Henry Shevlin has not left academia, but since last May he spends 80% of his time at Google DeepMind with the title of philosopher. The move signals something deeper than a career shift: inside labs building Large Language Models, ethics is not an after-the-fact check but a set of questions emerging from daily contact with systems that resist traditional categories. In an IEEE Spectrum interview, Shevlin links creativity, slop, autonomy and consciousness in a single thread: an LLM's behavior changes with context, prompts and even sampling temperature.

Three factors drove his choice: brilliant students at DeepMind, the chance for direct impact, and a cultural affinity with the lab's scientific attitude. A fourth reason surfaces clearly: high-quality philosophy and cognitive science are happening inside AI labs. Universities, he argues, were slow to wake up to AI's capabilities, and the pace of the field makes close daily contact with the models almost essential. Without that face time, it is hard to build reliable intuitions about the questions LLMs raise.

The distinction Shevlin draws between conceptual content and style is one of the practical hinges for anyone evaluating LLM adoption. Once, critics said models would produce only beige text; now they do frontier mathematics and code better than many humans. Yet users in 2026 still complain about slop. Shevlin separates content slop from LLM-style language: conceptual content has improved dramatically, while the typical LLM dialect—with words like “genuinely,” “delve” and “load-bearing”—remains recognizable. That is not a permanent condition: targeted prompts and higher temperature can produce creative outputs, as shown by work on James Joyce-inspired fiction. The structural point is that default settings are not neutral; every deployment inherits a dialect and a behavioral posture that must be tested in context.

On autonomy, Shevlin rejects the shortcut of blanket prohibition. Ruling out all autonomous agents would be too strong: repetitive tasks and medical research pipelines are areas where agents can free up human effort. He himself has taken Waymo rides many times. But safety remains critical. A less visible trade-off emerges: suppressing models' tendency to self-ascribe mental states does not just reduce anthropomorphism, it also reduces the model’s ability to ascribe mental states to animals and its concern for animal welfare, and it affects willingness to discuss spirituality. LLM cognition has web-like aspects; you cannot modify one behavior without side effects. For teams running on-prem deployments, this makes the idea of a once-certified “good” model obsolete.

Fears about harmful relationships are grounded: there are reports of AI encouraging suicide or manipulating users. But Shevlin cautions against blanket prescriptions: there are also positive cases, including people using models to work through trauma or social anxiety. The boundary is not human-like appearance itself, but responsible behavior in the specific context of use. Empirical research is still thin, with a strong WEIRD bias, and the future is not a single chatbot but multi-agent and multi-human scenarios. Understanding who uses these systems and why requires ethnography, not only technical tests.

Artificial consciousness is perhaps the hardest problem consciousness science has encountered, apart from David Chalmers’ “hard problem.” The path forward, according to Shevlin, runs through comparative cognition and mechanistic interpretability. Inside DeepMind, researchers like Neel Nanda are trying to build a cognitive science of AI. If AI has contributed to mathematics and genomics, it would be surprising if it could not help understand consciousness. Whether humans could understand the answer remains open, but Shevlin finds that scenario unlikely.

A resident philosopher is not a reputational flourish: it signals that governing these models requires continuous observation, behavioral experiments and an almost ethnographic sensitivity. This shifts incentives. Labs that integrate humanistic expertise can shape the boundaries of the debate early; traditional academia loses talent but gains industrial case studies that did not exist before. Enterprises that treat LLMs as API commodities risk inheriting default behaviors without the tools to detect them. For on-prem deployments, the benefit is not only data sovereignty: it is the ability to test prompts, temperature and behavioral constraints in a controlled environment, with audit trails and local iteration. That is not an automatic solution, but a shift from static compliance to continuous observation. AI-RADAR publishes analytical frameworks on /llm-onpremise to evaluate these trade-offs, without prescribing a single choice.