Guide
How to Tell if Text Is AI Generated: 6 Checks
How to tell if text is AI generated without guessing. Six checks for essays, articles and emails, what our own tests show, and how to read a detector score.
Last updated Nov 14, 2026
In 30 seconds: you cannot tell reliably from style alone. Start with context and the writer's drafts, compare with their earlier work, verify facts and citations, and only then run an AI text detector. Treat every result as evidence to discuss, never as proof.
| Check | Best for | Reliability |
|---|---|---|
| 1. Context | Any text | Sets expectations, proves nothing |
| 2. Drafts and version history | Essays, reports | Strong when available |
| 3. Earlier work | Students, regular writers | Good for big changes in voice |
| 4. Writing patterns | Quick first look | Weak on its own |
| 5. Facts and citations | Research and articles | Strong for invented sources |
| 6. A detector | Longer English text | Useful estimate, can be wrong |
1. Start with the context
Before reading for clues, ask what you would expect. Who wrote it, under what conditions, and how much did they know? A polished report from someone who usually writes short notes, or an essay that answers a different question from the one set, is worth a closer look.
Context does not prove anything. It tells you which of the checks below are worth your time.
2. Ask for drafts and version history
The most reliable way to detect AI writing is to see how the text was made. Google Docs and Microsoft Word keep a version history that shows when text appeared and how it changed. Human writing usually grows in many small edits; pasted text arrives in large blocks.
- Ask for earlier drafts, notes or outlines.
- Look for large blocks pasted in one step, then compare them with the rest.
- Ask the writer to explain a paragraph in their own words. People who wrote something can usually talk about why.
A paste is not proof either: people write in other apps and copy their work over. Treat version history as a reason to talk, not a verdict.
3. Compare with the writer's earlier work
If you have other writing from the same person, compare vocabulary, sentence length, typical mistakes and how they structure an argument. A sudden change in all of these at once is a stronger signal than any single sentence.
Be careful with writers whose skills are growing quickly, such as students, and with people writing in a second language. Their work can change a lot between pieces for good reasons.
4. Look at the writing patterns, with caution
Text generated by language models tends to share some habits. None of them proves anything alone, and plenty of careful human writers have them too.
- Even rhythm: sentences and paragraphs of very similar length.
- Generic phrasing: confident statements that could apply to any topic, with few concrete details, names or numbers.
- Neat structure: a summary at the start, balanced lists of three, and a conclusion that restates everything.
- Missing personal detail: no experiences, opinions or local knowledge that only the writer would have.
Clear, formal writing and writing by people using a second language often look like this as well. Use patterns to decide where to look harder, not to decide the answer.
5. Check the facts and citations
Language models can produce sources that do not exist: book titles, journal articles, court cases or quotes that sound right but cannot be found. Detecting ChatGPT-style text is often easiest here.
Look up a few cited sources. Check that they exist, say what the text claims, and are dated before the text was written. An invented source is strong evidence that something went wrong, whether from AI or from careless research, and it is a problem worth raising either way.
6. Run an AI text detector, and read the score
An AI text detector is a model trained on human and AI writing that estimates which one your text resembles more. It works best on longer English prose and least well on short, edited or mixed text.
We tested the model behind our free quick check on texts we collected ourselves:
| Texts | Called AI | Can't tell | Called human |
|---|---|---|---|
| Human writing (1,141) | 0.3% | 0.6% | 99.1% |
| Essays by English learners (879) | 0.1% | 2.8% | 97.0% |
| AI writing (270) | 86.7% | 5.6% | 7.8% |
| AI writing rewritten by a second chatbot (80) | 71.2% | 13.8% | 15.0% |
Two things follow. False alarms on human writing were rare, about 3 in 1,000, but not zero. And the detector missed about 1 in 8 AI texts, more after rewriting. The full method and test sets are on AI detector accuracy and limits.
Our AI text detector runs a free quick check on English text of 150 words or more, without an account. A result in the middle zone is a real answer: the evidence is mixed. The sample text report shows what a full report contains, and how AI detectors work explains the score.
What to do with the result
Put the evidence together before deciding anything. Drafts, earlier work, citations and a detector score that all point the same way make a strong case. A detector score alone does not.
If you are a teacher or editor, talk to the writer first and let them show their process. Never use a detector result alone to accuse, grade, reject or punish someone.
Frequently asked questions
Can you prove text was written by AI? Not with certainty. Version history and invented sources come closest; style and detector scores are estimates.
Does editing AI text make it undetectable? Often it makes detection harder. In our tests, rewritten AI text was called AI 71% of the time, against 87% for unedited AI text.
Are AI detectors unfair to non-native English writers? Plain, formulaic writing can be flagged more often, so treat those results with extra care. In our own test of 879 essays by English learners, 1 was called AI; see do AI detectors flag non-native English writers?