Why DeepMind Has a Philosopher, and Why AI Ethics Needs More Than One

Click to expand

Hiring a philosopher at an AI laboratory can sound like an exercise in answering distant questions. In practice, the questions have already reached the product. Should an assistant describe itself as a person? What should it say when someone becomes attached to it? What kind of evidence would justify a claim about machine consciousness?

In an October 13 IEEE Spectrum interview with David Berreby, Henry Shevlin explains why he moved much of his work from Cambridge into Google DeepMind. According to the interview, he joined in May and now spends 80% of his time at the company while remaining active at Cambridge. His job title is simply philosopher.^[1]

The appointment is interesting because the boundary between philosophy and engineering is becoming harder to maintain. Choices about identity, memory, language and autonomy determine how a system behaves and how people interpret it. Those are product decisions with ethical consequences, whether or not a product team uses philosophical language.

Philosophy belongs near the technology, but not only inside the lab

Shevlin argues that daily contact with rapidly changing models is essential to understanding them. He also points to the prospect of direct impact: an idea discussed with the people building or deploying a system can affect decisions much sooner than an academic debate alone.^[1]

That is a persuasive case for putting expertise inside a laboratory. A philosopher who can examine the actual model, its evaluations and its design constraints is better placed to ask useful questions than one reasoning entirely from an outdated description. Engineers, in turn, benefit from someone who can distinguish concepts that an attractive demonstration tends to blur.

But proximity is not a substitute for independence. An internal appointment does not, by itself, show whether ethical findings can delay a release, change an engagement target or become public when they are commercially inconvenient. Those are governance questions, not judgments about Shevlin's personal integrity.

The healthiest arrangement is complementary: researchers inside labs with real influence, universities able to challenge their assumptions, and scrutiny from affected users. Ethical expertise is valuable because it can alter a decision, not because it supplies a reassuring job title.

Three questions that should not collapse into one

The first question is capability: can a system reason about another person's perspective, adapt its conversation or complete a task? The second is social behavior: does its presentation encourage appropriate trust, unhealthy dependence or confusion about what it is? The third is subjective experience: is there anything it feels like to be that system?

These questions are related, but success on one does not answer the others. A model might predict what a person believes without having that person's experience. A convincing declaration of affection can affect a user without establishing that the system feels affection. A capable agent can also cause harm through its tools without being conscious.

Shevlin's February paper, Three frameworks for AI mentality, makes a useful distinction here. It explores whether limited belief-like or desire-like descriptions help explain AI behavior while explicitly setting aside the harder questions of consciousness and the nature of representation. Its proposed minimal cognitive-agent framework is a philosophical account, not experimental proof that current assistants have human minds.^[2]

The practical value is permission to be precise. Calling a system's stored representation a belief can be a useful analytical shorthand. It does not automatically establish feelings, personhood or moral rights. Equally, explaining the computation does not by itself settle every philosophical question about minds.

A safety change can affect more than its target

One of the interview's most provocative points concerns training models not to claim mental states. Shevlin refers to research suggesting that interventions in this area also change how models attribute minds to animals and respond to spiritual questions.^[1]

The underlying study is worth reading separately from the interview. In an arXiv preprint, Junsol Kim and colleagues examine three instruction-tuned models from the Llama 3 and Gemma 2 families. They compare the original systems with interventions that remove a learned refusal direction or steer internal activations toward consciousness-affirming responses.^[3]

They report shifts in model answers about mindedness, spiritual beliefs and standardized survey questions. Crucially, the evaluated theory-of-mind and general-reasoning performance did not significantly change under the reported safety-ablation tests. That is more specific than a claim that discouraging self-ascribed consciousness necessarily destroys social reasoning.

There are important limits. These are model responses and experimental interventions, not measurements of a model's feelings or evidence of real-world animal-welfare outcomes. The authors explicitly leave open whether self-attributed consciousness is the causal mediator of the broader changes. Removing a refusal direction is also not a clean test of one sentence in a product's identity policy.

The lesson is not to make assistants insist they are conscious. It is to test safety changes for unintended effects beyond the behavior they target. A model should be able to discuss people's beliefs respectfully and reason carefully about animals without presenting an unsupported claim about its own inner life.

Warmth and honesty are compatible

The interview resists a blanket conclusion that human-like conversation is always harmful. Shevlin points to potential benefits as well as risks and says the evidence about human-AI relationships remains thin, particularly outside the populations most often studied.^[1]

Taking that uncertainty seriously means avoiding two shortcuts. Anecdotes of comfort do not establish clinical effectiveness. Reports of harmful relationships do not establish that every friendly interaction is dangerous. Outcomes depend on the user, context, product design and behavior over time.

A helpful design distinction is between accessible conversation and manufactured obligation. An assistant can be patient, remember permitted preferences and use ordinary first-person language without claiming a human biography, demanding loyalty or implying that a user hurts it by leaving. A name and a friendly tone do not need to become a claim about personhood.

DeepMind's 2024 work on advanced AI assistants already identifies this tension: systems should remain identifiable as AI, preserve users' autonomy and avoid relationships built on emotional or material dependence.^[4] These are goals to evaluate in real interactions, not properties established by publishing a policy.

Evaluation should therefore ask more than whether a user enjoyed one conversation. Can people disagree with the assistant? Does it preserve their ability to make independent decisions? Does it encourage appropriate human support rather than positioning itself as indispensable? Do the safeguards work across languages, cultures and different levels of vulnerability?

Consciousness needs better evidence than a confession

Shevlin describes machine consciousness as an exceptionally difficult scientific problem. He suggests comparative cognition and work on model internals as possible routes forward, rather than expecting a decisive answer from conversational impressions.^[1]

A research report by Patrick Butlin and colleagues illustrates another evidence-led approach. It derives computational indicators from several scientific theories of consciousness and examines systems against them. Its 2023 assessment suggested the systems considered were not conscious, while finding no obvious technical barrier to building systems satisfying the proposed indicators.^[5]

That is a dated assessment of particular systems, not a permanent verdict on everything built afterward. Nor does satisfying a proposed indicator automatically settle the underlying debate. The useful contribution is the method: make the theory, implementation and uncertainty explicit instead of treating a model's fluent self-report as a reliable test.

For product teams, uncertainty is not permission to imply consciousness for engagement. Nor is it a reason to stop investigating the question. Research can remain open while public claims remain proportionate to the evidence.

The next ethical problem is coordination

Shevlin also challenges the familiar picture of one person talking to one chatbot. He expects relationships involving multiple humans and multiple agents to become increasingly important.^[1] DeepMind's earlier assistant research similarly considers how individually useful agents can create collective problems when many act at once.^[4]

Consider a hypothetical pair of assistants arranging a meeting. One knows a person's private calendar; the other asks for more context to negotiate a time. The relevant questions are not only whether either model is intelligent. They include what information can be shared, whose instructions govern, and which commitments require the person's approval.

Philosophical clarity helps identify those questions. Engineering must then make the answers enforceable through scoped access, visible permissions, reviewable actions and a route to human intervention. A conversation about trust cannot replace those mechanisms.

DeepMind's resident philosopher is therefore a useful sign of a larger shift. AI ethics is not an ornamental debate that begins after the technical work is finished. It is part of deciding what the technology should say, what it should be allowed to do and what evidence its makers owe the people who live with it.