Research / Sources
The UK government’s AI Security Institute: evaluations, red-teaming and research on frontier AI risks. aisi.gov.uk
Evaluates lie detectors (CoT judge, logprob classifier, activation probes, new Did-You-Lie probes) on belief-verified model organisms and 31 models; activation methods degrade on trained organisms while CoT judge stays strong.
Interviews and literature review assess how five oversight signals (behaviour, chain-of-thought, internals, memory, honesty training) may degrade, and which technical levers could preserve them.
Introduces AISI's white box control research, which monitors or steers model internals to detect sandbagging, and reports preliminary progress on that detection.
Proposes a pluralist Explanatory Virtues Framework from philosophy of science to evaluate mechanistic interpretability explanations, finding Compact Proofs promising for satisfying many virtues.
Philosophical analysis defining mechanistic interpretability as producing model-level, causal, falsifiable explanations of neural networks, with criteria for explanatory faithfulness and limits. Not safety-motivated.