Whether it's a blog post, a cover letter or a client report, you can often sense when writing came from a chatbot, but a hunch is hard to act on. This guide turns that hunch into specific things to check and explains how much weight each one deserves.
Key takeaways
- Several weak signs together mean more than one suspicious-looking phrase.
- Missing specifics such as names, numbers and firsthand detail often say more than word choice.
- The type of text matters: a product spec and a personal essay have very different normal styles.
- Sentence-level highlights help you focus a review, but the final call stays human.
Read for what's missing, not just what's there
AI-written text is rarely wrong in an obvious way. More often it's thin. Ask what a person who had actually done the thing would have included. A product review with no mention of how long the reviewer used the product. A cover letter that praises the company without naming a single project. A post on fixing a dripping tap that never names a part. An absence of lived detail is one of the most useful signs you'll find.
Compare two lines from a hotel review. 'The hotel offers a wide range of amenities, and the staff are attentive to every guest's needs.' Then: 'The lift broke on our second day, but the night porter carried our bags up four floors and wouldn't take a tip.' The first could describe any hotel anywhere. The second could only come from someone who stayed there.
Signs at four levels
It helps to look at a passage from close up and then from a distance. Word-level signs are the easiest to spot and also the easiest to change, so give more weight to the bigger-picture ones.
- Words: stock phrases ('plays a vital role', 'a wide range of'), hedges stacked on hedges, and upbeat adjectives that could describe almost anything.
- Sentences: similar length and structure one after another, few fragments or asides, and a habit of grouping things in threes.
- Paragraphs: each one opens with a topic sentence and closes with a mini summary, and the piece ends by restating what it already said.
- The whole piece: an even, pleasant tone, no opinion that might annoy anyone, and no clear sense of who the writer is.
Adjust for the kind of text
What counts as suspicious depends on the genre. Some formats are formulaic by design, so they can look machine-like even when a person wrote every word. Before deciding a piece stands out, compare it with human-written examples of the same genre.
- Formal or templated writing (legal boilerplate, product specs, lab reports) is naturally uniform. Expect weaker signals and be slower to conclude anything.
- Personal writing (reflections, application essays, reviews) should carry a voice and specifics. Generic text here is a stronger sign.
- Short text such as a two-line bio or a social post gives too little to judge reliably, by eye or by tool.
- Second-language writing can read as stiff or overly correct for reasons that have nothing to do with AI.
Checking a passage with Winston AI
Paste the text into the AI detector. You get an overall score plus a highlight on each sentence: red for likely AI, yellow for mixed signals, and no highlight for text that reads as human. Yellow often marks human writing that a tool has polished or paraphrased, so read those sentences yourself before deciding anything. If the highlights land on exactly the sentences that bothered you, that's useful confirmation. If they don't, take that seriously too.
Winston AI works in 14 languages and detects the language automatically. It reports whether text is likely machine-written, not which model produced it. If you're unsure how to read the number, see what an AI detection score means.
Deciding what to do with what you find
- Note the specific sentences and signs that concern you.
- Compare them with the writer's other work, if you have any.
- Ask for something only the real author would have: drafts, notes, sources, or a short conversation about the choices in the text.
- Match your response to the stakes. A blog draft might simply need rewriting with real examples, while a graded or contracted piece calls for a fair, documented process.
If you review submissions for a living, our page for editors shows how this fits into a normal editing routine.