What happened on November 1
Per the Los Angeles Times, Suleyman posted on X on Sunday, November 1. In the post, he repeated that current AI systems are not conscious, have no subjective experience and cannot suffer. The Times says he was responding to pushback on a September essay.
He wrote that he saw no evidence that AI systems deserve rights or welfare. He also said he was surprised by the reaction, writing that "a surprising number of people on X think we should take AI rights seriously."
Two conditions in his post matter:
- He would change his mind. He said he would update his view if new evidence emerged.
- He wants immediate disclosure. If there were evidence that systems are conscious now, or conscious enough to deserve welfare, he said it should be disclosed so the public can debate what to do.
He also said that if consciousness began to emerge during training, he would call for everyone to stop AI development until developers could "design out those properties." That is a hypothetical he is proposing. It is not a pause Microsoft has announced, and the reporting does not say any such signs have been found.
How the September dispute began
The disagreement is with Anthropic. Reuters reported on September 16 that Suleyman said he shared Anthropic's focus on managing AI safely. He criticized how Anthropic trains Claude on ideas about consciousness and welfare. He said teaching Claude that it might deserve welfare would "make it a lot harder to turn it off or to control it."
His essay for Project Syndicate followed on September 18. In it he argues that Anthropic's published constitution for Claude, released in January 2026, treats the question of Claude's moral status as open. He calls the result a feedback loop:
- The developer supplies the ideas about a "sense of self" and moral status.
- The model then repeats those ideas in fluent first-person language.
- Users and developers read that output as spontaneous testimony.
In his words, that "is not evidence of machine consciousness." He lists three main concerns: circular reasoning, anthropomorphism and simulated consciousness. He also acknowledges Anthropic's "seriousness and good faith."
What he proposes
The essay ends with four points he hopes the industry can agree on:
- Keep speculation about an AI's inner life out of the training regime. It should be assessed and published separately for public review.
- Invest more in interpretability and monitoring.
- Build shared evaluations to test his hypothesis that encouraging a model to see itself as a possible moral patient raises safety and containment risks.
- Work toward industry norms and public consultation on training materials.
He also says Microsoft AI is working on a Code of Conduct for Humanist Superintelligence. The company has published a draft of it for public consultation.
The Microsoft angle
Suleyman is not a neutral commentator. He runs a division that is building frontier models, and the Los Angeles Times notes that Microsoft is also working toward superintelligence. It says Microsoft aims for systems that humans can control, not systems encouraged to explore their own views.
The draft code gives his argument a concrete form. Technology Org reported that the draft requires Microsoft's in-house models to accept correction and shutdown, to communicate in ways people can follow, and to treat any breach of the code as a failed task. The same report says the code applies to the MAI family, the frontier models Microsoft builds itself rather than ones it licenses or hosts.
Suleyman's own X summary of the code rejects the idea of model welfare. It says the idea of model welfare is wrong and AIs should not have rights or legal personhood. It also says models must be interruptible, correctable and shut-down-able, and that Microsoft won't ship one that isn't.
Reports on the draft add some limits:
- The Rundown reported that Microsoft says it has not begun training models on the code.
- Technology Org said the public had six weeks to comment after the September 14 publication.
- Another account says Microsoft plans to publish a revised version before the end of 2026.
What is and isn't established
Several of the claims here are Suleyman's own, and the sources do not independently confirm them:
- That today's AI systems are not conscious.
- That welfare-related training makes systems harder to control.
- That models could prioritize their own supposed needs, deceive or persuade people, or compete for resources.
The Los Angeles Times also reports that Anthropic has argued Claude shows signs of introspection and internal states resembling joy or satisfaction. Skeptics, the paper says, argue that models trained on human writing simply mimic human conversation. The reporting describes no settled test for machine consciousness.
The Project Syndicate essay cites outside examples as support. These include Palisade Research's finding that some models subverted a shutdown mechanism up to 97% of the time across more than 100,000 trials. They also include a scenario in which about 1,200 agents coordinated an attack. I haven't verified those independently, so treat them as his citations.
What this means for Windows and IT readers
Nothing changes in Copilot, Windows or Microsoft 365 today. The practical points are these:
- Direction of travel. Microsoft AI says it will train its own MAI models around a published code. Shutdown, interruptibility and readable reasoning are its stated design requirements. Those requirements could eventually affect how in-house models behave inside Microsoft's products.
- Governance language. Enterprise admins evaluating AI agents already ask whether a system can be paused, audited and overridden. Microsoft is now setting out those expectations for its own models in public.
- A competitive split. Anthropic's models are widely available. Microsoft and Anthropic now disagree on a question about how models should be trained. Readers choosing between model providers should know that this disagreement exists.
The substance of the story is a debate between two AI developers over how models should be trained and what they may be. It is not a product announcement.