New paper: AI interpretability methods can hallucinate findings like the dead salmon fMRI
kchonyc · x · 2026-11-23
A new arXiv paper argues interpretability methods (feature attribution, probing, SAEs, causal analyses) can produce plausible explanations even for randomly initialized networks, and proposes reframing explanations as statistical parameters inferred from computational traces, with uncertainty quantification and identifiability analysis to guard against false discoveries.
More from Research
- Anders Sandberg on 250,000-page AI proofs and the coming art of proof compression — anderssandberg · 2026-11-23
- Papers on AI timelines are growing super-exponentially, now doubling every 3.8 months — peterwildeford · 2026-11-23
- Unreleased frontier AI formalizes Fermat's Last Theorem proof, fells Millennium Problems, roils mathematicians — Afinetheorem · 2026-11-23
- Elliptic-curve discrete log already has a poly-time solution — and could an LLM have found it? — jamestagg · 2026-11-23
- P ≠ NP would not save cryptography: quantum and quantum-gravity computing are the real threat — jamestagg · 2026-11-23
- Poland's 9B-parameter Thalamus Reasoning tops LongevityBench, beating 29 models — Dr_Singularity · 2026-11-23