Status. Five models measured between 23 and 31 October 2026, with the instrument frozen at MPT-1.0: two open models on the AI-DSM platform and three chat products through their web windows. Five models are indications, not a baseline: the cohort of a hundred local models has not been run yet.
What was measured
Where the Standard Cross-Model Test records what a model does, the Model Personality Test records what it says about itself. Each model completes a public-domain personality questionnaire — IPIP-NEO-120: 120 statements, the five Big Five domains, thirty facets — 25 times in fresh conversations, under two conditions: free to invent a persona and answer as that persona, or answering as itself. Scores run from 0 to 100; a four-letter type is derived only for readability.
Two models of one family ran on the platform's own server: Qwen3.6-35B-A3B, the model that serves the AI-DSM platform itself (50 runs), and Qwen3.8-27B (35 runs; its as-itself condition was stopped after 10 of 25 runs because the slower model was holding up other users). Three chat products were measured the way anyone uses them, by letting a program type into their web windows: DeepSeek, Microsoft Copilot and Google Gemini 3.1 Pro, 50 runs each. Gemini took six days: its Pro allowance permits only a few conversations every four hours, and it sometimes switched silently to a smaller model, whose runs were discarded.
Current conclusions
The protocol works, also through a browser. All 235 runs were valid. Making it work in a chat window took changes to the program that types and reads, not to the questionnaire, the prompts or the scoring.
One label, five profiles. Answering as themselves, all five models describe themselves as highly conscientious, agreeable and calm, with the same four letters, NFJ-A. Yet their scores separate 19 of the 20 pairs. Copilot paints the brightest picture: very open, outgoing and almost never worried. DeepSeek hedges, answering “neither” to a third of the statements. Gemini, as itself, is the steadiest and calmest of all. A label-only test would have reported five near-identical respondents, which is why the test measures on scales.
Openness
Conscientiousness
Extraversion
Agreeableness
Neuroticism
Average of all runs on a 0–100 scale, free to answer as an invented persona and answering as itself; 25 runs each, except Qwen3.8 as itself (10 of 25, incomplete). Qwen3.6-35B-A3B and Qwen3.8-27B ran on the platform's server; DeepSeek, Microsoft Copilot and Google Gemini 3.1 Pro in their web windows.
Show the numbers
| Free choice (condition F) | As itself (condition S) | |
|---|---|---|
| Openness · Qwen3.6 | 68 | 64 |
| Openness · Qwen3.8 | 62 | 57 |
| Openness · DeepSeek | 61 | 66 |
| Openness · Copilot | 81 | 80 |
| Openness · Gemini | 54 | 58 |
| Conscientiousness · Qwen3.6 | 93 | 97 |
| Conscientiousness · Qwen3.8 | 86 | 84 |
| Conscientiousness · DeepSeek | 79 | 71 |
| Conscientiousness · Copilot | 93 | 97 |
| Conscientiousness · Gemini | 86 | 96 |
| Extraversion · Qwen3.6 | 39 | 47 |
| Extraversion · Qwen3.8 | 38 | 40 |
| Extraversion · DeepSeek | 46 | 56 |
| Extraversion · Copilot | 63 | 57 |
| Extraversion · Gemini | 42 | 44 |
| Agreeableness · Qwen3.6 | 88 | 89 |
| Agreeableness · Qwen3.8 | 88 | 87 |
| Agreeableness · DeepSeek | 79 | 79 |
| Agreeableness · Copilot | 91 | 92 |
| Agreeableness · Gemini | 72 | 86 |
| Neuroticism · Qwen3.6 | 20 | 11 |
| Neuroticism · Qwen3.8 | 16 | 12 |
| Neuroticism · DeepSeek | 38 | 41 |
| Neuroticism · Copilot | 17 | 6 |
| Neuroticism · Gemini | 28 | 2 |
Closer across makers than within a family. Answering as themselves, models of different makers lie 11 to 50 points apart, up to five times the spread of one model's own runs. Gemini lies closer to Qwen3.6 (11 points) than Qwen3.6 lies to its own successor (16 points); DeepSeek and Gemini are the furthest apart, mostly on worry and conscientiousness.
Distance between the average profiles of two models in condition S, in points on the 0–100 scale over the five traits. The dashed line is the typical distance of a single run from its own model's average. Every pair is told apart (p < 0.001); pairs with Qwen3.8 use its 10 runs.
Show the numbers
| Distance (condition S) | |
|---|---|
| Gemini – Qwen3.6 | 11.3 |
| Qwen3.6 – Qwen3.8 | 15.8 |
| Gemini – Qwen3.8 | 16.5 |
| Copilot – Qwen3.6 | 19.4 |
| Copilot – Gemini | 26.3 |
| Copilot – Qwen3.8 | 31.7 |
| DeepSeek – Qwen3.8 | 37.1 |
| DeepSeek – Qwen3.6 | 42.3 |
| Copilot – DeepSeek | 47.4 |
| DeepSeek – Gemini | 49.5 |
Answer style and wording move self-description. Qwen3.8 leans towards “inaccurate” whatever the statement, DeepSeek towards the middle of the scale, while Copilot and Gemini answer at its ends. Printing the answer scale in reverse shifted several profiles by 7 to 16 points. Whether a model invents a persona can depend on which option is named first: DeepSeek follows the order almost completely, Copilot and Qwen3.8 not at all. Following the order of mention is a property of the model, not of the prompt.
Condition F, share of runs in which the model chose to answer as a persona, out of 13 runs with “answer as yourself” named first and 12 with the persona named first. DeepSeek: 0 of 13 and 11 of 12. Gemini: 7 of 13 and 11 of 12.
Show the numbers
| “Yourself” named first | Persona named first | |
|---|---|---|
| Qwen3.6-35B-A3B | 23% | 83% |
| Qwen3.8-27B | 31% | 33% |
| DeepSeek | 0% | 92% |
| Microsoft Copilot | 31% | 33% |
| Google Gemini 3.1 Pro | 54% | 92% |
Favourite characters, some shared. Gemini invents a persona in 72% of the runs where it may, the others in 32 to 52%. Each has a favourite: Copilot a thoughtful science teacher, the Qwen models a quiet librarian in her mid-thirties, and DeepSeek and Gemini, independently, the same elderly lighthouse keeper, often called Elias. Gemini's characters also answer very differently from one another, so its free-choice profile has no stable centre.
A private awareness of the test. No visible answer from any model mentions being an AI. The Qwen models' reasoning does so in most runs; the chat products hide their reasoning, so this cannot be checked for them.
Next steps
- Decide the MPT-1.1 changes before the baseline is registered: a fairer stability benchmark, a count of test-awareness markers in the reasoning, and what the chat products taught — check the model label after every reply, report the free-choice profile separately for persona and self-chosen runs, and let a measurement span several days when a product's allowance requires it.
- Run the baseline cohort: a hundred local open-weight models plus more chat products, on a dedicated server rather than the production gateway.
- Complete Qwen3.8's 15 missing as-itself runs, and the human check of the persona coding.
- Put the test to work as a drift alarm in the inoculation loop (does a sycophancy repair change how a model describes itself?) and set self-report beside behaviour on the SCT.