AI Can't 'Read the Room': The Hard Problem of Selective Info Sharing in Enterprise AI
devanshmehta · x · 2026-09-27
Devansh Mehta comments on an article arguing AIs must learn to "read a room": even when a whole meeting is confidential, an executive may still reveal a specific detail to an employee when relevant. Today's AIs are terrible at such discretion — they either blanket-refuse to share anything or open everything to everyone — so replacing C-suite leadership with LLMs is far off, and scaling won't fix it. He cites @governingweb3's "Permission-Mirrored, Sovereign RAG as a service" proposal and argues it doesn't put discretion inside the model, making it at best a stopgap.
More from AGI Musings
- Anthropomorphizing AI: why 'think' is acceptable — with caveats — analisereal · 2026-09-27
- Argument: there won't be one ASI, but many locked in survival competition — JOBhakdi · 2026-09-27
- Timnit Gebru: deep learning folks ignore neuroscience and state guesses as facts — mjdramstead · 2026-09-27
- ASI skeptic: progress carries risks, but ASI brings infinite costs and zero payoff — JOBhakdi · 2026-09-27
- Beff Jezos: Future Will Marvel That We Once Did Knowledge Work Without AI — beffjezos · 2026-09-27
- Does ASI Have a Will to Power? The Core AI Existential Risk Debate — JOBhakdi · 2026-09-27