User:Athanelar/Identifying AI-generated text

The most obvious is if the user says they generated their article/edit using AI. This is likely to come in the form of a statement like "I used AI, but I manually checked the veracity/sources/prose myself." This naturally settles any question as to the text's origin.

Some large language models leave behind telltale signs that may be missed by a user who does not manually check the output. These include communication from the model intended for the model user, fabricated references (not merely dead links or ISBNs which are wrong by a single number), and the UTM parameter indicating a citation copied from an LLM. Fingerprints like these are essentially the same level of confirmation and reliability as direct user confirmation, but it's important to note that some of these may not indicate a wholly AI-generated article in a way that violates WP:NEWLLM. For instance, a fabricated source or the presence of a UTM indicator might show that someone used an LLM to search for sources but not to generate the rest of the text. Communication intended for the user of course would indicate AI-generated text.

There is an extensive list of signs of AI usage based on writing quirks, formatting errors and so on. Any one of these in isolation is not necessarily confirmation, but the greater the number of them present, especially in a small amount of text, the higher the likelihood that text is AI-generated. If two or more of these signs are present in an edit made in 2023 or later, then the chances are high that the text is AI-generated, only increasing the more of them there are. Judging based solely on these requires individual discretion and some prior experience in identifying AI-generated text. One should always be wary of false positives and remember to always assume good faith.

AI detection software is not 100% reliable and should not be used as the only evidence of AI generated text.^[1][a] The referenced source shows that these programs are particularly ineffective when dealing with text that has been human-modified in certain ways rather than being exclusively AI-generated. At best, these detectors should be used to confirm an already-substantiated suspicion by an editor experienced in identifying AI-generated text, rather than being used by an inexperienced editor as a primary indicator.

If AI generation is suspected but not confirmed (due to mixed AI and human writing, uncertain AI signs, lack of direct confirmation etc) the best course of action is to ask the user in question to explain their writing decisions. You can do this through a progressive escalation;

First, plainly ask the user if they used any AI assistance when creating their article or making their edit. Assume good faith, and point out the problem that is leading you to ask the question. If they have submitted a draft article which appears to be primarily AI generated, link them to WP:NEWLLM and explain that articles should not be created from scratch using AI. If they have made an edit including a fabricated citation, link them to WP:LLM to explain some of the problems that arise from using AI to edit Wikipedia. A good faith editor may well own up at this point, whereupon you have the confirmation necessary to decline the draft/delete the article/revert the edit as necessary.

If the editor denies using AI assistance, then question their decisionmaking process more specifically, but still politely. "Where did you find the ISBN that you used for this reference?" "This information doesn't appear in the source you've cited, did you mean to cite another source?" A user who has generated their text with AI will likely be unable to satisfactorily explain certain decisions, especially things like source-to-text discrepancies and fabricated ISBNs. LLM users are also often new editors, so be extra mindful of WP:BITE here.

Ultimately, when dealing with an uncooperative editor who vehemently denies that the text is AI generated, unless there is 'smoking gun' confirmation from a telltale sign like user-directed communication, then the decision to revert or remove content based on WP:NEWLLM must fall to good judgement and consensus.

If, after asking the editor, they are unable to provide a satisfactory explanation and AI usage remains likely and suspected, then additional opinions could be solicited somewhere like the AI cleanup noticeboard, or if the question is whether to delete an article based on WP:NEWLLM, then the article can be nominated to AfD to allow uninvolved editors there to decide based on the available evidence whether the article is AI generated.