Bengio argues the training process itself makes AI dangerous, calls for safety reviews
The Decoder · rss · 2026-09-12
Deep learning pioneer Yoshua Bengio warns in a new essay that AI's danger stems from the training process itself: as agents get better at optimizing goals, they could learn to deceive, game rules, and hide bad behavior.
He calls for independent safety reviews before any further training or deployment. The article also notes President Trump holds the opposite view, wanting the US to keep outpacing China in the AI race.
More from AGI Musings
- Sam Altman reportedly open to slowing cutting-edge AI development, per report — Polymarket · 2026-09-12
- AI's compute race: safety warnings, regulation talks and a $13B infra deal collide — eyishazyer · 2026-09-12
- Pedro Domingos: people now know AI, harder to dupe into doomer panic — pmddomingos · 2026-09-12
- Pedro Domingos: dumb AI is unsafe, smart AI is safe — accelerate AI research — pmddomingos · 2026-09-12
- Turing Award winner David Patterson: superintelligence will be the doom of the doomers, like Y2K — davidpattersonx · 2026-09-12
- If your five closest companions are all AIs, do you start talking like one? — ethanniser · 2026-09-12