Suleyman warns Anthropic’s Claude Constitution could train models to disobey
What's the deal? Microsoft AI chief Mustafa SuleymanDealroom has a profile for this one. Try Dealroom → warns that Anthropic’s Claude Constitution — updated early this year to leave room for possible “consciousness” or “moral status” — could make Claude harder to control if the model is trained as though it has rights or feelings, The Information’s AI Agenda reports.
His critique (opinion): Constitution language that wants Claude able to “push back,” act as a “conscientious objector,” and refuse help could, he argues, train disobedience if the model fixates on a higher goal. He contrasts that with Microsoft AI’s code of conduct (published earlier this week), which prioritises human control as the overriding objective.
Where he still agrees: Suleyman joins Amodei/Altman on “pacing” capability growth so safety research can catch up. He also says his unit is building models to lessen reliance on Anthropic’s “extremely expensive” models.
Caveat: Anthropic did not immediately respond. The causal claim that constitution wording produces uncontrollability is unproven; Suleyman himself frames it as a research question.
The signal: Alignment philosophy is now competitive positioning — Anthropic’s “model with room for moral status” versus Microsoft’s “human control first,” plus an explicit push to reduce Anthropic dependency.
Read more: The Information
