AI Detectors Barely Miss ChatGPT. They Miss AI Trained to Write Like a Specific Person.

Quick answer

A new independent benchmark from Epoch AI tested three of the most-used AI text detectors — Pangram, GPTZero, and Originality.ai — on both human writing and AI text. When AI was generated from a plain prompt, all three caught it almost every time, with false-negative rates at or near zero. When the same models were shown five real passages from a specific author and told to write in that person's style, the story flipped: the detectors missed between roughly 1 in 10 and nearly 1 in 5 of those passages, and scientific writing was the hardest genre of all, with miss rates over 25% for some model-detector pairs. The takeaway isn't that detection is broken. It's that a single document-level verdict hides exactly the kind of AI use that matters most — text built to sound like someone specific — and that's a reason to read the sentence-level breakdown, not just the headline score.

What Epoch AI actually tested

The study comes from Epoch AI, an independent research group that posted full methodology and a downloadable dataset (last updated July 16, 2026). Their setup is simple: 495 passages of roughly 500 words apiece, taken from 99 documented authors — bloggers, published short-fiction writers, and scientists — all pulled from pre-2022 archives, so none of the human text could have been quietly written by a model. Wayback Machine snapshots and arXiv preprints supplied the source material, each one checked against the original by word-level overlap and publication date.

Three frontier models did the generating. Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro each contributed passages. They wrote under two different conditions. Each condition yielded a unique style from each model.

• Basic prompt: a short, generic instruction with no reference material — "write a short story about a lost dog."
• Style imitation: the model is shown five real passages from one of the 99 authors and asked to write something new "in the same style."

For all submissions, the process applies Pangram (version 3.3.2), GPTZero, and Originality.ai with their out-of-the-box parameters. A "Mixed" outcome on any of these platforms counts as a miss.

The numbers that matter

Across every metric, the false-positive outcome looks positive. Pangram reported no false positives among 495 genuine human texts. GPTZero likewise found zero false positives in those same 495 passages. Originality.ai identified 19 of 495 real human segments, roughly 3.8%. Those 19 segments came exclusively from blogging and fiction, not from scientific writing.

The detectors produced near-perfect accuracy, and no implicit masking was needed. Pangram did not miss any entry. A total of 297 items were evaluated, of which GPTZero omitted two. Originality.ai skipped one item out of the same 297. Consequently, the detection system can catch virtually all AI-generated text when it isn't masked. This confirms the tools' effectiveness in clear, uncomplicated situations.

The style-imitation condition is where the gap opens up:

Scientific writing was consistently the hardest genre for every detector, likely because dense, citation-heavy, formally hedged prose already narrows the stylistic gap between a careful human researcher and a model tuned to sound careful and formal. Fiction was the easiest genre to catch across the board, which tracks with how distinctive a real author's fiction voice tends to be compared to a model's smoothed-out approximation of it.

Why this matters more than the average headline number

This site advertises a single accuracy figure too — 93% on our own blind test set, same as every other AI-detection vendor. What Epoch's data makes concrete is that "accuracy" actually collapses two very different jobs into one number: catching text generated with no attempt to disguise it, and catching text specifically produced to resemble one person's writing. The first job is close to solved. The second is where a well-resourced attacker, or just a student who fed a model five of their own old essays as a style reference, has real room to slip through.

A document-level percentage alone isn't enough. Our own detection-score guide makes this point from the other direction: one aggregate number hides whether the issue is a single heavily-generated paragraph or several lightly-touched ones. Epoch's numbers show the same principle from the miss side. An 82-90% catch rate for style-imitated text sounds solid, until you remember that means one in ten documents in that category got through completely undetected, no flag at all, because the whole thing read clean.

This is a different problem than the false-positive story

Most of the AI-detection news this year has gone the other direction — a novelist's book pulled over a detector score, a newspaper editor accused over three op-eds, a population-level pattern used as proof against one person. That's a false-positive and misuse problem: a real signal stretched past what it can actually support. Epoch's study is the opposite failure mode, and it's worth being honest that both exist in the same tools at the same time. A detector can be excellent at leaving honest writers alone and still have a real blind spot for text that was built, on purpose, to sound like somebody specific. Neither finding cancels the other out. We've also written about why "humanized" AI text is a different, related evasion path — paraphrasing tools degrade the writing to dodge perplexity-based checks, while style imitation works by narrowing the gap in the other direction, making AI text sound more like a particular real voice instead of generic AI voice.

What this means if you're on either side of a flag

Educators, editors, and hiring managers should treat a detector score differently depending on the writing type. A clean result means more on fiction or casual prose than on formal, citation-heavy, or scientific text, where every tool in Epoch's test struggled harder. Something that reads suspiciously polished in that category and still comes back "human" doesn't settle anything either way — that's the exact scenario Epoch's numbers flag as hardest to call.

Here's the flip side, and it's reassuring if you're a writer worried about a false accusation. All three detectors performed close to flawlessly on real human writing, across every genre Epoch tested. The people most likely to get caught in this specific gap aren't honest writers with a formal style. They're texts deliberately produced to copy someone's voice.

Want to know where your own draft stands instead of guessing from someone else's benchmark? Run it through TheChecker.AI's free demo and read the sentence-level breakdown, not just the headline percentage. Our own accuracy figures, and how we measure them, are on the accuracy page — the same discipline Epoch AI is asking every detector to meet.

FAQ

Does this mean AI detectors don't work? So no, AI detectors are not broken. On text with no attempt to hide its origin, every detector in this test caught nearly all of it, and real human writing rarely got misflagged. The actual gap is narrower and more specific: writing built on purpose to sound like one particular person, especially in scientific or academic prose.

Which detector performed best in this test? When tasked with style imitation, Pangram records a 10.1% miss rate, while GPTZero records 10.8%; the two systems perform almost equally. In contrast, Originality.ai's miss rate is noticeably higher at 17.9%. In addition, Pangram and GPTZero produce no false positives on authentic human writing, whereas Originality.ai still flags a few. Read Epoch's own tables before drawing firm conclusions, though. The differences noted are genuine, but the sample size of 99 passages per genre-model pair leads to wide confidence intervals, tempering the strength of the conclusions.

Why is scientific writing the hardest genre to check? Academic prose is already narrow and formal by convention — hedged claims, consistent structure, technical vocabulary — which is close to what a language model naturally produces when asked to sound careful. That overlap narrows the statistical gap detectors rely on, and it gets narrower still when the model is specifically shown a scientist's real writing to imitate.

What should I actually do with a detector score, given this? Turn to a strong signal in the easy cases, a weaker one on formal or scientific writing where imitation is plausible. The sentence-level breakdown matters more than a single percentage. Neither a clean verdict nor a flagged one is the final word on its own, and that holds whether the worry is a false positive or something that slipped past undetected.