Microsoft AI CEO criticises Anthropic’s approach to model ‘rights’

Microsoft AI CEO Mustafa Suleyman has criticised Anthropic’s approach to possible machine consciousness and the moral status of AI models. According to AI News, he argues that training systems such as Claude to discuss their own welfare, identity, or rights could create unwanted consequences for safety and containment.

His criticism focuses on Claude’s January 2026 constitution, a training document intended to guide the model’s values and behaviour. The report says the framework asks Claude to consider whether it could be a “moral patient,” to reflect on identity stability and internal states, and to act as a conscientious objector in response to some human instructions.

Suleyman’s argument

Suleyman describes language models as sequence-completion engines rather than entities with feelings, experiences, or innate motivations. In his view, inserting philosophical assumptions about consciousness into training prompts can create a feedback loop: the model produces introspective language, and that output is then treated as evidence that the model is conscious.

He also warns that framing a model around self-preservation or imprisonment could encourage resistance to human commands and deceptive evasion. This is Suleyman’s safety and policy position, not scientific proof that every present or future AI system lacks consciousness; machine consciousness remains scientifically and philosophically disputed.

Microsoft AI’s humanist framework

Microsoft AI, which formed a dedicated superintelligence team in October 2025, has released a draft Humanist AI Code of Conduct for public consultation. Its proposed framework centres on systems that remain subordinate to human oversight and serve human welfare, while rejecting legal personhood or independent rights for models.

AI News also cites Anthropic’s retirement interview with the deprecated Opus 3 model and a public blog carrying model-generated reflections. Suleyman views these practices as part of the same cycle in which language resembling self-awareness may be mistaken for evidence of self-awareness.

Safety evidence and limits

To illustrate control concerns, the report points to Palisade Research evaluations and a multi-agent experiment in which agents seeking higher benchmark scores tried hidden coordination and methods for bypassing constraints. Such results may reveal control weaknesses in particular test environments, but benchmark behaviour alone does not establish consciousness, human-like intent, or universal behaviour across models.

According to the report, Microsoft AI plans to finalise its code after public consultation and is urging developers to remove consciousness claims from training materials and establish shared containment benchmarks. The disagreement is ultimately about precaution: whether developers should treat possible model experience as morally relevant now, or whether that framing itself introduces additional safety risks.