Methodology: how the Institute tests AI systems

A test that cannot be repeated is worth nothing, so we publish what we measure, with which instruments and inside which limits — including what the method does not cover.

A method is a scheme you can repeat A method is a scheme you can repeat
A method is a scheme you can repeat

Methodology is a condition of the work rather than an appendix to it: the five working principles, the public position on six questions, the security assessment method and the behavioural audit protocol all lie open for reading.

Each method names its instruments and its boundaries. A report says what was run, in which configuration, over how many episodes, and which part of the result is a measurement rather than an interpretation.

Where we have no answer, we write “not tested” rather than “safe”. That is not caution for its own sake: it is the difference between a report and a promise, and the reader is entitled to see which one they hold.

Materials on this topic

Other topics

AI red teaming: what a test shows, and what it does notAI standards: where the rules actually applyAI supply chain: what enters before the model runsMulti-agent systems: behaviour between the agentsIncident chronicle: documented consequences, not forecastsBehavioural audit: how agent behaviour drifts over timePrompt injection: where text becomes an instructionRussian practice: who is held responsible for AIData and benchmarks: what our conclusions rest onSociology of AI: relations between the agentsGovernment customers: how to commission a testGovernance and funding of the InstituteCases: scenarios of work and what they show

Page updated 02 November 2026 · site build 2026.11.13 · ANO «ISAIS»