For two years the honest answer to 'can AI detectors catch ChatGPT?' was: not reliably. They flagged human writing as machine-made often enough that no serious school or employer could act on the result alone, and a famous 2023 case saw a professor nearly fail a whole class on a detector's word. That answer is now changing. On July 29, 2026, Pangram Labs raised $9 million and shipped Pangram 4, a detector its makers say has a false-positive rate of roughly 1 in 24,000 documents. Independent researchers report numbers in the same range. So it is worth asking the question again, carefully — because the stakes for students, writers and job seekers are real, and the marketing is louder than the science.
This guide separates what these tools genuinely do from what people assume they do. The short version: the best 2026 detectors are dramatically more accurate than the free ones most people have tried, but even the best cannot see how a document was written. A detector produces a probability, not a receipt. Knowing the difference is the whole game.
Quick summary: Pangram Labs launched Pangram 4 on July 29, 2026 alongside a $9M raise led by Menlo Ventures. Pangram reports a false-positive rate around 0.0041% (about 1 in 24,000 documents) and a false-negative rate of 0.34%, down from 1.99% in Pangram 3. Independent testing by University of Chicago Booth researchers reported accuracy at or near 100% on most models, never dropping below 99.8%, and found false positives essentially zero on their test set; Pangram's own 1-in-24,000 and 1-in-10,000 false-positive figures are self-reported. Even so, no detector can prove how a document was made: 'humanized' AI text still slips through some of the time, and heavily formal or academic human writing remains the most likely thing to be wrongly flagged. Treat a detection score as evidence, never as proof.
Why Old Detectors Failed
The detectors most people formed an opinion on — the free web tools of 2023 and 2024, and OpenAI's own classifier, which the company quietly retired in 2023 for low accuracy — worked by measuring 'perplexity' and 'burstiness': how predictable and how varied the text was. The theory was that machines write more smoothly than humans. In practice, plenty of humans write smoothly too. Non-native English speakers, in particular, were flagged at alarming rates because their sentence structures looked 'too regular.' A widely cited Stanford study found detectors misclassified genuine writing by non-native speakers as AI more than half the time in some tests. When a tool is wrong that often on a real population, its output is not usable for any decision that matters.
The newer generation works differently. Instead of hand-picked statistical features, tools like Pangram are trained models — classifiers taught on very large, balanced sets of human and AI text across many models and writing styles, and updated as new models ship. That shift is why the accuracy numbers jumped. But it does not repeal the core limitation, and vendors are usually careful to say so even when their headlines are not.
The Two Numbers That Actually Matter
Accuracy is a misleading single figure. What matters for anyone who could be wrongly accused is the false-positive rate: how often human writing gets called AI. And what matters for anyone relying on the tool to catch cheating is the false-negative rate: how often AI writing sails through. Pangram reports a false-positive rate around 0.0041% — roughly one wrongly flagged document in 24,000 — and a false-negative rate of about 0.34%, improved from 1.99% in the previous version. Independent evaluation by University of Chicago Booth researchers reported accuracy at or near 100% on most models and found false positives essentially zero on their test set — while the specific 1-in-10,000 and 1-in-24,000 false-positive rates are Pangram's own reported figures, not the independent result. Source: Pangram (self-reported rates); University of Chicago Booth evaluation (independent accuracy).
Those are genuinely strong numbers, and far better than the tools that gave detection its bad name. But read them precisely. A 1-in-10,000 false-positive rate sounds like near-certainty until you apply it to a school that runs 10,000 essays a term — that is, on average, one innocent student flagged every term, with no way for the tool to tell that student apart from a real offender. The rate is low; it is not zero. And the numbers come mostly from controlled test sets, which tend to be cleaner than the messy reality of real submissions.
| What people assume | What detectors actually do | Why the gap matters |
|---|---|---|
| It proves I used AI | It estimates a probability from text patterns | A score is evidence to weigh, not a confession |
| 0% false positives | ~1 in 24,000 (Pangram, self-reported) | Rare, but real humans still get flagged |
| It catches everything | Misses some 'humanized' or edited AI text | Confident cheaters can still beat it |
| All detectors are equal | 2026 trained models beat 2023 free tools by a wide margin | The tool you tried in 2023 is not today's tool |
| It knows how I wrote it | It only sees the final text | AI-assisted editing of your own words is invisible to it |
The 'Humanizer' Arms Race
A whole category of tools now exists to launder AI text past detectors — 'humanizers' that reword machine output to look hand-written. Detectors and humanizers are locked in a cat-and-mouse game, and the current state of it is telling. Pangram reports catching humanized text as AI-generated about 97.7% of the time, and as either mixed or AI about 98.8% of the time. That is high, but the missing few percent is exactly where a motivated cheater operates. The practical takeaway cuts against both sides: detection is good enough that casual copy-paste is likely to be caught, and imperfect enough that someone determined to defeat it sometimes can. Building a disciplinary process on the assumption of perfection is a mistake in either direction.
Who Gets Wrongly Flagged
The false positives are not random. They cluster on writing that happens to share statistical fingerprints with AI: highly formal, structured, generic academic prose — the exact register that careful students and ESL writers are often taught to produce. If your writing is clean, evenly paced and low on personal voice, you sit closer to the boundary, not because you cheated but because good expository writing and AI writing overlap. This is the single most important thing for students to understand: sounding 'too polished' is a risk factor a detector cannot distinguish from guilt. Keeping specifics, personal detail and your own drafts is the best protection.
- If you're a student: keep your version history and drafts. Google Docs and Word both track edits, and a real writing trail is far stronger evidence than any detector's score.
- If you're an educator: use detection as a prompt to talk, not as a verdict. Every serious vendor, Pangram included, says its score is not designed to be the sole basis for an accusation.
- If you write for a living: a false flag can cost a client. Save drafts, and know that formal, generic prose is the most flaggable — voice and specificity lower your risk.
- If you're job hunting: some hiring tools now scan cover letters. A fully AI-written letter is the most detectable; a letter in your own voice, AI-assisted or not, is the safest.
- For everyone: never trust a single free detector's verdict on yourself or anyone else. Accuracy varies enormously between tools.
So — Can They Catch ChatGPT?
Yes, far better than a year ago, and well enough that pasting raw ChatGPT output into a graded essay is now a genuine gamble. But 'catch' is the wrong verb. A 2026 detector produces a strong probability that text is machine-written; it does not, and cannot, prove authorship or intent. The honest framing — the one the best vendors use themselves — is that these tools are a smoke alarm, not a court. They tell you where to look. What you do next has to involve a human, a conversation and actual evidence, because the one thing no detector will ever have is a recording of how the words came to exist.
01How accurate is Pangram 4, really?
Pangram reports a false-positive rate around 0.0041% (about 1 in 24,000) and a false-negative rate near 0.34%. Independent University of Chicago Booth testing reported accuracy at or near 100% on most models with false positives essentially zero; the specific 1-in-10,000 and 1-in-24,000 false-positive figures are Pangram's own, not the independent finding. Strong numbers — but measured largely on controlled test sets, not messy real-world submissions.
02Can a detector be wrong about my writing?
Yes. Even at a 1-in-10,000 false-positive rate, real human writing is sometimes flagged as AI — most often when it is very formal, generic or structured, which is common in academic and ESL writing. That is why a score should never be treated as proof on its own.
03Do 'humanizer' tools beat detectors?
Sometimes. Pangram says it catches humanized text as AI about 97.7% of the time, so most humanized text is still caught — but the remaining few percent is exactly where determined evasion succeeds. It is an ongoing arms race, not a solved problem.
04If I use AI to edit my own writing, will it be flagged?
A detector only sees the final text, not the process. Lightly editing your own words with AI usually leaves writing that reads as human. Generating whole passages from a prompt is what gets flagged. But policies differ — check what your school or employer actually permits before relying on this.
05What is the safest way to protect myself from a false accusation?
Keep your drafts and edit history. A visible writing trail in Google Docs or Word is far more convincing than arguing with a detection score after the fact.
The bigger picture is that detection and generation are improving together, and no single tool — on either side — deserves blind trust. That is also the practical case for testing models yourself rather than taking any vendor's headline at face value. LumiChats keeps many current AI models under one login at a pay-per-day price, so you can see how different systems actually write, and judge the claims about them with your own eyes instead of a marketing page.
