How to tell if something was written by ChatGPT

There is no reliable way to prove a passage was written by ChatGPT. That is the honest starting point, and everything else in this article follows from it. Automated detectors produce false positives at rates high enough to ruin lives and careers; the stylistic "tells" people trade online are weak signals that describe a lot of competent human prose; and the models themselves keep converging on the register of ordinary edited English. If you are trying to catch a specific document, you will usually be guessing. If you are trying to redesign a process so the question matters less, you have far better odds.

That does not mean the signals are worthless. It means they must be read with humility, weighted against base rates, and never treated as proof. Below is what actually works, what does not, and why the framing behind "how to tell if someone used ChatGPT" is wrong for most of the situations people ask about.

The honest answer: reliable detection is impossible

Text does not carry a fingerprint of its origin the way a photograph carries EXIF metadata. Large language models generate sequences of tokens that are statistically plausible continuations of a prompt, and by design those sequences resemble human writing. The better the model, the closer that resemblance. There is no hidden marker, no invisible signature, in a raw block of prose produced by ChatGPT, Claude, Gemini or any other current system. Watermarking research exists and some labs have experimented with it, but no widely deployed, robust, publicly verifiable watermark is in force across consumer chat products as of early 2026, and any watermark that survives copy-paste can usually be stripped by light paraphrasing.

The clearest institutional admission of this came from OpenAI itself. In July 2023 the company quietly retired its own AI Text Classifier, stating on the tool's page that it was shut down "due to its low rate of accuracy." A vendor withdrawing its own detector because it did not work is the most direct evidence available that the problem is genuinely hard, not merely unsolved by competitors.

So when someone asks whether ChatGPT can be detected, the accurate answer is: not reliably, not at the level of an individual document, and not to a standard you would want to act on if the consequences are serious.

Detector tool reality

Commercial detectors promise what the underlying science cannot deliver. GPTZero, Originality.ai, Turnitin's AI writing indicator and a rotating cast of smaller tools all output a probability or a percentage that reads like certainty. It is not.

The core problem is the false positive. A detector that flags genuine human writing as machine-generated does direct harm, and every large-scale detector produces these. A Stanford study led by James Zou and colleagues, published in the journal Patterns in 2023, found that several widely used GPT detectors flagged writing by non-native English speakers as AI-generated at strikingly high rates while classifying native-speaker essays correctly far more often. The mechanism is simple and damning: detectors lean heavily on "perplexity," a measure of how predictable the next word is. Non-native writers tend to use more common words and simpler constructions, which reads as low perplexity, which reads as machine-like. The tool is not detecting ChatGPT. It is detecting a smaller vocabulary, and penalising a population for it.

Vendors have iterated since, and some now report lower false-positive rates on their own test sets. Treat those figures with care. A detector's reported accuracy is measured on the vendor's chosen corpus, against the models the vendor chose to test, at the date they tested. It tells you little about how the tool performs on a paraphrased answer written by a bilingual student under exam pressure. Turnitin, to its credit, has publicly cautioned institutions against using its AI indicator as sole grounds for an academic misconduct case. That caveat is the honest part of the product.

For a fuller treatment of how these tools compete and how the evasion techniques evolve alongside them, see our analysis of the AI detection arms race between GPTZero, Originality and their rivals.

Writing pattern signals, and why they are weak

Ask people how to tell if something was written by ChatGPT and most will reach for a list of stylistic habits. There is truth in the list. There is also a trap.

The commonly cited tells are real tendencies in default model output:

• Certain vocabulary appears at elevated frequency: "delve," "tapestry," "navigate" used metaphorically, "underscore," "testament," "realm," "boasts," "landscape."
• Structural uniformity: paragraphs of similar length, sentences of similar rhythm, a tidy three-part enumeration where a human might sprawl.
• Rhetorical scaffolding: "It's important to note," "In conclusion," "On the other hand," and the balanced "not only X but also Y" cadence.
• A frictionless, slightly generic voice that hedges, qualifies and rarely says anything a reasonable person would disagree with.
• Heavy use of the em dash, which older ChatGPT outputs favoured for parenthetical breaks.

Here is why none of this is proof. First, every one of these habits is also present in good human writing. Business consultants love "leverage" and "landscape." Careful academics hedge. Plenty of people use the em dash correctly and often. Second, these are defaults, not constraints. A single instruction – "write in short, blunt sentences, no lists, opinionated" – erases most of the surface tells. Third, human writers who use these tools absorb their rhythms, so the style bleeds into work that was genuinely composed by a person who merely reads a lot of AI output. The signal degrades from both directions at once.

What the tells can do is raise a hypothesis. A document that combines several of them – uniform paragraphs, the stock vocabulary, the hedging voice, a suspiciously even structure – is more likely than average to have involved a model. That is a prior, not a verdict. Acting on it as though it were a verdict is where the damage happens.

Why detection mostly fails

Four forces work against anyone hoping to catch ChatGPT reliably.

Convergence. As models improve, their output moves toward the centre of well-edited human prose. The distributional gap that early detectors exploited keeps narrowing. A 2023 paper by Vinu Sankar Sadasivan and colleagues, "Can AI-Generated Text be Reliably Detected?", argued on both empirical and theoretical grounds that as language models approach human-level fluency, even the best possible detector's performance approaches that of random guessing.

Paraphrasing. Running model output through a second model, a paraphrasing tool, or a human editor's light rewrite is enough to defeat most detectors. The features they rely on are fragile. Change the surface and the signal collapses.

The false-positive floor. Because so much human writing legitimately resembles model output, any detector tuned to catch AI aggressively will also catch innocent people. Lower the threshold to protect the innocent and you miss most of the actual AI text. There is no setting that solves both problems, and the group most often caught in the crossfire is non-native speakers, exactly as the Stanford work showed.

Adaptation. Detection is adversarial. The moment a heuristic becomes known, it becomes an instruction to avoid. Anyone motivated to evade a detector can, with a few minutes of prompting or one editing pass.

When you can actually be confident

Confidence is possible in a narrow band of cases, and it almost never comes from stylometry. It comes from content errors that only a model would make.

The strongest tell is leftover refusal or assistant language: a paragraph that begins "As an AI language model, I cannot," or a helpful "I hope this helps! Let me know if you'd like me to expand." These are copy-paste failures, and they are close to conclusive. Similarly damning is fabricated citation: confidently formatted references to papers, cases or page numbers that do not exist, a hallucination pattern characteristic of models asked for sources they were not given. A document whose factual boundary stops abruptly at a model's training cutoff – no awareness of events after a certain date, delivered in an otherwise current-sounding piece – is another soft tell, though a lazy human writer produces the same gap.

Even here, confidence is about the artefact, not the author. A stray "As an AI" proves a model produced that text. It does not prove the person had no legitimate use for it, and it certainly cannot quantify how much of a longer document was machine-written.

The wrong framing

For most of the contexts where people ask how to tell if someone used ChatGPT – a classroom, a hiring pipeline, a content team – the detection question is the wrong question, and pursuing it produces worse outcomes than abandoning it.

The reason is arithmetic. If you scan a hundred essays with a detector that is, say, ninety-something per cent accurate, and only a fraction of the class actually used AI, a meaningful number of the flagged students will be innocent. You will then have to accuse people on evidence you cannot defend, and the accusations will fall disproportionately on the students who write in a plainer register. That is not a hypothetical; it is the documented pattern.

The better move is to redesign the process so that authorship is either verifiable or beside the point. In education that means in-class writing, oral defence of submitted work, drafts and version history, and assignments that require personal experience or specific local knowledge a model cannot fabricate convincingly. In content and knowledge work it means judging output on correctness and usefulness rather than provenance, and building workflows where a model is a declared, governed tool rather than a contraband one. This is the same shift we describe in our guide to using ChatGPT effectively through better prompting: the value is in the process and the verification around it, not in pretending the tool does not exist.

The deeper answer to provenance is not detection but authentication – proving what something is at the point of creation rather than guessing afterward. That is the premise behind emerging standards discussed in our explainer on the content authenticity concept. Signing content at source is a tractable engineering problem. Reverse-engineering origin from finished prose is not.

For specific contexts

Education. Faculty coverage in Inside Higher Ed and The Chronicle of Higher Education has moved steadily away from detection and toward assessment design, and for good reason. A false accusation is a serious harm, appeals are costly, and the tools cannot bear the evidentiary weight. If an institution uses a detector at all, it should be one input among several, never sole grounds, and students should be told it is in use. Turnitin's own guidance points the same way.

Journalism and publishing. Here the relevant question is rarely "was ChatGPT involved" but "is this accurate, sourced and disclosed." A newsroom policy that requires disclosure and holds every claim to the same verification standard, whoever or whatever drafted it, is more robust than any scanner. Fabricated quotes and phantom citations are catchable by fact-checking, which you should be doing regardless.

Hiring. Cover letters and take-home assignments are now routinely AI-assisted, and detecting that is both unreliable and mostly pointless. What a structured interview, a live problem-solving session or a supervised work sample reveals is whether the candidate can actually do the thing. That is the signal worth measuring. Screening for the ghost of a chatbot in a cover letter measures nothing you care about, and it penalises exactly the candidates – non-native English speakers again – most likely to lean on a tool for polish.

The through-line across all three is the same. Detection asks a question the technology cannot answer and, increasingly, need not answer. The durable strategies do not try to catch the machine after the fact. They make authorship observable, or they make it irrelevant by measuring what a person can actually do. As models keep converging on ordinary competent prose, that will only become more true, and the detectors will keep promising a certainty they cannot deliver. Verify their claims against the vendor's current documentation, check the date on every accuracy figure, and treat any percentage that looks like proof as the marketing it usually is.