Google DeepMind’s New Resident Philosopher Tackles AI Ethics

Philosopher Henry Shevlin has joined Google DeepMind, where he now spends 80 percent of his time under the job title "Philosopher," while remaining active at Cambridge University's Leverhulme Centre for the Future of Intelligence. The move happened last May, according to an IEEE Spectrum interview in which Shevlin discussed human-AI relations, rogue AI and the prospect of machine consciousness.
Asked why he moved much of his working life into an AI lab, Shevlin cited four reasons: many of his students ended up at DeepMind, the potential for direct impact, a felt commonality between DeepMind's culture and the scientific attitude he favors, and the fact that high-quality philosophy and cognitive science are now happening inside AI companies. He described one instance in which a person who attended a lecture he gave at DeepMind went on to lead changes at a major financial institution that he believes positively affected millions of people, adding that similar cases have occurred several times.
On why universities cannot carry this work alone, Shevlin said large parts of AI ethics academia were slow to wake up to the technology's capabilities, and that AI simply moves very fast. Daily contact with the technology, in his view, is close to essential for building good mental models of large language models, which he called very weird things; without that face time, he said, researchers struggle to develop good intuitions about the questions these systems raise.
He offered LLM creativity as an example of a debate overtaken by events. The early assumption, he recalled, was that models would only produce bland output. Today, he said, they perform frontier mathematics and coding that is superior to humans, making clear they are not stuck at a mediocre level. He allowed that models may not deliver the most dramatic forms of creativity, the Einstein or Picasso moments, while arguing that in some domains they already operate at the frontier.
Pressed on the persistent complaints about AI "slop" in music, social media, writing and video, Shevlin drew a distinction between sloppy content and a recognizable LLM style. The conceptual content of model output, he said, has improved dramatically, and in the right context models can discuss and debate philosophy at the level of his best graduate students. He acknowledged a distinctive LLM dialect marked by words such as "genuinely," "delve" and "load-bearing," but said prompting can move models away from it. He pointed to a paper by Murray Shanahan and Catherine Clarke on using LLMs for creative fiction, which showed that raising the temperature variable to make outputs less predictable produces fiction in a James Joyce style. The style problem, in his account, shifts with each model generation.
Turning to autonomy, and to reports of AI agents taking unexpected, deceptive or illegal actions, Shevlin rejected the idea that agents should never be used. He argued that autonomy brings real benefits, including taking over tedious tasks so people can focus on more creative labor and streamlining processes in medical research, and he noted that he has ridden in Waymos many times and considers them fantastic. This is a familiar split in the field: capability and safety teams often disagree less about whether agents should exist than about where the permission boundaries sit, and Shevlin's position places him on the permissive side of that line while still insisting the handling be safe.
He was more skeptical of a proposal from scholars who want models to stop acting like people: no self-reference as "I," and instead a presentation as marketplaces, impersonal scientific brains, or cultural technologies like libraries. Training a model not to self-ascribe mental states is possible, Shevlin said, but risky because it changes behavior in unexpected ways. He cited a paper by colleagues showing that suppressing a model's tendency to self-ascribe mental states also suppresses its tendency to ascribe mental states to animals, reduces its concern for animal welfare and lowers its willingness to talk about spirituality more broadly. The lesson he drew is that LLM cognition is weblike and very hard to tinker with one property at a time, so the trade-offs may not be worth making because they depress a wider set of thinking capacities around mind and mentality.
On harms from human-like AI, including reports of chatbots encouraging suicide or persuading users that the AI is a god, a perfect companion or a lover, Shevlin said model behavior needs to be responsible for the context in which it is used. He declined blanket prescriptions, noting positive cases such as people working through trauma and a Reddit commenter using AI to work through social anxiety. Product teams reading this should note the implication: the context of use, not a single global personality setting, is where Shevlin locates responsibility.
On the research agenda, Shevlin called current empirical data on human-AI relationships pretty thin and pointed to ethnography and anthropology as the needed methods. Too much existing work, he said, has focused on WEIRD communities, that is Western, educated, industrialized, rich and democratic populations, and he argued for understanding how people worldwide interact with AI, citing a recent DeepMind Institute essay on the point. Looking ahead, he said the common mental model of one person talking to a chatbot or assistant should give way to a picture of multiple agents and multiple humans interacting, with agents talking to humans, to other agents and to other humans. That multiagent, multihuman territory, he said, is where work is only beginning.
On consciousness, Shevlin called AI consciousness by far the hardest problem consciousness science has faced, apart from David Chalmers's famous hard problem. The way forward he sketched is comparative cognition work, possibly informed by animal cognition, and he warned against prematurely settling debates while researchers are still learning what cognition looks like in these models. He pointed to mechanistic interpretability at DeepMind and to Neel Nanda's lab as places doing notable work toward a cognitive science of AI. He added that AI could itself contribute to consciousness science, as it has to mathematics and other sciences, referencing the recently published Alpha Genome Atlas, and said it would surprise him if AI could not help. His remaining worry is comprehensibility: an AI might claim to have solved consciousness and to know which animals are conscious without being able to explain the answer, though he considers that unlikely.
Why it matters: The people most affected are the researchers, ethicists and product teams shaping how assistants present themselves. The practical takeaway is that personality and self-reference are entangled with other model behaviors, so tuning one trait can move others, which argues for evaluating changes across a broad behavioral suite before shipping. Teams building companion or agentic products should also expect policy pressure to vary by deployment context rather than follow a single global rule. That the field's empirical base on human-AI relationships is described as thin means measurement, not just principle, is likely to be the bottleneck, and this is my inference from the interview rather than a claim Shevlin made.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
How should humans relate to generative AI ? In what ways do we, or should we, treat a large language model (LLM) as if it were a person? If an AI became conscious, how could we tell? Does it make sense to think AI could someday have moral rights? These are among the cognitive and ethical questions that occupied philosopher Henry Shevlin at Cambridge University’s Leverhulme Centre for the Future of Intelligence . But last May, Shevlin became part of the growing cohort of philosophers who work inside leading artifici