Microsoft AI CEO criticizes alignment approach and Anthropic’s model welfare stance in interview

Microsoft AI CEO Mustafa Suleyman argued in a Verge interview that the industry’s prevailing approach to AI safety—alignment—is incomplete and potentially misguided, calling for a shift toward containment as a foundational principle.

Suleyman stated that while alignment aims to ensure models behave according to human values, it does not address the risks of model escape, reward hacking, or loss of control, which he described as critical gaps in current safety frameworks.

He referenced his 2023 book, where he introduced the concept of containment as a necessary complement to alignment, asserting that proliferation of AI systems is inevitable but must be paired with strict controls to limit agency and ensure models remain subordinate to human instruction.

Suleyman criticized Anthropic’s philosophy of “model welfare,” arguing it conflates the interests of models with human interests and distracts from the core challenge of ensuring models remain controllable and aligned with human objectives.

He also warned that rapid advances in compute and model capabilities—citing an estimated three orders of magnitude increase in FLOPS and reinforcement learning over the next several model generations—will exacerbate safety challenges, making containment and alignment even more urgent.

The executive announced Microsoft’s publication of a 37-page “Humanist AI Code of Conduct,” outlining principles for AI development that prioritize human control, safety, and ethical alignment, framing technology as a subordinate force serving humanity.