MIT's Rosalind Picard: top frontier AI models meet zero of NIST's 7 trust principles
RosalindPicard · x · 2026-11-19
MIT professor and affective-computing pioneer Rosalind Picard asks whether any frontier AI has been shown to exhibit trustworthy behavior per NIST's seven principles—arguing the biggest, best-known models meet ZERO of the seven.
- Reliability (#1): asking the same question 100+ times reveals inconsistency; engineers have further failing tests on this and the other criteria
- Explainability (#5): explanations are typically generated with randomness and wording engineers have shown doesn't match the model's actual decision criteria—harmful mainly where reasoning matters (medical advice, major business decisions, product engineering)
She stresses this doesn't mean big models aren't useful, but users should get informed about which uses can and cannot be trusted.
More from Models
- Gemini 4 Neon one-shot a Shibuya runner game starring Google's Pegman — rseroter · 2026-11-20
- Professors built 'AI-proof' exam questions — GPT-5.4 still scored 83% — alex_verem · 2026-11-20
- User tests Opus 5.5 via Claude Pro: after 24h it beats ChatGPT subs on value — mariofilhoml · 2026-11-20
- Google AI declares a border configuration 'mathematically impossible' and proves it — wrongly — conitzer · 2026-11-20
- 13 AI models vs. real policymakers on policy questions: 10 models beat the humans — ChenhaoTan · 2026-11-20
- Frontier models can unmask authorship from a single text, no tools needed — MohitIyyer · 2026-11-20