Short answer: when people and AI chat models answer the same prompts, the clearest differences are habits of the whole answer, not single words. In WriteVoice research, AI answers tend to be broad, signposted and wrapped up: general advice in neat sections, transitions like "first" and "additionally", and a closing summary. People tend to be specific, draw on their own experience and take a side. On several of these habits the gap is about one standard deviation. Even so, the habits vary between models and change between versions, and plenty of people write this way too, so no single sign proves a text is AI-written. See the research summary.
Two kinds of tells
People use "AI tells" for two different things:
- Stylistic and content habits you can spot by reading: inflated significance, stock vocabulary, "not just X, but Y" constructions, triplets, a helpful-assistant tone. Wikipedia's editors keep the most thorough public catalogue of these (Wikipedia: Signs of AI writing).
- Habits of the whole answer you notice across many texts: how specific the writing is, whether it draws on experience, whether it takes a side, how it is structured and how it ends. This is what our research looked at, comparing people and AI answering the same prompts.
The first kind is useful for editing. The second kind explains why AI text "feels" off even when no single phrase gives it away.
The familiar stylistic tells
These categories come from Wikipedia's "Signs of AI writing" page, written by editors of WikiProject AI Cleanup:
- Undue emphasis on significance and legacy: claims that a topic "plays a pivotal role" or "has shaped discussions about" something larger.
- High density of "AI vocabulary": a recognisable set of favourite words, used far more often than people use them.
- Avoiding plain "is" and "are": "serves as", "stands as", "represents" in place of a simple copula.
- Negative parallelisms: "not just X, but also Y", "not X, but Y".
- The rule of three: lists of three adjectives or three clauses, over and over.
- Vague attributions: "experts say", "it is widely regarded" without a source.
- Collaborative communication left in: "I hope this helps", "let me know if you'd like...".
The same page is clear about the limits: it is advice, not policy, no single sign is proof, and it warns against relying on AI-detection tools because of their error rates.
The habits our research found
When we compared people and current AI chat models answering the same prompts, five habits separated them most reliably. They are the five WriteVoice dials:
| Habit | What default AI tends to do | What people tend to do |
|---|---|---|
| Specificity | General truths and advice that would read the same anywhere | Pin claims to a when, a where, a who and a real number |
| Lived experience | Argue in general terms, never "when I..." | Make the point through something that happened to them |
| Stance | Weigh both sides: "it depends", pros and cons | Take a side, in their own words |
| Signposting | Announce structure: "first", "additionally", neat sets of points | Let one thought lead into the next |
| Endings | Close with a recap or takeaway | Stop when the point is made |
Two things matter most for anyone hunting for tells:
- The gaps are large, but they are tendencies. On lived experience, stance and signposting, typical AI answers sit about one standard deviation from typical human answers. Many individual people still sit closer to AI than the average does.
- Models differ, and they change. Most chat models lean general, while some recent Claude models are as specific as people, or more; the newest models signpost more than older ones. A list of tells describes particular models at a particular time.
Why AI detectors and tell-lists go stale
A tell is a habit of a particular model at a particular time. When the model changes, so do the habits: in our research, how specific and how signposted AI answers are depends on the model family and the version. A detector or checklist trained on last year's models is measuring last year's habits.
Human writing varies too. Plenty of people signpost, weigh both sides and wrap up with a summary, which is one reason detectors misjudge real people. WriteVoice shows where you sit on those habits next to default AI, as a description, not a verdict (see Direct to Signposted and Open-ended to Wrapped-up).
What to do with this
- If you edit AI drafts: cut stock vocabulary, swap "serves as" for "is", break up triplets, delete the closing recap, and put back the specifics (the date, the place, the number).
- If you want AI to sound like you: know your own habits, give the model your real details, and correct drafts with specific feedback; or train your double in WriteVoice. See how to make AI write like you.
- If you are judging someone else's text: treat tells as a reason to ask, not as proof.
Frequently asked questions
What are the most common tells of AI writing?
Readable tells include inflated significance, a recognisable "AI vocabulary", "serves as" instead of "is", "not just X, but Y", triplets and vague attributions. In our research, the habits that separate AI answers from people's most reliably are being general rather than specific, leaving out lived experience, weighing both sides, signposting and ending with a recap.
Can you reliably tell if text was written by AI?
Not from one sign. Tells are model-specific, change between versions, and also appear in human writing. Wikipedia's editors explicitly warn against relying on AI detectors because of their error rates.
Do different AI models have different tells?
Yes. In our research, most chat models lean general while some recent Claude models are as specific as people, and how much models signpost changes between versions. The shared habits (balance, signposting, a closing recap) are common across model families.
Is AI writing getting more human-like?
Not in a simple way. Some habits move toward people (some recent models are more specific), while others move away (the newest models signpost more, not less). "Different from human" changes shape between versions rather than steadily shrinking.