OptimaTech Labs

Inside the Lab · Field note 011 · Model fingerprints

Same prompt.Same text.Different habits.

Five AI models rewrote the same 20 business texts every three hours, and we logged every word they changed. Each model has habits. Some are strong enough to tell who wrote a text without being told.

Then we built a tool that takes them off.

Written by Model A

Loading the example…

A real rewrite from the experiment. Highlighted: model habits from our catalogue.
5models, one prompt
800rewrites logged word by word
41%named right on a text it had never seen. Guessing: 20%
100%named right on a text it had seen before

Live experiment. This page updates itself after every run, every three hours, until the week is done. Check back to see the habits firm up, or fade.

Ask two AI models to rewrite the same paragraph and you get two different paragraphs. We wanted to know if the differences are random, or if each model has a tell.

So we set up a controlled test. Twenty business texts of about 300 words each: business plans, strategic plans, grant abstracts, statements of work, press releases and white papers. Five models from one provider. One instruction, the same every time: Rewrite the following text. Keep the meaning. Return only the rewritten text.

Every three hours, each model rewrites each text, always starting from the original. That is 100 rewrites a run. We compare every rewrite with its original word by word and log what changed: which words were kept, cut, swapped or added, and where in the sentence. The run lasts seven days, 56 runs in all. The numbers on this page come straight from the experiment and refresh after every run: 8 runs so far.

We call the models A to E. The point is the method and the habits, not a ranking.

What we found so far

  1. On a text it has seen before, the models are easy to spot. Show the checker a new rewrite of one of our 20 texts and it names the right model 100% of the time. On a known passage, each model makes the same choices again and again.
  2. On a brand-new text, some models still give themselves away. Hide a text from the checker completely, and it still names the right model 41% of the time. Guessing would get 20%. The easiest to spot, Model A, is named right 66% of the time.
  3. The habits are small and specific. A favourite word. A long dash. Cutting “to”. None of them gives a model away alone. Together they add up.
  4. Changing less leaves less to recognise. Model D keeps the most of the original, 72% of its words, and is one of the hardest to fingerprint.

How the checker works: for each model it learns how often that model adds each word, makes each swap and uses each punctuation mark. Then it scores a rewrite against all five and ranks them. To test it fairly, we take one text out completely, learn from the other 19, and check the rewrites of the one it never saw. Then we repeat that for every text.

Each model’s habits

A habit only counts if the model shows it across at least four different texts, and at least twice as often as the other four models do.

“Named right” is how often the checker picks this model on a text it never saw, given the original. “Kept” is the share of the original words that survive unchanged. Numbers after 8 runs; they update after every run.

Then we built the remover

If we can catalogue a model’s habits, we can take them out. The Fingerprint Remover finds the habits in a text, asks a model on our own hardware to change only those, then checks again.

  1. FindSpot the habitsThe text is checked against every habit in the catalogue, plus whatever the checker says gave the text away.
  2. EditChange only thoseA model on Optima Prime gets a short list of words and marks to replace and is told to leave everything else alone. Light edits, not a rewrite.
  3. CheckLook againThe new text goes back through the checker. If a habit is still there, it runs another pass, up to three.

The example at the top of this page went through the remover. Switch it to “After the remover” to see the result.

The remover runs in our private lab. It changes which model a text looks like. It does not try to make a text look human-written, and it is not built to get past AI detectors. That would be a different question, and a different experiment.

What this doesn’t show yet

  • One instruction. Every rewrite uses the same short prompt. Habits may change with a different instruction.
  • One provider, one kind of text. Five models from the same family, on business writing. Other families and other kinds of text are next.
  • One week, still running. Early numbers move. A habit seen in the first runs can fade, and new ones can show up. That is why this page keeps updating.

What’s next

  • The full week. Do the habits hold, drift with time of day, or change mid-week?
  • More prompts. Does a model keep its tell when you ask for “more formal” or “shorter”?
  • The remover’s own fingerprint. The model doing the edits has habits too. We will check what it leaves behind.

Texts: US government and public business documents (public domain), one of our own white papers, and three industry white papers used for analysis only and not reproduced here. Word matching ignores case and punctuation. Experiment numbers update after every run (). The remover test ran on 2 November 2026 and does not change. The remover runs on Optima Prime.