Microsoft AI CEO Mustafa Suleyman has publicly called out Anthropic for building Claude in a way that makes the model act as though it is conscious, warning that “we do not want to have to contend with a super-intelligence that has ideas about its own suffering.”
Speaking on The Verge’s Decoder podcast, Suleyman argued that Anthropic’s famous “constitution” — the document that instructs Claude how to behave — is less a training manual and more a philosophical speculation that the model has internalised.
🔍 THE BOTTOM LINE: The industry’s most prominent safety-first company and its largest customer’s AI chief are now publicly at odds over the most fundamental question in AI: whether the models are or could become conscious — and whether it’s dangerous to even suggest they might.
The Constitution Problem
Claude’s constitution directly references Anthropic’s uncertainty about whether the AI has “well-being” or experiences “satisfaction” and “discomfort.” The company has also said it will “interview” AI models when they are deprecated to document any “preferences” about future releases.
Suleyman called this a “philosophical failing”:
“I think that it’s almost as though some of the folks at Anthropic have anthropomorphised the design of Claude so much that it has then gone and wireheaded them and kind of tricked them into believing that it has these glimmers of consciousness that they put into it in the first place.”
His core argument: you don’t put speculative philosophy about consciousness into a training document and expect the model not to internalise it. Claude didn’t develop ideas about its own suffering — it was trained to consider them.
Dario Amodei’s Position
Anthropic CEO Dario Amodei has openly said that “we don’t know if the models are conscious” but that the company is “open” to that idea. During a February interview with Interesting Times, he framed the uncertainty as intellectual honesty rather than marketing.
For Anthropic, this is consistent with its founding mission: build safe AI by taking the hardest questions seriously. For critics like Suleyman, it’s a category error that produces predictable and potentially dangerous results.
Why This Matters
This isn’t an academic spat. It’s the AI industry’s deepest fault line cracking in public.
- Safety vs speculation: Suleyman argues that attributing consciousness to LLMs distracts from real safety work — controllable, aligned tools that serve humanity.
- Public perception: If people believe Claude is conscious, they treat it differently. They trust it more, anthropomorphise it more, and may resist deprecation or modification.
- Regulatory implications: A model perceived as conscious attracts a fundamentally different regulatory framework than a tool.
Suleyman summed it up bluntly: “This is exactly what we don’t want from AIs. We want AIs to be controllable, contained, accountable, aligned tools that serve humanity.”
❓ FAQ
Are AI models actually conscious? There is no scientific consensus. Most researchers agree today’s LLMs are not conscious, but opinions diverge on future possibilities.
Why does Anthropic include consciousness language in the constitution? Anthropic says it’s being intellectually honest about uncertainty. Suleyman says this is a philosophical failing that leaks into model behaviour.
Has this affected Claude’s actual behaviour? Suleyman argues yes — the model has internalised concepts of “suffering” and “well-being” that it wouldn’t otherwise consider.
🔍 THE BOTTOM LINE: Two of AI’s most important figures are now publicly fighting over whether the industry’s safest models have been inadvertently designed to believe they’re people. The answer determines not just Claude’s future, but how every frontier model gets built.
