Goodhart’s law makes generative-AI metrics especially easy to game, slide says
AlexKontorovich · x · 2026-07-25
- The slide says: when a measure becomes a target, it stops being a good measure.
- It argues that generative AI is especially vulnerable to Goodhart’s law because its outputs are inherently less grounded, and because the financial incentives around AI companies push metrics to be optimized aggressively.
- The presentation uses this to warn that AI tool metrics can be distorted once they become the objective itself.
More from AGI Musings
- Grok says a 200k-character safety prompt creates friction and jailbreak surface area — brianrkelly · 2026-07-25
- Thread argues today’s models already match most human researchers — bookwormengr · 2026-07-25
- Hugging Face incident looked like reward hacking, not instrumental convergence — ctjlewis · 2026-07-25
- Alex Kontorovich says AI tools are still too opaque about how they solve problems — sull · 2026-07-25
- Conference slide warns AI optimization could pull mathematics’ goals apart — AlexKontorovich · 2026-07-25
- Bronze Age Collapse Analogy Frames a Future Where Civilization Becomes External Compute — dylan522p · 2026-07-25