Research and publications
Understanding changes in AI behaviour.
FASO’s research explores changes in AI behaviour and how to build a reliable account of what has happened. Here you can read about our papers and proposed studies, with the current status and limits of each made clear.
Research papers
Paper 1 and Paper 2
Two connected areas of FASO research: how evidence supports a conclusion, and how information can pass between model runs.
A clear basis for interpreting evidence
Separating Observation, Validation, Analysis and Safety State in AI Evaluation
Paper 1 sets out why recording a model’s behaviour, checking the evidence, interpreting it and making a statement about safety must remain separate steps. It presents FASO’s proposed approach, supported by author-reported implementation and internal testing. The manuscript is complete; it does not claim independent validation or prove that an observed model is safe.
Information carried between model runs
Causal Information Transfer Across Language Model Runs: A Gated, Auditable Framework
Paper 2 investigates whether information gained during an evaluation can persist outside one model run and influence a later run, including a different model. It examines how to distinguish the effect of transferred information from other explanations. A working manuscript and pilot testing are in place; the research does not yet establish a confirmed causal result.
Further research
Working hypotheses and proposed studies.
Working hypothesis · 27 July 2026
From Instruction Attenuation to Operating-Regime Reversion
Could shortening or losing a governing instruction make an AI system less careful about accuracy as work continues?
Read the complete publicationWorking hypothesis · 27 July 2026
Evaluation-Induced System Drift
Could repeated evaluation help change a system when information from testing is carried forward through memory, updates or changes to working practices?
Read the complete publicationWorking hypothesis · 27 July 2026
Cross-Model Evaluation Propagation and Multiplier Effects
Could information gained from evaluating one model influence other models through shared systems, human changes or later training?
Read the complete publicationValidation protocol · 27 July 2026
Controlled Validation Protocol for Instruction Attenuation
A proposed study comparing different ways of supplying and checking governing instructions, to test whether these affect a model’s ability to follow them accurately.
Read the complete publicationFurther reading
Related research and publication guidance
How the pilot and research fit together
How FASO publishes its research
FASO-RP-001 · Version 1.0 · Method published