Is this text AI-generated? What can be verified, and what can only be estimated
You cannot verify from the text alone whether AI wrote it. Plain text carries no signature or record of its origin, so AI text detectors can only estimate, and published tests show real error rates. Evidence comes from the writing process instead, such as dated drafts, version history and hashes of each draft.
A file can be checked against a published hash, and an image can carry a signed content credential. A paragraph pasted into an email or a learning platform carries neither. That is the core problem with asking whether text is AI-generated: there is nothing inside plain text that records where it came from.
This page sets out what can be verified about a piece of writing and what can only be estimated. It sits in our wider set of guides to verifying files, images and text, which rank evidence from strongest to weakest. For text, the strong evidence lives outside the words, in how they were written.
Why text has no hash of origin
A hash is a fingerprint of exact bytes. Change one character and the hash changes completely, so a hash can show that a document is unchanged since a known point. It says nothing about who or what typed it.
Signatures and content credentials add a name to the hash, but someone has to apply them when the file is made, and almost no writing tool does that for ordinary text. Copying text into a new box, a new file or a new format leaves no trace of the old one behind. The only text-level origin signal in use today is a provider watermark, covered below, and it only exists if the provider added it.
What AI text detectors measure
A detector is a classifier. It looks at statistical features of the text, such as how predictable each word is given the words before it, and compares them with examples of human and machine writing it was trained on. The output is a probability, and it can be wrong in both directions.
The clearest public evidence comes from a detector maker that measured its own tool. OpenAI released a classifier in January 2023 and reported that, on its challenge set of English texts, it correctly identified 26% of AI-written text as likely AI-written and labelled human-written text as AI-written 9% of the time. OpenAI also said the classifier was very unreliable on texts below 1,000 characters.
OpenAI withdrew its AI text classifier on July 20, 2023, citing "its low rate of accuracy". The notice is still on the original OpenAI announcement page, above the figures it published.
Who gets flagged by mistake
Error rates are not spread evenly across writers. In a 2023 study, Liang, Yuksekgonul, Mao, Wu and Zou at Stanford ran seven widely used GPT detectors on 91 TOEFL essays written by non-native English speakers.
Seven detectors flagged 91 human-written TOEFL essays with an average false positive rate of 61.22% (Liang et al., 2023). Of the 91 essays, 18 (19.78%) were flagged as AI-generated by all seven detectors, and 89 (97.80%) were flagged by at least one.
The same study then had ChatGPT enrich the word choices in those essays. With the vocabulary made more varied, the average false positive rate fell to 11.77%. In other words, the detectors were reacting to plain, predictable phrasing, which is common in careful writing by people working in a second language.
These figures describe those detectors on those essays in 2023. Detectors have changed since then, and so have the models they try to catch. The lesson that carries over is about method: a detector's error rate depends on who wrote the test texts, and a score on one person's essay is a single draw from that error rate.
The evidence ladder for text
Ranked from strongest to weakest, these are the kinds of evidence available for a piece of writing, and what each can show.
| Rank | Evidence | What it can show | What it cannot show |
|---|---|---|---|
| 1 | Third-party timestamps on drafts | A specific draft existed by a specific time | Who typed that draft |
| 2 | Version history kept by a service or in Git | How the text grew, edit by edit, with dates | That no text was pasted in from elsewhere |
| 3 | Your own hash log of drafts | That each logged draft is unchanged | When the log was made, unless it is timestamped |
| 4 | A provider watermark | That one provider's model generated the text | Anything about other providers |
| 5 | Notes, sources, outlines | The research and thinking behind the text | Authorship on their own |
| 6 | A detector score | A probability, with an error rate | Origin |
Version history is strong because it is a record of process. Google Docs, Microsoft 365 and similar services keep revisions on their own servers, and Git records every committed version under a content hash. The Pro Git book explains that each Git object is named by a 40-character SHA-1 checksum of its content, so a committed draft cannot be altered without its name changing.
Watermarks are narrower. Google DeepMind says SynthID can watermark text by adjusting the probability scores of the words a model picks, and it describes this for text from Gemini. A positive check is useful evidence about Gemini output. A negative check says nothing about text from any other model, and our page on how watermarks differ from hashes and signatures explains why that limit applies to every watermark.
What does count as evidence
If you write for a living, for a class or for a client, the useful step is to create evidence as you go. Save drafts under new names, keep the outline and the sources you read, and let your editor keep version history. Then record a SHA-256 hash of each saved draft in a simple log.
Hashes prove a draft has not changed since you logged it. A trusted timestamp from an independent authority proves when the hash existed, which closes the gap a self-kept log leaves open. Our step-by-step guide to hashing and timestamping your drafts shows both with the exact commands, and the page on RFC 3161 trusted timestamps explains how a timestamp authority signs that record.
If you are reviewing someone else's writing
Teachers, editors and clients face the same evidence problem from the other side. Start with the writer's process instead of the finished text. Ask for drafts, notes and version history, and compare the piece with earlier work you know the person wrote, written under similar conditions.
If you use a detector, record which tool and version you used, the score, the length of the text and the date. Keep in mind the two figures above: a tool that misses most AI text can still flag plenty of human text, and plain, careful prose by second-language writers is where false flags cluster.
Then talk to the writer. A short conversation about how the piece was planned, which sources were used and why a paragraph was cut tells you more than any score. Write down what you asked and what you saw, so the decision rests on evidence another person can review.
If you are accused
Ask what the claim rests on. If it rests on a detector score alone, ask which tool, which version and what threshold was used, and point to the published error rates above. A score is an estimate, and the people who built the best-known classifier withdrew it for low accuracy.
Stay factual and keep copies of everything you send. Then bring your process. Export the version history of the document, gather your drafts with their dates, and include any timestamped hashes. Offer to talk through how the piece developed and why you made particular choices. Writers can usually explain their own sentences, the sources behind them and the cuts they made, and that conversation is evidence too.
No AI detector score should be the only evidence that someone did not write their own work. That holds for teachers and editors reviewing a submission, and it holds for anyone deciding whether to trust what they read.
Sources
Questions
How accurate are AI text detectors?+
Less accurate than their scores suggest. OpenAI's own classifier caught 26% of AI-written text on its challenge set and flagged human text as AI 9% of the time, and OpenAI withdrew it in July 2023. Any detector's figures depend on the texts it was tested on.
Are AI detectors biased against non-native English writers?+
The best-known study found they can be. Liang and colleagues at Stanford ran seven detectors on 91 TOEFL essays by non-native writers and recorded an average false positive rate of 61.22%.
How can I prove I wrote something myself?+
Show the process. Dated drafts, the document's version history, your notes and sources, and a hash log of each draft are records an AI score cannot produce. Start keeping them before anyone asks.
Why are detector results on short texts so unreliable?+
Short texts give a classifier too little to measure. OpenAI described its own classifier as very unreliable below 1,000 characters, and said even longer texts were sometimes labelled wrongly.
Can text be watermarked like an image?+
Some providers do watermark generated text by adjusting word choices, and Google describes this for Gemini output. A watermark check only covers that one provider, and heavy rewriting can weaken the signal.
Related reading
- How-to
Hash and timestamp your drafts
Keep a dated record that you wrote it
- How-to
Trusted timestamps
How a third party vouches for when a file existed
- Explainer
AI watermarks vs hashes
What each one can show, and what it cannot
- Explainer
Is this image AI-generated?
A checklist that runs from strong evidence to weak