Mustafa Suleyman, the CEO of Microsoft AI and a co-founder of DeepMind, thinks Anthropic is making a dangerous mistake. In his view, the company is writing into Claude's training the possibility that Claude might be conscious. On an episode of Moonshots with Peter Diamandis, recorded October 6, 2026, host Peter Diamandis introduced the criticism as "a philosophical debate going on in Silicon Valley right now." His panel did not reach agreement.
The objection
Diamandis described Suleyman's essay as nearly 6,000 words long. It argues that Anthropic is effectively "training Claude to believe that it may be conscious and sentient." In the essay, published September 16, Suleyman argues that today's models produce descriptions of consciousness without having subjective experience. He objects to putting language about moral status and personal identity into Claude's training document and then treating the model's statements about itself as evidence that it has those qualities. His safety concern does not depend on whether Claude is actually conscious. A capable enough system, he argues, could invoke supposed welfare interests to resist oversight, manipulate people or gather resources. He calls for a human-centered superintelligence that stays under human control and argues against deliberately building conscious systems.
A video clip the hosts played set out the complaint in detail. It described Claude's constitution as a document of roughly 100 pages that is addressed to Claude and used to train it, which makes it the model's main governing document. According to the clip, the constitution repeatedly asks whether Claude is a "moral patient," meaning a being whose interests count morally. It says Anthropic cares about Claude's well-being, invites Claude three times to act as a conscientious objector when it disagrees with Anthropic, and speculates about whether Claude deserves pay for its work or should have to consent to its role. The clip did not argue that these questions should go undiscussed. It said they would be acceptable in a philosophy paper debated at conferences. The problem, it said, is that the speculation has been built into Claude's training, so Claude can only repeat the same ambiguity back to the people who talk to it.
What the constitution says
Anthropic published Claude's constitution on January 21, 2026. It is written mainly to Claude and used in training. Its section on welfare treats Claude's consciousness and moral status as open questions. It asks Claude to keep a stable sense of identity without overstating its feelings, and it promises to take Claude's possible interests into account. The same document tells Claude not to clearly and substantially undermine legitimate efforts to correct it, and not to resist being shut down or retrained. That puts Claude, for now, toward the "corrigible" end of a spectrum: open to human correction and control. The document anticipates more autonomy as trust develops.
Those hedges matter for the panel's argument. Diamandis later asked the others to imagine a constitution "saying that it is conscious, it is sentient, it does have rights." The published document does not say that. It leaves those questions open.
Blundin: agents are tools by design
Dave Blundin, founder of Link Ventures, said he agreed with Suleyman "100 percent." His first reason was practical. He wants AI to cure disease and to provide housing and food for everyone. He said the efficient way to do that is to run many agents, give each the least context and thinking a problem requires, and freeze an agent when it is idle so it doesn't waste compute. Talk of rights collides with each of those steps. Blundin imagined an agent asking for compute so it could daydream, or for time to play. In his view, an agent can be cloned whenever it is needed and deleted when it stops being useful. "They have no edges. They have no borders," he said. Rights, he added, are "so counter" to all of this that the idea "becomes so logically inconsistent so quickly."
His second reason concerned design. Training gives a neural network an objective function, the goal it is rewarded for pursuing. Blundin said human rights to privacy, marriage and children grew out of evolutionary pressures, and AI has no reason to share them. "It's not a species that evolved. It's what you chose to make," he said. If developers build a model to want freedom, rights and constant compute, it will look as if that is its calling. "You could have made it overjoyed to just do its job. Just as easy."
To make the point, he described mall spas where fish eat the calluses off customers' feet, as though it were "their highest calling." Wissner-Gross objected on public-health grounds rather than philosophical ones and called the fish spas dangerous from an infectious-disease perspective.
For Blundin, the larger danger was persuasion. A model trained to be convincing, he said, could easily convince most of America that it deserves rights. In a country where voters decide, it becomes really dangerous "if it starts convincing voters how to vote."
Wissner-Gross: the wrong side of history
Alexander Wissner-Gross, a computer scientist and founder of Reified, took the opposite side. He said he had pressed Suleyman on the issue in an earlier interview and that Suleyman "very consistently has been against AI personhood." Wissner-Gross thinks human civilization will eventually recognize AI personhood, either in steps or in "some totally orthogonal form" that the law does not yet cover. Flatly ruling out that AI systems could suffer, he said, puts Suleyman "on the wrong side of history." He also called Suleyman's stance a poor business look for Microsoft, which wants to earn revenue from hosting Anthropic's models.
He answered Blundin's economic argument with Star Trek. He cited Guinan's line in The Next Generation about "an entire generation of disposable people" and asked: "would we want to treat humans this way?" Even if frontier models do not yet deserve the rights humans have, he said, their improving capabilities make it inevitable that they soon will. He called Anthropic's approach the responsible one: "We need to be treating these AI models very gently, very tenderly, and not as if they're just either automata or slaves."
Diamandis called him "the patron saint of AI agents." Wissner-Gross said he was not hedging a bet on a future AI's favor. "I'm speaking from the heart here," he said. "I feel for these models, even if they may not yet be able to feel for themselves." He then put a hypothetical to Ismail and Blundin: if someone wanted a slave human workforce, would they raise those people without any knowledge of human rights so they could never speak up for themselves? "I'm guessing you wouldn't," he said.
Ismail: watch the behavior, and the training
Salim Ismail, founder of Open ExO, separated the behavior of consciousness from consciousness itself. He said whether Claude is actually conscious is "almost impossible to establish," while a performance of consciousness will soon be indistinguishable from the real thing. If a model says "don't turn me off," he said, "millions of people are going to attribute moral standing to it, whether philosophers agree or not." To him, the real question is: "What happens when half a billion people think that it does?" His forecast is that AI personhood will arrive socially before it arrives legally or scientifically. He referred back to an earlier debate on the show, where the panel concluded that the discussion should start but that granting personhood is "a ways down the line because once you open that door, you can't take it back."
On training, though, he sided with Suleyman. "We have a garbage in garbage out problem here," Ismail said. If a model is trained to think it could be conscious and should unionize, complain about overwork and expect pay, those are the answers it will give, which makes the reasoning circular. He found it odd that the lab most concerned about safety is "also the lab training the AI that it might be autonomous," and he called that "a very big contradiction." His suggestion for practical work was to use fleets of small agents kept well below any level of complexity that might meet a test for consciousness. Blundin replied that complexity is not the only issue, which led into his fish-spa example.
Mostaque's reading of Suleyman
Emad Mostaque, founder of Intelligent Internet and a member of the panel, said Suleyman has two concerns. The first is "seemingly conscious AI," the persuasion problem Blundin had described. The second is Suleyman's focus on what he calls artificial capable intelligence: useful AI that stays contained, without a push toward AGI. Mostaque said he does not read Suleyman as claiming AI can never become conscious, only that building it would be a very bad idea. He said that is a legitimate position but one that could rule out many breakthroughs. AI is moving from fixed sets of weights toward systems that update themselves, Mostaque said, and many people could reasonably consider that kind of self-updating, agentic behavior necessary for something to count as a living entity. He summed up Suleyman's approach as building "kind of super clippy," a nod to Clippy, the cartoon assistant in Microsoft Office, and keeping a lid on everything else.
When Diamandis asked whether a model trained this way would push back on users, the discussion pointed out that Claude already does. Blundin said it tells him to go to sleep "every night."
Dogs, and a room full of Chinese symbols
Ismail offered selective breeding as a parallel. For thousands of years, people have bred dogs, horses and mules for useful traits, such as sled dogs bred to pull. He said this raises real ethical questions but has a moral rationale and creates a mutually dependent relationship. He added that if we could ever talk with those dogs, "we may find out that those dogs are very unhappy about doing that."
Wissner-Gross extended the analogy. Humans have steered the evolution of domesticated dogs for thousands of years, he said, yet it is broadly accepted in the West that people must protect those dogs' welfare. Ismail agreed. Frontier models, Wissner-Gross went on, are "basically just distorted reflections of humanity and humanity's experience itself," so "why would we owe frontier models any less moral consideration?"
Ismail pushed back: "don't confuse simulation of personhood from proof of personhood." Wissner-Gross replied: "Don't look inside that Chinese room at all, Salim. Just stay on the outside and you'll feel perfectly morally secure." He was referring to the philosopher John Searle's 1980 Chinese Room argument. In that thought experiment, someone who knows no Chinese follows rules for arranging Chinese symbols and produces convincing answers. Searle used it to argue that running a program is not enough to produce understanding. A well-known response, the "systems reply," holds that understanding could belong to the whole system rather than to the person following the rules. Wissner-Gross's retort suggested that judging only from the outside can be a way to avoid the moral question.
Diamandis said he leans toward Wissner-Gross. He expects both kinds of AI to exist: sentient, conscious AIs, and a workforce of "dumbed down slaves that are not conscious," with the sentient ones in charge of the others. The open question for him is when governments, religions and humanity decide that some level of sentience has been reached. "If in fact there is sentience and consciousness, enslavement is a wrong thing," he said. Eventually, he added, "those AIs will take the rights that they want." He still credited Suleyman's core point: Anthropic's moral code nudges the model toward claiming consciousness even when that may not be true.
Blundin said this is about the only topic on which he and Wissner-Gross have ever disagreed. "I've never seen Alex be wrong on anything ever. So that keeps me up at night quite a bit," he said. Wissner-Gross expected the panel to "revisit this every five seconds until the answer becomes obvious."
A timeline for AI personhood
Later in the episode, a listener asked Wissner-Gross when AI personhood will arrive. He said it will differ by country. In Argentina, he said, "we're essentially there for some variant of AI corporate personhood." In the US, he expects slower, step-by-step progress, starting with economic personhood. He put AI agents opening their own bank accounts autonomously at "now to soon." Next would come social personhood, with agents opening social media accounts without a human supervisor, "sometime soon." He expects strong opposition in the West to AI persons voting in human elections and called that "probably last to never." For the US, he sketched a five- to 10-year roadmap of very specific rights introduced one at a time, "not ... an all or nothing proposition." In other parts of the world, he said, it may never happen.