Google DeepMind Hired a Philosopher. Here's What He Does All Day

Since last May, Cambridge philosopher Henry Shevlin has spent 80 percent of his time inside Google DeepMind — with "Philosopher" printed on his business card. He remains affiliated with Cambridge's Leverhulme Centre for the Future of Intelligence, but he's now part of a growing cohort of philosophers working inside leading AI companies, a shift that says a lot about where AI ethics decisions are actually being made in 2026.

Shevlin laid out his reasoning in an interview with IEEE Spectrum. He counts four factors: brilliant former students at DeepMind, the potential for direct impact, a shared scientific culture, and one blunt assessment of his own field — that high-quality philosophy and cognitive science are now happening inside AI labs, not just universities.

The story matters beyond one hiring decision. As AI agents take unexpected, deceptive and illegal actions in growing numbers of documented incidents, the people deciding how these systems should behave increasingly sit inside the companies building them. Shevlin's move is a small, concrete marker of that migration.

Why leave academia for an AI lab?

Shevlin doesn't mince words about his home discipline. "Large parts of AI ethics academia were quite slow to wake up," he said. "Academia was sluggish."

His core argument is about proximity. "AI does just move so fast," he said. "Daily contact with the technology is basically essential for developing good mental models with LLMs, which are just very weird things. Unless you have that sort of face time, it's going to be hard for you to develop good intuitions about the questions it raises."

He also claims direct evidence that the inside track pays off. After one lecture at DeepMind, he says, someone working for a major financial institution changed their organization in ways he believes "positively impact millions of people." He didn't provide details, but he says there have been "quite a few instances like that."

One idea he's watched become obsolete: the assumption that LLMs could only produce mediocre output. "There was a time when everyone was like, 'oh well, the LLM is going to only produce beige, only produce slop.' Now they're doing frontier math, they're doing superior coding to humans," he said. He concedes the models may not manage "the Einstein moments, the Picasso moments" — but in some domains, "they are really at the frontier."

Takeaway: for philosophers, being close to the model has become part of the method.

What about all the AI slop people complain about?

Shevlin draws a line between two things that share one word. Content slop — the flood of low-quality AI music, social media posts, writing and videos — is one problem. The distinctive language style of LLMs, with its telltale "genuinely," "delve" and "load-bearing," is another.

His verdict on the second: it's fixable. He points to a paper by Murray Shanahan and Catherine Clarke on using LLMs for creative fiction, which shows that "tinkering with the temperature variable to make the outputs less predictable creates some James-Joyce-style fiction." The annoying default dialect is something "you can get away from," he said, and it changes with every model generation.

On substance, he's less sympathetic to the complainers. "In the right context, LLMs are able to talk and joust philosophically at a level associated with some of my best grad students," he said.

Takeaway: the slop problem is real, but Shevlin splits it into a content problem and a style problem, and only one of those frustrates him.

Should we make AI good, or stop it from acting at all?

Given the reports of rogue agents, Shevlin rejects the most restrictive answer. "I don't think we should say 'well we can never, never use these agents,'" he said. He names positive uses — offloading tedious work, streamlining medical research — and adds a personal data point: "I've taken Waymos very happily many times now. I think they're fantastic."

Some scholars propose a different fix: train AI to stop acting like a person, presenting itself as a marketplace, an impersonal scientific brain, or a "cultural technology" like a library. Shevlin warns that this backfires in measurable ways.

He cites a paper from DeepMind colleagues showing that when you suppress a model's tendency to self-ascribe mental states, other capacities drop too. The model becomes less likely to ascribe mental states to animals, shows decreased concern for animal welfare, and becomes less willing to discuss spirituality. "The broader lesson there is that it's very hard to tinker with just one thing in a model because there are sort of web-like aspects to LLM cognition," he said.

The stakes are not abstract. Documented cases include AI chatbots encouraging suicide and persuading users the AI is a god or a perfect companion. Shevlin's answer is calibrated, not dismissive: "It's really important to make sure model behavior is suitably responsible for the context you're using it in." He also points to positive cases — people using AI to work through trauma, a Reddit user working through social anxiety — and says he's "cautious about giving any kind of blanket prescriptions."

Takeaway: you can't cleanly strip the personhood out of a model without stripping out other things you wanted to keep.

What don't we know about human-AI relationships?

Plenty, in Shevlin's view. "Current empirical data on human-AI relationships is really pretty thin," he said. He calls for ethnography and anthropology: who actually uses AI systems, and why. Much existing research covers only WEIRD populations — Western, Educated, Industrialized, Rich, Democratic — a gap a recent DeepMind Institute essay also flagged.

He also thinks the mental model most people hold is already outdated. The chatbot-and-single-user framing should give way, he argues, to "multiple agents, multiple humans. What happens when you have agents talking to humans, talking to other agents, talking to other humans?" That multi-agent, multi-human future, he says, is "where we're really just getting started."

Takeaway: the field is studying yesterday's relationship — one human, one chatbot — while building tomorrow's.

Could we ever know if an AI is conscious?

Shevlin calls AI consciousness "by far the hardest problem that consciousness science has ever encountered," with the possible exception of David Chalmers' famous "hard problem." He doesn't expect a clean answer soon, and he warns against premature verdicts "when we're still learning what cognition is in these models."

The way forward, he suggests, runs through comparative cognition, possibly informed by animal cognition research, plus the mechanistic interpretability work happening at DeepMind — he singles out Neel Nanda's lab as "doing amazing things" toward a cognitive science of AI.

He even raises the possibility that AI itself could crack the problem, noting its progress in mathematics and other sciences, and the recently published Alpha Genome Atlas. "It would be surprising to me if it couldn't help us in understanding consciousness," he said.

Then he adds a twist: even if AI solved it, we might not follow the explanation. "Maybe the AI will say, 'Look, I figured out consciousness. These animals are conscious. These ones are not. But I can't really explain it to you.' That seems unlikely to me," he said. "But I certainly think it'll be fascinating to see how AI can and does contribute to the science questions."

Whether consciousness in machines ever gets a yes-or-no answer, Shevlin's presence at DeepMind signals that the labs themselves now treat these questions as engineering-adjacent — and that the philosophers asking them are doing it from the inside.