We had Hemmingway-1.5, our newest model, write 1,848 texts in 13 kinds of writing, and ran each one through GPTZero and four of the top open AI detectors, once. 90% came back human. Then we ran 1,670 texts that people wrote through the same detectors, and 86% came back human. Hemmingway-1, our earlier model, passed 7% of the same test.
This is the test behind Hemmingway-1.5, released on 25 October 2026, and behind what we call undetectable AI on this site. We ran it ourselves, on the model we make, so this page sets out how it was done and every number. The same numbers are on the model card.
The short answer
- Hemmingway-1.5: 90% of its writing passed GPTZero and four other AI detectors on the first try, at every thinking level.
- People: 86% of real human writing passed the same detectors. The other 14% was flagged as AI.
- GPTZero alone: passed 90% of Hemmingway-1.5's texts, at each of the four thinking levels.
- Other big models: on the same 80 everyday asks, the detectors passed between 13% and 26% of their answers.
How the test was run
- What it wrote. 1,848 texts in 13 kinds of writing, among them emails, stories, speeches, blog posts, forum comments and product descriptions. Each was asked for in plain words, the way a person would ask.
- How it was set. All four thinking levels: off, low, medium and extra high, at temperature 1.0, the setting the model is made for.
- One try each. Every text went through the detectors once. Nothing was picked out, and nothing was run again.
- The detectors. GPTZero, and four of the top open AI-detector models: Fakespot from Mozilla, Vanguard, WasItAIGenerated and MELD. A text passed when they read it as human with at least 90% confidence.
- The human yardstick. 1,670 emails, essays, blog posts, reviews, stories and comments written by people went through the same detectors, the same way.
- The other models. Fable 5.1, GLM-5.3, GPT-6 Astra, Kimi K3, Qwen3.8 Max and DeepSeek V4 Pro answered the same 80 everyday asks, taken from CommunicationBench, our own benchmark, and went through the same detectors.
The results
Next to people
| Writer | Passes the detectors |
|---|---|
| Hemmingway-1.5 | 90% |
| Real human writing | 86% |
| Hemmingway-1 | 7% |
GPTZero, by thinking level
| Thinking | Hemmingway-1.5 passes GPTZero |
|---|---|
| Off | 90% |
| Low | 90% |
| Medium | 90% |
| Extra high | 90% |
By kind of writing
| Kind of writing | Hemmingway-1.5 passes |
|---|---|
| Product descriptions | 100% |
| Blog posts | 96% |
| Forum comments | 96% |
| Stories | 94% |
| Speeches | 94% |
| Emails | 93% |
All six kinds shown passed more often than the 86% that human writing got.
The same 80 asks, other models
| Model | 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% |
What the results say
- Detectors catch most AI writing. The answers from six of the biggest models were flagged most of the time, between 74% and 87% of them.
- They also catch people. One in seven texts that people wrote was flagged as AI. A detector reads patterns in the text. It cannot know who typed it.
- Training changes the result. Hemmingway-1 passed 7%. Hemmingway-1.5, built on it at the same size, passed 90%.
- Thinking does not hurt it. The result held at all four thinking levels, so a longer think before writing did not make the text read as more machine-made.
How good the writing is
We score the quality of the writing separately. With thinking on, our quality judge gives its writing 4.2 out of 5, and 4.5 out of 5 for doing what was asked. 64% of its first tries pass the detectors, write well and do what was asked, all at once.
Check it yourself
- Read real outputs from the test, each a first try that read as human, on the model card.
- Download the weights from Hugging Face and run your own texts through the detectors you use.
- Try your own eye on Spot the AI, which puts a person's message next to Hemmingway's.
For what the signs of AI writing are, and why neither a person nor a detector can be sure, read how to tell if something was written by AI.
Try HemmingwayDownload the app
Common questions
How accurate are AI detectors?
Not accurate enough to be the only evidence. In our test, the detectors flagged 14% of writing that people wrote themselves, and they passed 90% of Hemmingway-1.5's writing. They caught most of what the other big models wrote, between 74% and 87% of it.
Which AI detectors did you test?
GPTZero, and four of the top open AI-detector models: Fakespot from Mozilla, Vanguard, WasItAIGenerated and MELD. A text passed when they read it as human with at least 90% confidence.
Can I see the texts?
Real outputs from the test are on the model card, each marked with what was asked. The weights are open, so you can write your own and test them.