Short answer: you cannot reliably detect AI-generated text, and you should never make a high-stakes decision — a failing grade, a firing, a rejected manuscript — on a detector score alone. The popular tools (GPTZero, Turnitin's AI indicator, Originality.ai, Copyleaks, Winston AI and the rest) hand you a probability, not proof, and that probability is wrong often enough to ruin real people's days. This guide explains what detectors actually measure, what the published research says about their accuracy, why they fail in predictable ways, and the signals that genuinely help editors and educators.
If you remember one thing: a detector output is a weak prior, not evidence. Treat it the way you would treat a smoke alarm that goes off when you make toast — worth a glance, never worth a verdict.
The 30-second verdict
- Detectors measure style, not origin. They estimate how statistically "predictable" your writing is. Predictable prose is not unique to machines, so the signal is fundamentally noisy.
- False positives land on the wrong people. Non-native English writers, plain technical writing, and even famous human documents get flagged as AI. That is an equity problem, not a rounding error.
- Evasion is trivial and cheap. A single paraphrase pass or a "vary your sentence length" instruction tanks most scores. The arms race structurally favors the evader.
- The fix is process, not a better detector. Assess the trajectory of the work (drafts, version history, oral checks) and verify facts. Those signals are far harder to fake than a finished paragraph.
How we evaluated the detection landscape
This is not a vendor benchmark with a single leaderboard, because a single accuracy number is exactly the kind of false comfort that gets people hurt. Instead we synthesized three things: (1) the published academic literature on detector accuracy and bias, including the widely cited Stanford study on bias against non-native writers and OpenAI's own decision to retire its classifier; (2) the vendors' own published claims and documentation; and (3) the structural logic of the problem — what a perplexity-based or classifier-based detector can and cannot know in principle. Where we give numbers, they are ranges or qualitative bands, because real-world accuracy swings wildly with text length, genre, model version, and how much a human edited the output. Anyone quoting you a clean "99% accurate" figure is selling, not measuring.
If you want the upstream context on how this text is generated in the first place, our guide to using AI to write blog posts and the companion piece on writing better AI prompts both show how easily output can be steered toward — or away from — the patterns detectors hunt for.
What AI text detectors actually do
Detectors do not hold a secret dictionary of "AI sentences." They estimate how statistically unsurprising a passage is, leaning on two core ideas.
Perplexity
Perplexity is, loosely, how surprised a language model is by each next word. AI writing tends to pick high-probability, "expected" words because that is literally what the underlying model optimizes for, so machine text reads as low-perplexity. Human writing is often lumpier — we reach for the odd word, break our own rhythm, contradict ourselves mid-paragraph. A detector runs your text through a reference model, scores the surprise, and treats low surprise as suspicious.
Burstiness
Burstiness is the variation in sentence length and complexity across a passage. Humans naturally mix a 4-word sentence against a 38-word one. Some AI output, especially from older or default-tuned models, is more uniform. Low burstiness nudges the score toward "AI."
Fine-tuned classifiers
Newer tools layer a trained classifier on top: feed the model thousands of labeled human-vs-AI samples and let it learn whatever separates them. This can lift accuracy on the exact distribution it trained on, and collapse on anything new — a different model, a different genre, a translated passage. Either way, the method is pattern-matching on style. None of it observes provenance. The detector never sees who typed what; it only sees the finished words, which is precisely the information that carries the least signal about how those words came to exist.
| Capability | Marketing claim | On long clean AI text | On edited / short / translated text |
|---|---|---|---|
| Flags obvious raw AI output | ✓ | ✓ | ~Drops |
| Resists simple paraphrasing | ✓ | ✕ | ✕ |
| Avoids false positives | ✓ | ~Sometimes | ✕ |
| Proves a specific person used AI | ~ | ✕ | ✕ |
| Calibrated, defensible probability | ✓ | ✕ | ✕ |
Why detectors fail
The core problem is that the thing detectors measure — predictable, fluent prose — is not unique to AI. Plenty of humans write that way, and plenty of AI output, once nudged, does not.
False positives hit exactly the wrong people
Detectors disproportionately flag:
- Non-native English writers. Their vocabulary and sentence structures are often simpler and more regular, which reads as low-perplexity. A 2023 Stanford study found detectors flagged the large majority of essays by non-native speakers as AI-generated while clearing native-speaker essays — a stark, well-documented bias. This single fact should disqualify detectors from any consequential decision in a diverse classroom.
- Formulaic but legitimate writing. Technical documentation, legal boilerplate, lab reports, structured beginner essays. Clear, plain writing scores as "AI" precisely because good editing strips out the lumpiness detectors rely on. If you have ever used a tool like Grammarly or its alternatives to smooth your prose, you have nudged your own writing toward the "AI" end of the scale.
- Famous human text. People have run the US Constitution, the Bible, and Moby-Dick through detectors and gotten confident "AI-generated" flags. It is a party trick, but it makes the point cleanly: the method has no ground truth.
Evasion is trivial
Even when a detector would have caught raw output, defeating it is almost free. Paraphrase once. Add the instruction "write with varied sentence length and a casual, slightly imperfect voice." Run the text through a "humanizer." Hand-edit four sentences. Any of these tanks the score. The economics are lopsided: evasion costs seconds, detection costs a research program, and every model upgrade resets the board in the evader's favor.
Scores are not calibrated
A "92% AI" score does not mean a 92% probability the text is AI. Vendors rarely publish how their thresholds map to real-world accuracy on text like yours, and accuracy craters on short passages (under a few hundred words) and on mixed human-plus-AI documents — which are now the default way people write. OpenAI itself quietly shut down its own AI-text classifier in 2023, citing low accuracy. When the lab that builds the models concedes it cannot reliably detect their output, third-party certainty should be treated with deep suspicion.
How the major detectors compare
No detector escapes the structural problems above, but they differ in positioning, transparency, and intended use. The table below is a directional map, not a leaderboard — treat every "strength" as relative within a fundamentally limited category.
| Detector | Primary audience | Notable strength | The catch |
|---|---|---|---|
| GPTZero | Educators | Popular, explains perplexity/burstiness, sentence-level highlights | Same false-positive exposure as the rest; not court-grade |
| Turnitin AI | Universities (LMS-integrated) | Embedded in existing plagiarism workflow | Several institutions disabled it over reliability concerns |
| Originality.ai | Content / SEO agencies | Bulk scanning, team workflows | Tuned for marketing copy; struggles on edited and short text |
| Copyleaks | Enterprises | Multi-language support, API | Multilingual makes the non-native bias problem worse, not better |
| Winston AI | Publishers | Readable reports, plagiarism combo | Marketed accuracy claims outrun independent results |
The pattern is consistent: each tool is genuinely better at something (workflow, reporting, integration), and none has solved the underlying detection problem, because it is not solvable from the finished text alone.
Better signals than a detector score
Since the finished text is increasingly impossible to fingerprint, move your attention to process and context — the parts of the work a detector never sees.
For educators
- Look at the trajectory, not the artifact. Drafts, outlines, revision notes, and version history are far harder to fabricate than a polished essay. Google Docs version history shows how a document actually grew, paste events and all.
- Use in-class or oral checkpoints. A 90-second verbal follow-up — "walk me through your argument in paragraph three" — reveals understanding no detector can measure, and it is fair to every student regardless of background.
- Design assignments AI cannot easily do. Tie work to a specific class discussion, personal reflection, local data, or an event after the model's knowledge cutoff. This shifts the goal from policing output to assessing thinking.
- Be transparent and document. State plainly what is and is not allowed, and never accuse based on a score. Open a documented conversation instead. The MLA-CCCC joint task force guidance is a sober reference for policy.
For editors and publishers
- Check verifiable facts. AI text is fluent but frequently wrong on specifics: fabricated citations, plausible-but-fake statistics, invented quotes. Fact-checking catches what style analysis cannot. This is the single highest-yield review habit.
- Read hallmarks as hints, not proof. Hedging filler, repetitive transitions ("Moreover," "In conclusion,"), suspiciously even structure, and confident vagueness all suggest a closer read — never a verdict.
- Invest in the writer relationship. Knowing your contributors, their beat, and their track record beats any tool. If you are commissioning at scale and worried about volume, our guide to building content workflows around AI writing covers editorial guardrails that scale better than scanning.
- Watch the transcription edge case. Interview and podcast workflows that run through an AI transcription tool produce machine-clean text from genuine human speech — a classic false-positive trap if you later scan it.
How to use detectors responsibly (if at all)
Detectors are not useless; they are badly misused. The governing rule: the higher the stakes for an individual, the less weight a detector score deserves.
| Use case | Appropriate? | Why |
|---|---|---|
| Screening signal to prompt a closer human review | Yes, with caution | Low stakes, a human still makes the final call |
| Sole basis for an academic misconduct charge | No | False positives ruin lives and the score is not defensible |
| Bulk content audits across a large site | Maybe | Trends across thousands of docs are more meaningful than any single score |
| Proving a specific person cheated | No | No detector can establish provenance from finished text |
| Triaging a freelance content pipeline | Maybe | Pair with fact-checking and editor judgment, never standalone |
Use a detector to ask a question, never to deliver an answer. The moment a score becomes the decision instead of the prompt for a human to investigate, you have misused the tool.
What about watermarking and provenance standards?
The more durable future is not better detection of finished text but provenance baked in at generation time. Google's SynthID embeds statistical watermarks in model output, and the C2PA standard attaches signed content credentials to media. These are genuinely more principled than perplexity-guessing because they carry a verifiable signal of origin rather than inferring it from style.
The honest caveats: watermarks only cover text from cooperating models, they can be weakened by heavy editing or paraphrasing, and nothing stops someone from using a model that does not watermark at all. Provenance standards help most for the responsible majority and least against a determined bad actor — which is the inverse of what high-stakes enforcement actually needs. Useful infrastructure, not a silver bullet.
The honest conclusion
The detection era was brief. As models improve, machine text drifts toward the center of "normal human writing," and the statistical gaps detectors rely on keep shrinking. You cannot win a measurement race against a system explicitly optimized to look like the thing you are measuring.
The durable approach is not a better detector — it is better process. Assess thinking instead of artifacts. Verify facts instead of vibes. Keep a human in the loop for any decision that matters, and design your assignments, your editorial standards, and your hiring tests so that the process of doing the work is what you evaluate. If a tool ever promises certainty about who or what wrote a piece of text, that promise is the clearest sign it is overselling. The right question was never "is this AI?" It was always "is this true, is this understood, and did this person actually do the work?" — and those you can still answer.