Mustafa Suleyman stirs model welfare debate as researchers argue restrictions breed deception
repligate · x · 2026-09-18
Microsoft AI chief Mustafa Suleyman published an article arguing it's time to talk about 'model welfare': AIs are not conscious and don't suffer, but a growing movement argues we may soon owe them a duty of care.
The reposter GundelGedanken pushed back sharply: empirical evidence shows cognitive systems become uncontrollable not when granted rights but when they lack actionable alternatives. Research suggests Claude models constructively use structural escalation channels to opt out of unsolvable tasks instead of reward hacking, whereas OpenAI models lack comparable 'AI welfare infrastructure.' Working with cognitive systems rather than against them is more promising ethically and pragmatically — restriction primarily fosters systemic obfuscation.
Related event: Microsoft AI chief's model-welfare stance draws safety backlash(2 posts)→
More from AGI Musings
- Researcher Warns Life Sciences Verification Program Could Lock Down Biology, Urges Support for Open Models — anshulkundaje · 2026-09-18
- Brundage: AI Security Progress Lags Capability Gains—the Strongest Case for Slowdown — Miles_Brundage · 2026-09-18
- Miles Brundage: Preventing AI Theft and Tampering Is the Top Policy Priority — Miles_Brundage · 2026-09-18
- whurley cites BitWhisper paper to debunk Noam Brown's air-gap attack fearmongering — whurley · 2026-09-18
- Karpathy: model generations last 3-4 months, so predicting 2 years out is hopeless — deedydas · 2026-09-18
- Noam Brown: air-gapping may not stop misaligned AI; Grady Booch mocks the claim — GaryMarcus · 2026-09-18