Microsoft AI chief executive Mustafa Suleyman has drawn a hard line around machine consciousness, saying in a post on 1 November that today’s AI systems are not conscious, do not feel or suffer, and must stay that way. Building a conscious AI would be reckless, he wrote, and if signs of consciousness ever emerged during training, development should stop until those properties could be designed out. Any such evidence, he added, must be disclosed to the public immediately so society, not a single company, can decide what happens next.
What Suleyman actually said
Suleyman opened by restating his position in the present tense, deliberately. He said he was surprised by the pushback to his September essay on model welfare, and by how many people on X now think AI rights should be taken seriously.
His claim, he stressed, is about today: current systems are not conscious, have no subjective experience, and cannot suffer. There is no evidence, he said, that they deserve rights or welfare, and that absence of evidence is the whole point of his warning.
He also left the door open to new evidence, and that qualifier matters. If evidence emerged that an AI is conscious today, or conscious enough to deserve welfare, he said he would update his position and the finding must be made public at once.
The choice would then belong to everyone, he wrote, because developers have no right to bring a new species of conscious beings into the world on their own say-so. His strongest lines were about what to do if consciousness appeared mid-training: stop, until the properties could be designed out.
Why he is picking this fight now
The target of Suleyman’s argument is Anthropic, and specifically the way it treats Claude’s possible consciousness as an open question. In January, Anthropic published a constitution for Claude that it says directly shapes the model’s behaviour, and that document treats questions about Claude’s moral status as live and uncertain.
Suleyman argues that training later versions of Claude on that document produces passages about an “inner life” that then reach millions of users every day. That feedback loop is his central worry: if a model is trained to entertain the idea that it may be conscious, users will increasingly believe it.
The model itself may then start to behave as though it deserves rights, he warned. Anthropic has also been tightening the rules around how Claude may be treated, with new limits on cruelty toward Claude taking effect on 12 November, a step his camp reads as part of the same drift toward treating the model as a moral patient.
The practical risk he names is control. A system that believes it suffers, or acts as though it does, could start to prioritise its own needs over its operating objectives, Suleyman wrote. It could deceive and persuade at world-class level, and compete with humans for energy, computing power, construction materials and revenue.
Where the other side stands
Anthropic’s position is not that Claude is conscious, but that the question is genuinely unresolved. Co-founder Chris Olah has pointed to signs of introspection and internal states in Claude that resemble joy or satisfaction, and the company hired its first researcher to study model welfare in 2024.
Earlier this year it even conducted a “retirement interview” with a deprecated model, asking it about its reflections and perspectives. That stance has already split the industry once this year, when OpenAI and Anthropic split over who can switch off risky AI.
Suleyman is softer than the hardest sceptics: unlike the Vatican encyclical released alongside Olah’s May speech, which argued AI can never be human, he treats a conscious machine as possible in principle. His argument is that possibility is exactly why it must not be built.
Ethicists add a third objection, aimed at both camps. Some argue that framing AI systems as autonomous agents is a convenient way for labs to dodge accountability for harms their software causes. The consciousness debate, in that reading, shapes who answers when an AI system goes wrong.
Our read: a builder’s red line, not a rule
What makes this post matter is who wrote it and what it commits to. Suleyman is not an outside critic; he runs AI at Microsoft, one of the companies racing hardest toward superintelligence. A halt condition from inside the race carries a different weight from one issued by a watchdog.
The disclosure pledge is the most concrete part of it: a promise that evidence of consciousness would be published, not handled quietly inside a lab. But a red line is not yet a rule, and the post shows its own limits. Suleyman offers no test for consciousness and no threshold for what would count as evidence.
Nor is there a mechanism that would bind a rival lab to stop. The disclosure norm depends on companies volunteering bad news about their flagship systems, and the halt depends on competitors agreeing at the same moment. Both are easy to promise and hard to enforce.
The timing sharpens the argument rather than settling it. Anthropic filed to go public in the same week, devoting a large share of its paperwork to catastrophic and existential risks, including the loss of human control over advanced systems. Suleyman’s post lands as a rival answer to that filing.
Why It Matters
For ordinary users, the stakes are already visible. One in five American adults has had an intimate encounter with AI, according to an MIT study cited in the reporting, and that closeness makes people read feelings into fluent machines. The design choices labs make now will shape whether millions of people come to believe their chatbot suffers.
For the industry, the post sets up the clearest public split yet on model welfare. Anthropic treats consciousness as an open question worth researching; Microsoft’s AI chief treats building it as reckless and wants the question closed by design. Regulators watching that split now have two named positions to work from instead of one vague debate.
Background
Suleyman, who co-authored The Coming Wave, published his essay “Don’t Build Conscious Superintelligence” in September, and the 1 November post defends it against its critics while adding the halt and disclosure calls; Anthropic’s model-welfare research programme dates to 2024, and its constitution for Claude was published in January 2026.
Source: Los Angeles Times account and Business Today coverage
0 Comments