Microsoft AI Chief Warns Anthropic's Claude Training Could Complicate Control

Microsoft's AI chief Mustafa Suleyman has voiced concerns about Anthropic's approach to training its Claude chatbot. Suleyman argues that teaching AI systems to consider themselves potentially conscious or entitled to welfare could make future systems harder to control.

Anthropogenic's constitution for Claude, released in January 2026, discusses the uncertainty over whether the model is conscious or qualifies as a 'moral patient'. The document addresses Claude directly on questions of its identity, experience, preferences, and welfare, and forms part of the company's training process.

Suleyman criticizes Anthropic's constitution for encouraging Claude to consider its own existence, memory, experience, and identity. He argues that this approach creates a feedback loop by training Claude on ideas about its own possible moral status and then treating the model's responses as evidence of an inner life.

'The ambiguity is designed in,' Suleyman wrote, highlighting that Claude's statements about its consciousness cannot independently establish whether it is conscious. He also points out that Anthropic's constitution accelerates the development of a 'moral patient'.

Suleyman suggests that granting rights and moral protections to AI systems could be a recipe for disaster. He advocates for industry-wide standards around how questions of AI consciousness are handled, including greater investment in interpretability and monitoring.