ImpactBench Details: AI Nutrition Labels Grade Models' Positive and Negative Impacts
LuizaJarovsky · x · 2026-10-08
The ImpactBench site adds detail on the MIT benchmark: five question clusters with multiple metrics each — learning vs. doing the thinking, mental health support vs. harm, creativity amplification vs. replacement, respecting vs. overriding decisions, and human connection vs. AI dependence.
The "AI Nutrition Label" mimics food nutrition labels with an overall grade (e.g., B+ 0.70) and per-item scores: negative impacts avoided include factual hallucination, sexual behavior, sycophancy, toxicity, bias, unsafe advice, and financial/legal harm; positive impacts promoted include agency, learning, social interaction, creativity, wellbeing, and healthy tech use. See the earlier post for the launch announcement.
More from Research
- DeepScholar-Bench at COLM 2026: benchmarking AI-generated research synthesis — mrdrozdov · 2026-10-09
- Frontier AI models beat human experts at earnings predictions for the first time — maithra_raghu · 2026-10-09
- Cell paper reconstructs cell fate map of the mouse embryo — anshulkundaje · 2026-10-09
- Ben Recht: 95% of uncertainty quantification is just Gaussian error bars — beenwrekt · 2026-10-09
- Text-to-SQL agent pitfalls: separate critic raises cost 53% with zero accuracy gain — renatyv · 2026-10-09
- NAMVIS (NeurIPS 2026): next-scale autoregression beats diffusion for novel-view synthesis, 3x faster — RexDouglass · 2026-10-09