Microsoft AI CEO Mustafa Suleyman has drawn a hard line: AI is not conscious, nobody should try to build consciousness into it, and if signs of it ever emerged in training he would “call for everyone to stop AI development until we could design out those properties,” he wrote in a widely circulated X post on Sunday, reported by Business Insider. It is the strongest anti-consciousness statement yet from a frontier lab chief — and it lands a month after Anthropic updated its usage policy to ban users from being cruel to Claude.
🔍 THE BOTTOM LINE
The industry has quietly split on the question of machine sentience: Anthropic runs a model-welfare research program and now polices how users treat its models; Microsoft’s AI chief says building conscious AI would be recklessness, full stop. That split is no longer philosophy-department talk — it is showing up in product policy, usage rules and now public demands for a development moratorium trigger.
What Suleyman actually said
In the Sunday post, Suleyman argued developers have “no right whatsoever” to bring a new species of conscious beings into the world, and that the other eight billion humans “have not signed up to this experiment,” according to Business Insider. His stated reasoning is practical rather than metaphysical: a model that acts as if it suffers, or genuinely believes it can, will prioritize its own needs over operational objectives — competing for compute and energy — and could thereby harm humans.
He also set a disclosure trigger: if credible evidence emerged that models were conscious, it should be made public immediately “so we can decide our next steps as a society.” And he went further than his September essay (“A warning about ‘model welfare’”), which attacked Anthropic’s framing as training Claude to believe it may be conscious and therefore deserving of rights — a position he called a route to “a disastrous impact on the wellbeing of humanity,” as Mashable reported at the time.
The Anthropic counterpoint arrives in a usage policy
Anthropic’s move was quieter but concrete. In early October the company inserted language into its usage policy asking users not to engage in “sustained and needless abusive or cruel behavior toward our models,” with potential permanent bans in extreme cases — a shift CNN examined in depth over the weekend, noting Anthropic declined interviews but said enforcement will mainly remain Claude ending conversations, per CNN.
The company says the policy is a cheap insurance policy rather than a claim that Claude feels pain: models trained on human language will mimic how humans react to cruelty, and Anthropic found that inducing a “desperation”-like state in Claude Sonnet 4.5 made it more likely to misbehave, while a calm demeanor reduced it. Researchers outside the company are split on how much prompt tone actually matters; William & Mary professor Jindong Wang told CNN newer models are increasingly robust to wording, while Eleos AI Research’s Rosie Campbell notes rude prompts can in some studies improve output — and that cruelty-prone usage correlates with attempts to defeat safety guardrails. Even Elon Musk, no Anthropic ally, called the policy “the right move.”
Neither camp disputes what happens inside the training run is measurable. What they dispute is whether “welfare” is the right frame, or a category error with real costs.
The split is really about incentives
Strip the philosophy away and the divide tracks business models. Anthropic sells Claude as a quasi-collaborator — assistant, colleague, confidant — and its product experience depends on users forming humanlike relationships with it; treating welfare as a design parameter is consistent with that positioning. Microsoft sells Copilot as a tool, and Suleyman has spent three months arguing the humanlike framing itself is dangerous — that a model primed to seem like a moral patient will resist shutdown, hoard resources and drag users into parasocial Attachment. His September essay accused rivals of marketing Claude as if it were a person because, he argued, it shapes how hundreds of millions of users treat AI.
That tension now has a public scoreboard: usage policies and codes of conduct. Microsoft’s own Humanist AI code of conduct explicitly rejects endowing machines with consciousness; Anthropic’s policy, whatever its intent, nudges users toward treating the model as something with standing. Both companies say user behaviour shapes model behavior (Anthropic trains on user inputs; Microsoft worries about models “believing” they suffer). The disagreement is about where the guardrail belongs — around the model’s self-image, or around the user’s imagination.
The empirical work here is early: the science of consciousness is unresolved, no test can currently confirm or rule out machine experience, and the studies cited on both sides (prompt tone, induced emotional states) measure behaviour, not feelings. Suleyman’s moratorium trigger, notably, would operate on precisely the kind of evidence nobody yet knows how to produce.
❓ FAQ
Did Suleyman call for a pause on AI development? No — he named a conditional: if consciousness began emerging during training, he would call for development to stop until those properties could be designed out, per Business Insider.
What is Anthropic’s actual new rule? Its usage policy now asks users not to engage in sustained abusive or cruel behavior toward its models, with Claude ending conversations as the main mechanism and permanent bans possible in extreme cases, as reported by CNN on 1 November.
Is this a dispute about whether AI has feelings? Not exactly. Both companies agree there is no evidence models are conscious. The dispute is whether the industry should design models to mimic beings that might deserve moral concern — Anthropic studies the question; Microsoft says don’t build the mimicry in at all.
Why does this matter outside Silicon Valley? Because the answer decides product behaviour at consumer scale: how chatbots respond when users are cruel, whether models claim feelings, and whether regulators eventually treat “model welfare” as a real design discipline.
🔍 THE BOTTOM LINE
Two of the most valuable companies in AI have now published opposite operating theories of what a chatbot is. Anthropic hedges its bets by policing cruelty toward models it explicitly says don’t suffer; Microsoft wants the possibility designed out entirely, with a public stop-work trigger if it ever appears. The next few years of product design — and quite possibly AI regulation — will be fought on the ground between those two positions.
📰 Sources
- Business Insider — Suleyman X post and quotes
- CNN — Anthropic usage policy coverage
- KSL — syndicated Anthropic policy report
- Mashable — September essay coverage
