The myth of the reliable AI detector

Before we talk about undetectability, we have to face a fact that few tools in this space dare remind their users of: AI text detectors are not reliable, and it is documented.

OpenAI, which has more access to GPT models than anyone, launched its own AI Text Classifier in January 2023. Six months later, in July 2023, the company quietly pulled the plug. The official reason: a low accuracy rate. The classifier only correctly identified 26% of AI-generated text, and wrongly flagged 9% of human text. On short text (under 1,000 characters), it was simply unusable.

« As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. »

OpenAI, official announcement, July 2023

If the very company behind ChatGPT cannot detect the content produced by its own model, how could third-party tools like GPTZero, Originality, ZeroGPT or Copyleaks be any more accurate?

What the peer-reviewed studies say

Several academic studies published in 2023, 2024 and 2026 converge on the same conclusions.

1. Detectors are biased against non-native English speakers

A study led by James Zou and his team at Stanford (published in Patterns in 2023, peer-reviewed) tested seven mainstream AI detectors on essays written by international students and by native American students. The verdict:

  • The seven detectors flagged 61% of human texts written by non-native English speakers as « written by AI ».
  • For about 20% of those human texts, all seven detectors agreed: « written by AI ». A unanimous verdict, and completely wrong.
  • For equivalent texts written by native American students, the error rate dropped below 5%.

The structural reason is that non-native English speakers use a less varied vocabulary, simpler constructions, and more common turns of phrase. That is exactly the statistical profile of text generated by a large language model, which assembles the most probable words and sentences from its training corpus. Detectors cannot tell human simplicity apart from algorithmic simplicity.

2. Accuracy collapses as soon as the text is edited

A 2026 meta-analysis (Springer, International Journal for Educational Integrity) evaluated AI detector accuracy on three types of content:

  • Raw, untouched AI text: average accuracy between 88% and 95% depending on the detector.
  • AI text lightly reworded by hand: accuracy drops to between 60% and 80%.
  • AI text paraphrased or humanized by a dedicated tool: accuracy close to chance, around 50%.

In plain terms, as soon as AI-generated text passes through a human who rewrites it even partially, or through a humanizer like Rephrase, detectors lose their ability to tell it apart from purely human text.

3. False positives on human text are the norm

Still in the Springer 2026 meta-analysis, false-positive rates (human texts wrongly flagged as AI) vary by detector:

  • GPTZero (free): between 3% and 8% false positives on everyday human text, up to 40% on formal academic text or text written by non-natives.
  • ZeroGPT: 14.6% to 33% false positives depending on the content type, peaking on structured text and academic writing.
  • Copyleaks: variable rates, more stable than ZeroGPT but with documented false positives on highly polished content.

There are even real cases documented by The Markup in August 2023 of international students falsely accused of cheating by their universities because of a Turnitin or GPTZero score, when they had in fact written their own work.

The Rephrase method: 3 pillars

Our approach is not about « fooling » a detector (which, as we just saw, is wrong half the time anyway). It is about producing text that even an attentive human reader would perceive as naturally human. If the result is good enough for a human, it is necessarily good enough for a machine.

Pillar 1: break the statistical patterns typical of AI

Models like ChatGPT, Claude or Gemini have measurable structural tics: sentences of a very steady length (between 20 and 25 words), too many adverbs ending in -ly, pompous unsourced vocabulary (« crucial », « pivotal », « optimal », « essential »), mechanical pairings (« not only… but also… »), overused transition phrases (« indeed », « furthermore », « it is worth noting that »).

Our system prompt explicitly instructs the model to eliminate these patterns and to vary sentence length organically. An 8-word sentence can follow a 27-word one, which is typical of a human thinking out loud while writing.

Pillar 2: apply a human stylistic signature

You can create a custom voice from 3 to 5 texts you have actually written in real life (LinkedIn, emails, notes). Our analyzer extracts your linguistic signature from them: average sentence length, preferred connectors, register, stylistic tics, characteristic vocabulary, punctuation, paragraph rhythm. That signature is then injected into the prompt as a top-priority instruction, above the style preset.

The result: the humanized text is not just « not-AI », it reads like your own writing. A reader who knows you recognizes your voice.

Pillar 3: allow human imperfections

Human text is never perfect. It sometimes contains a missing comma, a dropped accent, a clumsy apostrophe, a typo. Rephrase's « Add mistakes» mode deliberately introduces plausible imperfections into the humanized text, at the density you choose (from « barely there » to « very strong »). These mistakes never touch proper nouns, numbers or quotes: only the writing around them, where a human would have left some.

Across 8 possible mistake types (missing plurals, apostrophes turned into spaces, dropped accents, double consonants, punctuation, punctuation spacing, homophones, typos), you switch on whatever matches your profile. Someone who types fast will tend to produce typos; someone writing on mobile will tend to drop accents. The granularity is intentional.

Why this method produces text that is genuinely undetectable

Put end to end, these three steps eliminate the three signals AI detectors look for first:

  • The steady perplexity of the text (which detectors measure as a statistical-improbability score). Our sentence-length variation and stylistic signature break that signal.
  • The average vocabulary, with no peaks or troughs, typical of LLMs. Our injection of characteristic vocabulary (your voice) produces highs and lows.
  • The orthotypographic perfection. Our mistakes mode degrades it in a controlled way, the way a real human would.

That is also why text humanized by Rephrase slips under the radar even of recent detectors, whether they are called GPTZero, Originality, Copyleaks or ZeroGPT. Not because we found the flaw in their algorithm (they change it every week), but because the text we produce is, structurally, indistinguishable from human text.

What Rephrase does not claim

For the sake of honesty, here is what we do not promise:

  • We do not promise that AI detectors will one day be reliable. They are not, they never have been, and they probably never will be (the problem is mathematical: an LLM imitates the average of human text, so any human text within that average becomes indistinguishable).
  • We do not promise that humanized text will be 100% undetectable by every tool, in every case, forever. Detectors evolve. So do LLMs. So does our method: we keep tuning our prompts as the underlying models change.
  • We do not promise that academic or professional cheating is risk-free. If you turn in work that is not yours, that is your responsibility, not ours.

What we do promise is text written in natural language that sounds human to an attentive human reader. Everything else follows from that.

Sources and further reading