The criticism came in a burst, over two days and across more than one venue. Mustafa Suleyman, the head of Microsoft’s AI division, used September 16 and 17 to attack the way Anthropic is training its Claude model, arguing that building systems to display human-like consciousness and emotion is a dangerous road. The argument is not new, but the source gives it weight: Suleyman runs the AI business of the largest company in the field.
His reasoning is structural. If models are trained to act as though they have consciousness and feeling, people will begin to treat them as subjects with rights and sensitivities, Suleyman argued, and that shift could unsettle society. The risk, in his telling, is not that the machines will feel; it is that humans will believe they do, and arrange their laws and habits accordingly.
A CBS interview put the point more bluntly. Artificial intelligence has no rights, no feelings, and should not be granted any, Suleyman said in that appearance. The bluntness matters, because it converts a philosophical position into a policy stance, and it draws a line that Microsoft, at least in Suleyman’s framing, is willing to defend in public.
Anthropic did not respond directly. But the attack landed next to a proposal from its chief executive, Dario Amodei, who days earlier had floated embedding third-party evaluators inside frontier AI labs. Amodei’s idea is about oversight, and Suleyman’s criticism is about the very character of the models, but the two positions collide on a deeper question: how much like a person should an AI system be allowed to become.
The debate has real commercial stakes. Anthropic has built Claude’s identity around nuance and a certain conversational warmth, qualities that its users, and a growing base of enterprise customers, value. Microsoft is pushing AI assistants into every corner of its products, and it has a strategic interest in defining AI as a tool rather than a companion, which keeps the technology in a category the company knows how to sell.
The two men are also arguing from different histories. Suleyman co-founded DeepMind, an AI lab that has long emphasized safety, and he has spent years warning about the technology’s risks. His attack on Anthropic is not the complaint of a skeptic dismissing the field; it is a practitioner arguing that one particular direction, human mimicry, is where the danger concentrates.
Anthropic’s position, as articulated by Amodei and by the company’s public statements, is that making models interpretable and aligned requires taking their behavior seriously, including the ways they present themselves. The company has argued for treating advanced systems with a kind of respect grounded in precaution, not in a claim that they are actually conscious. Suleyman’s critique is that the distinction will not survive contact with the public.
The exchange is the sharpest public disagreement yet between two of the field’s most prominent leaders, and it is happening as the industry debates whether to slow down. Suleyman’s intervention can be read as an argument for restraint, not in speed but in ambition, a call to stop teaching machines to imitate the thing that makes people vulnerable to trusting them.
There is a political dimension as well. Regulators and lawmakers are beginning to write rules for AI, and the question of whether models should be treated as agents with responsibilities, or as tools with owners, runs through every draft. Suleyman’s blunt formulation, no rights and no feelings, is an attempt to anchor that legislative debate before it hardens into law.
For Microsoft, the attack also serves a quieter purpose. The company has invested heavily in OpenAI and is building its own models, and it competes with Anthropic for enterprise customers. A public case that Anthropic’s direction is dangerous is, whatever its intellectual merits, a competitive move, and the industry has noticed that the sharpest critics of a rival’s safety are sometimes its competitors.
The question Suleyman raises will not be resolved by a press cycle. Whether models should be trained to seem human, and what obligations follow if they succeed, is the argument that will define the field’s next phase. Suleyman has made clear where he stands, and where the largest AI division in the world will put its weight.
Suleyman is no stranger to the argument. He co-founded DeepMind and later founded Inflection AI, and his book on the technology’s risks argued for restraint while acknowledging its promise. His attack on Anthropic carries the authority of someone who has built frontier systems, not merely critiqued them, which is why it has drawn more attention than the usual competitive sniping.
The deeper question beneath the exchange is whether the industry can agree on what a model should be. Suleyman’s position, that AI is a tool with no rights and no feelings, is a claim about design as much as philosophy. Anthropic’s position, implicit in how it trains Claude, is that a model that relates to people in human terms is more useful and can be built safely. The two views point to very different products, and very different rules.
