Today we're introducing personal models: AI models trained for one person, on their own writing.
Every frontier model is trained once, for everyone, so what it writes sounds like the model. A personal model learns how one person writes. We tested ours on 50 writers against Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon, each given 50 of the person's texts to learn from.
A blind judge could not tell our models' writing from the person's own. An AI detector read all of it as human.
It writes like the person
We followed the protocol of PersonalBench, a benchmark for writing as a specific person,1 on a licensed set of 50 writers. Each system wrote new texts as each writer from a one-line summary of something the writer had actually written. A blind judge read three of the writer's real texts and the new one, and decided whether the same person wrote it.
An authorship-verification model, built only to tell one writer from another, agrees. It placed our models' writing next to the person's own and every frontier model well behind.3
It reads as human
Most people check whether writing came from AI with a detector, and Pangram is the most accurate one we know of. We sent it one text per writer from each system. It flagged nearly everything the frontier models wrote, even while they imitated a real person from 50 examples. It flagged none of ours.
Why prompting falls short
The usual way to make a model write like you is to paste in examples of your writing. In PersonalBench, every prompting method scored below a random stranger's writing.1 We gave the frontier models 3 of a person's texts, then 50. Seventeen times more writing moved them only part of the way. A personal model trained on 25 texts went further than any of them.
Writing well matters as much as sounding right. In our system, each personal model works with our models for substance and fact-checking, and a blind rater scored that writing 3.70 out of 5, above the people's own 3.11.5
Results
| Evaluation | Personal model | Opus 5.5 | GPT-6 Astra | Gemini 4 Argon | The person |
|---|---|---|---|---|---|
| Blind judge: same writer | 82% | 60% | 60% | 50% | 82% |
| Authorship model (LUAR) | 0.79 | 0.72 | 0.72 | 0.70 | 0.80 |
| Pangram: human | 100% | 4% | 10% | – | 100% |
| Blind judge, personal model from 25 texts | 78% | 60% | 60% | 50% | 82% |
Frontier models were given 50 of the writer's texts. A different person's writing scores 15% on the judge and 0.62 on the authorship model. The personal models' lead over every frontier model holds at 95% confidence on both likeness measures.
- PersonalBench (Sawant, August 2026) measured how well models write as a specific person using LUAR, a published authorship-verification model. Few-shot prompting, profile extraction and contrastive prompting all scored below a random other person's writing.
- 50 writers from a licensed public corpus, 250 new texts per system, written from one-line summaries of held-out texts no system saw. The judge is Claude Sonnet 5.5, shown three of the writer's texts and one candidate, never which system wrote it. Raw rates: the person 82.0%, personal models 82.4% (78.3% from 25 texts), Opus 5.5 59.9% (45.3% from 3 texts), GPT-6 Astra 59.6% (52.4%), Gemini 4 Argon 49.7% (39.2%), a different person 14.7%.
- LUAR similarity to the writer's other texts: the person 0.798, personal models 0.788, Opus 5.5 and GPT-6 Astra 0.718, Gemini 4 Argon 0.698, a different person 0.622.
- Pangram, one text per writer per system. Two of GPT-6 Astra's texts were under the detector's minimum length. Gemini 4 Argon with 50 texts was not scored.
- A blind rater, Claude Sonnet 5.5, scored each text 1 to 5 as a contribution to its discussion; the person's real text was scored the same way. With this system the judge still attributed 75.8% of texts to the writer.