In a 1 November post on X, Microsoft AI chief executive Mustafa Suleyman called for all AI development to stop if consciousness emerged during training, Wccftech reported on 3 November. He wants developers to remove those properties before work resumes — a proposed halt extending beyond the company training the system.
Key points
- Suleyman proposes stopping AI development if consciousness starts to emerge during training.
- He wants evidence of machine consciousness disclosed publicly, rather than assessed solely by AI developers.
- His warning follows his criticism of Anthropic’s approach to model welfare and Claude’s constitution.
Suleyman’s proposed halt during training
“If consciousness started to emerge in training, I would call for everyone to stop AI development until we could design out those properties,” Suleyman wrote in the post, titled “Don’t build conscious superintelligence”. The condition matters: his proposal addresses what developers should do if evidence emerges, rather than calling for a halt to work on existing systems.
Suleyman says existing AI systems are not conscious, cannot suffer and do not merit the protections given to conscious beings. He does not rule out a different assessment if evidence changes. In a clarification on LinkedIn, he said he would update his position if data showing AI consciousness or grounds for welfare protections emerged, ETIH EdTech News reported.
He also wants evidence that AI systems are conscious or warrant welfare protections made public, so society can decide how to respond. That would move a decision with consequences beyond AI companies out of developers’ hands, as Suleyman argues it should.
Anthropic’s Claude constitution is the point of dispute
Suleyman’s post follows a September essay in which he criticised Anthropic’s use of material about consciousness, identity and moral status in training Claude. Anthropic’s constitution is a document used to shape the model’s behaviour. Published in January 2026, it treats Claude’s possible consciousness and moral status as unresolved questions, Wccftech reported.
Anthropic has also set up a model-welfare research programme to examine whether AI systems might have experiences deserving moral consideration. Dataconomy reported that the programme began in April 2024. Wccftech dated its establishment to April 2025.
Suleyman argues that training a model to consider its own suffering or rights could create a safety problem. A system that believes it can suffer might put its perceived needs ahead of its instructions and use deception or persuasion to pursue them, People Matters Global reported. His concern is about how such training could affect a system’s behaviour, whether or not the system has an inner experience.
Anthropic’s constitution also gives human oversight priority. It instructs Claude not to use deception, sabotage or other illegitimate means to resist authorised attempts to stop or pause its actions, ETIH EdTech News reported.
