False accusations of AI use can drive away new editors and foster an atmosphere of suspicion. Before claiming AI was used, consider whether Dunning–Kruger effect and confirmation bias is clouding your judgement. Here are several somewhat commonly used indicators that are ineffective in LLM detection—and may even indicate the opposite.
Perfect grammar: While modern LLMs are known for high grammatical proficiency, many editors are also skilled writers or come from professional writing backgrounds. (See also § Sudden shift in English variety use.)
Combination of casual and formal registers, or language that sounds both "clinical" and "emotional": This may indicate the casual writing of a person in a technical field, such as computer science. It may also indicate youth, a preference for mixed registers, playfulness, or neurodivergence. In the case of a wiki, it may simply be the result of multiple editors adding to a page.
"Bland" or "robotic" prose: LLM output has specific traits, as detailed above, and it skews positive and verbose by default. While these tendencies are formulaic, they may not scan as "robotic" to those unfamiliar with AI writing.
"Fancy", "academic", or "formal" prose: While LLMs disproportionately favor certain words and phrases, many of which are longer and have more difficult readability scores than some of their synonyms, these are specific words. The correlation does not extend to all formal, academic, or "fancy"-sounding prose.
Letter-like writing (in isolation): Although many talk page messages written with salutations, valedictions, subject lines, and other formalities after 2023 tend to appear AI-generated, letters and emails have conventionally been written in such ways long before modern LLMs existed. Human editors (particularly newer editors) may format their talk page comments similarly for various reasons, such as being more accustomed to formal communication, posting as part of a school assignment that requires this tone, or simply mistaking the talk page for email. Other tells, such as vertical lists, placeholders, or abrupt cutoffs, are stronger.
Transition words (in isolation): Older AI text tended to formulaically overuse certain transitions like Additionally, Consequently, and Notably, often to begin sentences. However, only a few transition words and phrases are known to be overused by AI in this way. This pattern also has precedence in essay-like writing by humans and is accepted by many style guides, so this is not a strong tell.
Unsourced content: More than 570,000 articles are tagged as needing citations, and most of them predate LLMs. Meanwhile, since modern LLM chatbots can search the web and view sources a user provides to it, citations are fairly common now in AI-generated text. This does not mean they are accurate citations, but they are there.