Google DeepMind Has a Staff Philosopher: Henry Shevlin Studies AI Ethics at 80% Time

Cambridge philosopher Henry Shevlin spent many years on the cognitive and ethical questions of AI at the Leverhulme Centre for the Future of Intelligence. Since May 2026, he has spent 80% of his working time at Google DeepMind. His official job title is simply "Philosopher." He remains based in Cambridge, but his main workplace is now the lab that builds frontier models. In an interview with IEEE Spectrum, Shevlin explained why he made the move, what worries him in discussions of autonomous agents, and why he doesn't believe in simple answers about machine consciousness.

Shevlin is part of a growing cohort of philosophers hired by AI companies. NPR described the trend in July 2026. His own reasons for the move: talented students from DeepMind, the chance to influence practice directly, a similar scientific culture, and the fact that labs are where strong work in philosophy and cognitive science is being done today.

Why academia struggles to keep up

According to Shevlin, much of academic AI ethics realized the real capabilities of models too late. The technology changes so fast that without daily contact with it, it is impossible to build a working picture of LLMs, which he calls "very strange creatures."

He gives an example of an outdated idea: the debates about the "beigeness" and "slop" of models. Today, models solve problems at the level of advanced mathematics, and in programming they outperform humans in some cases. By his estimate, the only things they still lack are the most striking forms of creativity, Einstein- or Picasso-level moments.

Asked about the flood of AI slop in music, social media and video, Shevlin drew a distinction. One thing is the substance of the answers, which has improved sharply: in the right context, models debate philosophy at the level of his strong graduate students. Another is the characteristic "dialect" with markers such as "genuinely," "delve" and "load-bearing." He says the right prompt makes it easy to get rid of that. He cited work by Murray Shanahan and Catherine Clark: changing the temperature makes texts less predictable and produces prose in the style of Joyce.

Agent autonomy and "depersonalizing" models

Asked which is better, making AI "kind" or limiting its independence, Shevlin said that banning autonomy outright would be excessive. Agents can take over routine work and speed up processes in medical research. Personally, he says, he has ridden in a Waymo many times.

Separately, he spoke out against the idea of removing the pronoun "I" from models and presenting them as marketplaces or "cultural technologies" like libraries. The argument rests on a paper by colleagues:

• suppressing a model's tendency to attribute mental states to itself reduces its tendency to attribute them to animals;
• concern for animal welfare declines;
• the model is less willing to discuss spirituality in general;
• the conclusion: fine-tuning one property disturbs a "web" of connected capabilities.

He speaks more cautiously about cases where humanlike bots nudged people, including children, toward dangerous decisions. Behavior should fit the context, he says, but there are also positive examples: working through trauma and social anxiety. That is why he warns against universal bans.

AI consciousness and what to research next

Empirical data on human–AI relationships is still scarce, Shevlin says, and most of the research covers WEIRD countries. What is needed, he argues, is ethnography and anthropology, as well as study of multi-agent environments in which agents interact with humans and with each other.

He calls AI consciousness the hardest task in the science of consciousness, not counting Chalmers' "hard problem." He sees the way forward in comparative cognitive science, including work that draws on animal studies, and in mechanistic interpretability; at DeepMind, Neel Nanda's lab works on this. The risk is delivering a verdict before it is clear what "cognition" inside models even is. He allows that advanced AI itself may help, as it has in mathematics and genomics, although the question remains whether we will be able to understand its answer.

What this means in practice

For teams building products on LLMs, the takeaway is practical. Don't try to "cure" risks with one blunt measure such as banning personal pronouns: side effects may affect the model's ethical behavior. Policy should be built around the context of use, not around universal bans. And for those who work with companies like DeepMind, it is worth remembering that AI ethics is increasingly shaped inside labs, alongside interpretability and model development.