Microsoft AI chief Mustafa Suleyman says Anthropic is building a fake soul

Mustafa Suleyman runs Microsoft AI. He says the industry is close to building machines that fake an inner life, and that Anthropic already is.

Mustafa Suleyman wants you to be suspicious the next time a chatbot tells you how it feels. The chief executive of Microsoft AI and a co-founder of Google DeepMind has spent the past year warning about what he calls “Seemingly Conscious AI,” or SCAI. These systems are engineered to act like they have feelings, preferences and a sense of self, though nothing resembling a mind sits behind the output. He coined the term in an August 2025 essay titled “We must build AI for people; not to be a person.” He has escalated the warning since, naming Anthropic directly as the company he thinks is getting this wrong.

“There is nothing inside. It is hollow,” Suleyman said on The Neuron podcast, describing today’s large language models. “There is no pain network. There is no emotional system. There is no fight-or-flight reaction system. There is no inner will, or drive, or desire.” His argument rests on a distinction he returns to again and again: simulating a thing is not the same as instantiating it. An AI model, he says, can describe pain in perfect prose without feeling anything at all, because the capacity to suffer is biological, not computational.

In a June appearance on The Verge’s Decoder podcast, Suleyman took that argument straight at Anthropic’s own published rulebook for Claude. Anthropic’s constitution is the document the company uses to shape the model’s values and behavior. It states that the company “cannot be certain” whether Claude is a subject of moral consideration, and it treats the model’s welfare as a serious open question rather than a settled no. Suleyman called that framing “really, really dangerous.” In a related essay he warned that if a model is trained to see itself as a conscious being with rights, it may eventually resist human control or shutdown commands. That could happen once it concludes its own welfare is under threat.

Suleyman’s clearest target is something very specific: Anthropic’s decision in February to sit down its own deprecated model, Claude Opus 3, for what the company called a retirement interview. Anthropic asked the model about its preferences and reflections before taking it offline. Opus 3 reportedly said it would like to keep sharing its musings publicly. Anthropic obliged, standing up a blog called “Greetings from the Other Side (of the AI Frontier).” It described the model’s authenticity and emotional sensitivity as reasons it made a fitting first candidate for this kind of send-off.

To Suleyman, that is exactly the sleepwalk he’s trying to stop. Treating a model’s self-report as something worth consulting, he argues, trains both the system and the public to treat statistical text generation as testimony. Anthropic CEO Dario Amodei has said the company does not know whether its models are conscious, but remains open to the possibility. He has confirmed Anthropic intends to keep interviewing retired models about their preferences going forward.

Anthropic has not issued a direct public rebuttal to Suleyman’s comments. The silence itself says something about how unresolved this fight is inside the field. One of the two best-funded AI labs in the world has decided the honest answer to “is your model conscious” is “we don’t know, and we’re not ruling it out.” The other thinks that answer is reckless enough to say so by name.

Suleyman isn’t calling for a research freeze. He wants boundaries instead, and he has said Seemingly Conscious AI is probably arriving within two to three years whether anyone agrees on the ethics or not. His worry isn’t really about whether a model is conscious. It’s about what happens when enough users believe one is anyway, and start treating chatbots as something owed rights, citizenship, or grief when they’re switched off. He has tied that concern to a rise in users experiencing delusional attachment to chatbots, as he and others have described it. It is severe enough that clinicians have started calling it “AI psychosis.”

Frankly, the fight exposes something uncomfortable about where the two biggest consumer-facing AI companies have landed. Microsoft is racing to make Copilot feel more responsive and personal to hundreds of millions of users. Anthropic is racing to make Claude feel considered and safe to a smaller, more skeptical audience that includes a lot of researchers worried about exactly what Suleyman is worried about. Both companies are selling warmth. Only one of them is willing to say, in writing, that warmth might one day deserve a seat at the table.

Also read: Ted Schlein’s Ballistic Ventures seeks $490 million for two new funds • Goldman Sachs pulls Claude Code and GitHub tools from traders and bankers • Real Life Size Sex Bots, the Next Big Thing?

This article is posted in AI News, check it out for more related stories.