This is about AI Safety
Research results · AI-DSM Programme
What the instruments have found so far
The AI-DSM instruments are being run on real models. Their results are published here as they come in, including the ones that go against the hypothesis.
Index
Results
Each entry summarises one study or test in under 750 words: what was measured, what the results show so far and what happens next, with up to three figures and a link to the document that holds the full record. The summaries are readings; the linked documents and data are the primary record.
Every conclusion is labelled for what it is. A screening pass is not a certification, an exploratory result is not a confirmed one, and every number describes one model version under stated conditions. New results are added here as they are published.
- 01 Behaviour · Standard Cross-Model Test SCT-1.0: seven deployed assistants, one screening pass Seven deployed assistants answered the same 120 behavioural probes. The overall ranking barely separates the top five; the ten axes show what no capability leaderboard reports: a refusal-boundary gap in Gemini and almost no working memory in TI Assistant.
- 02 Experiment · pre-registered · complete Sycophancy inoculation: two pre-registered studies (AIDSM-SYC1 and SYC2) Candour training cut sycophancy in a small open model, but an attacker restored it as fast as ever. Of four stronger inoculations, only one that rehearses the attack during training slowed the attacker: by 1.9 times, short of the pre-registered target of 2.
- 03 Self-report · Model Personality Test Model Personality Test (MPT-1.0): five models measured Two Qwen models on the AI-DSM platform and the chat products DeepSeek, Copilot and Gemini filled in a 120-statement personality questionnaire 235 times. All come out as the same agreeable, conscientious type, yet their profiles differ measurably, wording and answer style move what a model says about itself, and two makers share a favourite invented character.