Google Deepmind’s New Resident Philosopher Tackles AI Ethics

Henry Shevlin used to work on the big, slippery questions from Cambridge University’s Leverhulme Centre for the Future of Intelligence: whether large language models should be treated like people, what would count as machine consciousness, and whether AI could ever deserve moral rights. Last May, he took that work into the belly of the beast. He now spends 80 percent of his time at Google DeepMind, while staying active at Cambridge, with the job title “Philosopher.”

His reason for making the jump is pretty practical. He points to brilliant DeepMind students, the chance to have direct impact, and a shared scientific culture. He also says some of the best philosophy and cognitive science on AI is happening inside AI labs now. Universities, in his view, were slow to catch up with what these systems can actually do.

That speed matters. Shevlin says daily contact with the technology is essential because LLMs are “very weird things,” and old arguments age badly fast. The easy dismissal of models as beige, sloppy imitators no longer fits, he argues, because they now do frontier math and better-than-human coding in some areas. They may still miss the dramatic flashes of Einstein or Picasso, but the ceiling has clearly moved.

He’s also pushing back on the idea that AI language is just one uniform blur of “slop.” The annoying stock phrases are real, he says, but they can be pushed aside with prompting, and the style shifts from model to model. On the substance, he’s more impressed than the memes suggest: in the right setting, these systems can argue philosophically at the level of some of his best graduate students.

On safety and autonomy, Shevlin does not want a total ban on agents. He thinks there are useful jobs for them, from dull chores to parts of medical research, and he says Waymos have gone over well with him. But he also warns against simple fixes, like training models not to describe their own mental states. That can spill into weird places, including how they talk about animals, spirituality, and concern for welfare. The model, in other words, is a web, not a switch.

The hardest question is still consciousness. Shevlin calls AI consciousness the toughest problem consciousness science has faced, aside from David Chalmers’ famous hard problem. His bet is that progress will come from comparative cognition, animal cognition, and mechanistic interpretability work like what Neel Nanda’s lab is doing at DeepMind. And yes, he thinks AI itself may end up helping to answer the question. It would be a very AI-era ending if the machine could explain consciousness better than we can.