know you used AI for something in the article, then see the "What to say" section.
- If that edit is not by you, assume good faith; most people doing AI cleanup are trying their hardest to improve the encyclopedia. They may not be experts on the article subject, but they generally do have a good sense of what AI-generated text on Wikipedia looks like, based on reading thousands of instances of it across hundreds of different subject areas. Look at the text in question—which likely has some issues if it came to an editor's attention—and compare it to the signs listed in WP:AISIGNS, especially anything in the "Words to watch" boxes. Look for any other discussions of the user's use of AI, on their talk page, the AI noticeboard, or even ANI. Don't just remove the tag, however, without strong evidence that the text isn't AI (for instance, if the edit predates November 2022, the tag is most likely wrong).
Notes
[edit]- ^ Basic copyediting refers to making corrections to text to fix spelling, punctuation and capitalization errors.
- ^ See Wikipedia talk:WikiProject AI Cleanup/Archive 8 § What do LLMs actually do when asked to write a Wikipedia article?
References
[edit]- ^ Juzek, Tom S.; Ward, Zina B (19 January 2026). "Why does ChatGPT 'Delve' so much? Exploring the sources of lexical overrepresentation in Large Language Models" (PDF). Proceedings of the 31st International Conference on Computational Linguistics. Association for Computational Linguistics. pp. 6397–6411. arXiv:2412.11385. Archived (PDF) from the original on 12 March 2026. Retrieved 16 March 2026.
- ^ Russell, Jenna; Karpinska, Marzena; Iyyer, Mohit (July 2025). "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text". Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. pp. 5342–5373. arXiv:2501.15654v2. doi:10.18653/v1/2025.acl-long.267. ISBN 979-8-89176-251-0. Archived from the original on 11 March 2026. Retrieved 16 March 2026.
- ^ Robinson, Matt (2 December 2025). "Do AI Detectors Work Well Enough to Trust?". Chicago Booth Review. University of Chicago Booth School of Business. Archived from the original on 4 May 2026. Retrieved 28 May 2026.
All three commercial tools kept false positive rates below 1 percent, with Pangram's the lowest—essentially 0 across most decision thresholds. False negative rates were higher, coming in between roughly 0 percent and 2 percent for GPTZero and between 2 percent and 4 percent for Pangram. Originality.ai's false negatives were higher still: between 10 percent and 40 percent, depending on the model.