Most AI writing does not pass. In our test, the answers that six of the biggest models gave to 80 everyday asks were flagged by GPTZero & Pangram and four other detectors most of the time: they passed between 13% and 26%. Hemmingway-1.5, our own model, is the exception we built: GPTZero & Pangram alone read 90% of its writing as human, at every thinking level, and 90% passed all the detectors together.
This page is what we found and how. The full method is in our AI detector test, and the case for Hemmingway-1.5 as undetectable AI is on its own page.
What GPTZero & Pangram are
GPTZero & Pangram are AI detectors: you paste in text and they give back how likely it is that an AI wrote it. Their sites say they flag text from models including ChatGPT, Gemini and Claude. GPTZero's site says teachers and students use it, that no AI detector is 100% accurate, and that its results should not be used to punish anyone or as the final verdict.
What our test found
| Writer | Passes the detectors |
|---|---|
| Hemmingway-1.5 | 92% |
| Fable 5.1 | 26% |
| GLM-5.3 | 21% |
| GPT-6 Astra | 20% |
| Kimi K3 | 18% |
| Qwen3.8 Max | 18% |
| DeepSeek V4 Pro | 13% |
That is the same 80 everyday asks for every model, each answer run once through GPTZero & Pangram, Fakespot, Vanguard, WasItAIGenerated and MELD. A text passed when they read it as human with at least 90% confidence.
On their own, GPTZero & Pangram passed 90% of Hemmingway-1.5's 1,848 test texts, at each of the four thinking levels: off, low, medium and extra high. We did not run the other models through GPTZero & Pangram alone, so we give their results only for all the detectors together.
Why AI writing gets caught
Detectors read patterns. What gives most AI writing away is not one word but its habits:
- Even rhythm. Sentences of the same length, paragraphs of the same size.
- Stock words. "Moreover", "pivotal", "it's worth noting", "I hope this finds you well".
- The wrapper. An opening that restates the ask, and a closing that sums up or offers more help.
- Nothing specific. General lines where a person would give the name, the day or the number.
A model trained to write like a person leaves those habits out. Hemmingway-1, our first model, passed the same detector test 7% of the time. Hemmingway-1.5, built on it at the same size, passed 90%.
Does editing AI text help?
Sometimes. Cutting the opening and closing lines, swapping stock words for plain ones and putting in real details all make text read more like a person's, and how to make AI text sound human walks through it by hand. A model can do the rewrite too: by Hemmingway-1.5's Hugging Face card, 89% of its rewrites of AI text passed the same detectors.
Do GPTZero & Pangram get it wrong?
Yes, in both directions. In our test, the detectors together flagged 14% of 1,670 texts that people wrote themselves. A detector score is a guess with a number on it, and GPTZero's own site says it should not be the final verdict. How to tell if something was written by AI covers what to do instead of trusting a score.
Try HemmingwayDownload the app
Common questions
Can GPTZero & Pangram detect ChatGPT?
Their sites say they flag text from ChatGPT, Gemini and Claude. In our test, the answers of GPT-6 Astra, the ChatGPT model we tested, passed GPTZero & Pangram and four other detectors together only 20% of the time.
Does Hemmingway pass GPTZero & Pangram?
Nine times in ten. GPTZero & Pangram read 90% of Hemmingway-1.5's texts as human, at every thinking level, with each text run once.
Are GPTZero & Pangram accurate?
Not enough to be the only evidence. GPTZero's own site says no detector is 100% accurate. In our test, the detectors together flagged 14% of writing by real people.
How do I make AI text pass GPTZero & Pangram?
Start from a model that writes like a person, or rewrite the draft so it has a real voice: real details, uneven sentences, no wrapper. Hemmingway-1.5 does both, and GPTZero & Pangram read 90% of its writing as human in our test.