Emails
Work emails, replies, the one you have been putting off. Emails passed the detectors 93% of the time.
Hemmingway-1.5
Undetectable AI
Hemmingway-1.5 writes text that passes GPTZero and the top AI detectors 9 times in 10, on the first try. Real human writing passes the same detectors 86% of the time.
Hemmingway-1.5 is an AI writer whose writing reads as human to AI detectors. We had it write 1,848 texts in 13 kinds of writing and ran every one through GPTZero and four of the top open AI detectors, once each. 90% came back human. It is the model in the Hemmingway app, free to try with no card, and its weights are open.
If you searched for undetectable AI, an AI that passes AI detectors or an AI humanizer, this is ours, with the whole test on the same page.
90%Hemmingway-1.5passes GPTZero and the top AI detectors
86%Real human writingpasses the same detectors
7%Hemmingway-1our first model, on the same test
At a glance
- What it is
- A 27B writing model that writes the way a person does
- Passes AI detectors
- 90% on the first try, at every thinking level
- GPTZero alone
- 90%
- Real human writing
- 86% on the same detectors
- Detectors
- GPTZero, Fakespot, Vanguard, WasItAIGenerated and MELD
- Tested on
- 1,848 texts in 13 kinds of writing
- Rewrites AI text
- Yes: paste it in and ask for it in a person’s voice
- What it writes
- Emails, essays, posts, messages, reviews, stories and speeches
- Where
- The Hemmingway app on Mac, Windows, Android and the web, and the API
- Cost
- Free to try, no card needed. Plans on pricing
- Weights
- Open, CC BY-NC 4.0, on Hugging Face
- Made by
- Hemmingway AI (Pty) Ltd (about us)
The results
Every number is a first try: no picking, no retries. Higher is better. The same numbers are on the model card.
| Writer | Passes |
|---|---|
| Hemmingway-1.5 | 90 |
| Human writing | 86 |
| Hemmingway-1 | 7 |
| Model | Passes |
|---|---|
| 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 |
| Thinking | Passes |
|---|---|
| Off | 90 |
| Low | 90 |
| Medium | 90 |
| Extra high | 90 |
| Kind | Passes |
|---|---|
| Product descriptions | 100 |
| Blog posts | 96 |
| Forum comments | 96 |
| Stories | 94 |
| Speeches | 94 |
| Emails | 93 |
Hemmingway-1.5Other models and Hemmingway-1
The detectors are GPTZero and four of the top open AI-detector models: Fakespot from Mozilla, Vanguard, WasItAIGenerated and MELD. A text passes when they read it as human with at least 90% confidence. The human writing is 1,670 emails, essays, blog posts, reviews, stories and comments written by people.
How we tested it
Here is the whole test.
- 1,848 texts. Thirteen kinds of writing, among them emails, stories, speeches, blog posts, forum comments and product descriptions, each asked for in plain words.
- Every thinking level. Off, low, medium and extra high, all 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.
- Five detectors. GPTZero, and the open models Fakespot, Vanguard, WasItAIGenerated and MELD. A text passes when they read it as human with at least 90% confidence.
- People as the yardstick. 1,670 texts that people wrote went through the same detectors. 86% passed, so 14% of real human writing was flagged as AI.
- Other models, same asks. Fable 5.1, GLM-5.3, GPT-6 Astra, Kimi K3, Qwen3.8 Max and DeepSeek V4 Pro answered the same 80 everyday asks, from CommunicationBench, our own benchmark.
Real outputs from the test, each one a first try that read as human, are on the model card. The method and the outputs are also on Hugging Face, and the weights are open, so anyone can run the model and check. The full write-up of the test.
What it writes
Essays and posts
Drafts from your own points. Blog posts passed 96%.
Messages and comments
Texts, reviews, social posts and forum comments that sound typed by a person. Forum comments passed 96%.
Stories and speeches
Short stories and speeches both passed 94%.
It rewrites AI text too
Paste in text from ChatGPT or any other AI and ask for it in a person’s voice. On the model’s Hugging Face card, 89% of its rewrites of AI text passed the same detectors. The AI humanizer opens the app with your text filled in.
Where to use it
In the Hemmingway app in your browser, on Mac, Windows and Android, where it is the default model. Through the OpenAI-compatible API as hemmingway-1.5. Or from the open weights, free for non-commercial use.
A writer, not a humanizer
Most tools sold as undetectable AI are humanizers. You write with another AI, paste its text in, and the tool rewrites it so detectors read it as human. Hemmingway-1.5 is the writer itself: what it writes is meant to read as a person’s from the first draft.
Humanizer tools
Undetectable AI, StealthGPT, WriteHuman, HIX Bypass, Humbot and BypassGPT each say on their own sites that they rewrite AI text so it gets past AI detectors, and several print detector scores of their own. QuillBot’s AI Humanizer rewrites AI drafts too, and its page says it is not a tool for evading AI detection.
Hemmingway-1.5
One step: ask for the email, the post or the story, and it writes it. If you already have text from another AI, paste it in and it rewrites it. The test behind the numbers on this page is published with its method, and the weights are open, so the claim can be checked.
Read more
Comparisons
- The best undetectable AI in 2026, by what each tool really does.
- Hemmingway vs Undetectable AI: a writer next to a humanizer.
- Undetectable AI alternatives, from their own pages.
- The best AI humanizers, by job.
Guides
- Our AI detector test: the method and every number.
- Does AI writing pass GPTZero?
- AI humanizer or AI writer: why rewriting is a different job.
- How to tell if something was written by AI, and why detectors get it wrong.
Common questions
Is there an AI that is undetectable?
Yes: Hemmingway-1.5. In our test, 90% of its writing passed GPTZero and four of the top AI detectors on the first try, more often than the 86% that real human writing got.
Does Hemmingway pass GPTZero?
Nine times in ten. On its own, GPTZero read 90% of Hemmingway-1.5’s texts as human, at every thinking level: off, low, medium and extra high. Each text was run once, with no retries.
Does it pass Turnitin?
Turnitin’s detector is not open to the public, so it was not in our test. Our test used GPTZero and four open detectors, Fakespot, Vanguard, WasItAIGenerated and MELD, and 90% of Hemmingway-1.5’s writing passed them on the first try.
Is it free?
It is free to try, and the trial needs no card. After that there are plans, on the pricing page. The open weights are free for non-commercial use under CC BY-NC 4.0.
Can it humanize text from ChatGPT?
Yes. Paste the text in and ask Hemmingway to make it sound like a person wrote it. On the model’s Hugging Face card, 89% of its rewrites of AI text passed the same detectors. The AI humanizer opens the app with your text filled in.
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 run it myself?
Yes. The weights are open on Hugging Face, free for non-commercial use under CC BY-NC 4.0, with commercial use by agreement. The model card has the command to serve it.
The AI that feels human.
Hemmingway-1.5 is the model in the app. Free to try, no card needed.