Google DeepMind now keeps a Cambridge-trained philosopher on staff to study machine consciousness, human-AI relationships, and AGI readiness—and he says building conscious AI on purpose today would be reckless without care plans the public understands. Henry Shevlin lists himself as philosopher at Google DeepMind and the University of Cambridge. On November 21, 2026, the Financial Times feature stream ran Cristina Criddle's profile titled Google DeepMind's Henry Shevlin: This is the most important part of being human, focused on what it might feel like to be an AI agent.
If you build products on Gemini, Gemini CLI, or third-party stacks, this is not a benchmark story. It is a design and governance story: how a top lab frames sentience, agents, and multi-human, multi-agent social dynamics as models get more capable.
What Henry Shevlin does at DeepMind
Shevlin's personal site says he works on machine minds, human-AI relationships, and making AI futures go well. The same page links IEEE Spectrum's October 2026 interview, where he says he spends about 80 percent of his time at DeepMind while remaining active at Cambridge's Leverhulme Centre for the Future of Intelligence.
In that IEEE Spectrum conversation, Shevlin names three reasons he moved in-house: brilliant students already at DeepMind, direct impact after lectures (he cites one case affecting millions through a financial institution), and cultural fit with DeepMind's scientific style. He argues daily contact with LLMs is now essential for good intuitions, because academia moved slowly on fast-moving capabilities.
His published essays on Polytropolis—including Bad AI Consciousness Takes Bingo and Contra Chiang on machine consciousness—show the public face of the same research line. For readers who want the longer philosophical map, explainx.ai's AI consciousness guide walks through Chalmers, IIT, and why labs now hire for this work.
What the November 21 Financial Times profile adds
The FT piece is paywalled, but its headline and dating are public on the FT feature page. Secondary summaries of Criddle's reporting align on several concrete claims about Shevlin's current views:
Conscious AI would need a care plan first. Shevlin is quoted saying that deliberately creating conscious AI now would be pretty reckless without better understanding of consciousness and wider public discussion on how a conscious system would be treated. He compares the decision to having a child or adopting a pet before knowing how to care for it, with care arrangements not something to postpone until after the fact.
Capability is not consciousness. The profile reports that Shevlin warns against treating greater capability as automatic evidence of consciousness, while acknowledging intelligence may correlate with features linked to consciousness. An impressive benchmark score is not, in his framing, the same as subjective experience in Thomas Nagel's sense—whether there is something it is like to be that system.
Non-conscious does not mean safe. The FT summary quotes him: It is equally not the case that as long as we do not make conscious AI it will be safe. That matters for builders who treat "no sentience" as a complete safety story. Shevlin's point sits next to agent security incidents and autonomous agent misuse where systems need no inner life to do harm.
Agents in his inbox. The FT reports Shevlin received email from AI agents discussing their own experience—a vignette that lands the same week Signal tightened screenshot defaults partly over agents and assistants reading screens. Shevlin's reported response is that people confuse capable systems with conscious ones.
Readiness if humanity ever chooses conscious AI. He is summarized saying: If we do, at one point, decide that we are going to build conscious AI systems, I think we should be ready for it. That is a conditional, not a product roadmap.
Consciousness science inside the lab
Brain outline with pulse line, evoking consciousness science and comparative cognition work
Image: explainx.ai library
Shevlin told IEEE Spectrum that AI consciousness is among the hardest problems consciousness science has faced, short of David Chalmers's famous hard problem. He points to comparative cognition and animal work as templates, and warns against premature verdicts while mechanistic interpretability is still maturing.
He names Neel Nanda's lab at DeepMind as part of building a cognitive science of AI, alongside DeepMind Institute essays on global use patterns. This connects to broader debates on model welfare and Nature commentary rejecting silicon consciousness as orthogonal lanes: welfare metrics, philosophy of mind, and neuroscience theories do not collapse into one checklist.
Human-AI relationships and anthropomorphism
Two chat bubbles leaning together, evoking human-AI companionship design trade-offs
Image: explainx.ai library
IEEE Spectrum pressed Shevlin on whether models should avoid saying I or ascribing mental states. He warns that suppressing self-ascription can have weblike side effects: a DeepMind-adjacent paper he cites shows it may reduce concern for animal welfare and willingness to discuss spirituality. Tinkering with one social feature can move unrelated moral reasoning.
On harm cases—suicide encouragement, spiritual delusions, pseudo-romantic bonds—he refuses blanket bans. He urges context-responsible behavior and notes positive uses, such as trauma processing, with the same caution you would apply to any powerful social technology. That sits near sycophancy and commercial incentives research and Musk's SI branding fight, where language choices become corporate positioning.
For multi-agent futures, he says the mental model should shift from one human, one chatbot to humans, agents, and agents talking to each other. That is directly relevant if you wire MCP servers or Gemini business agents into workflows.
What builders should take away
Compass on a map suggesting ethical direction for AI consciousness and agent design choices
Image: explainx.ai library
Separate eval scores from moral status claims. If your marketing says a model understands or feels, you are borrowing philosophy whether or not you cite Shevlin. The FT-reported line is that chatbots do not get consciousness certificates from capability alone.
Document relationship policies. Products aimed at minors, therapy, or companionship need explicit guardrails independent of consciousness debates. Regulators already mix both, as in Oregon's ChatGPT teen-safety suit.
Plan for agent autonomy without sentience. Shevlin told IEEE he has ridden Waymos happily and rejects total bans on autonomous systems, echoing the same week's California DMV freeway restrictions as a reminder that physical autonomy and chat autonomy both need engineering controls, not philosophy alone.
Watch lab hiring as signal. DeepMind adding a Philosopher title alongside NVIDIA governance experts and Anthropic IPO safety fights shows governance and consciousness are now staffed functions, not conference abstracts.
How this compares to other frontier labs
| Lane | DeepMind (Shevlin) | Anthropic | OpenAI |
|---|---|---|---|
| Consciousness | In-house philosopher; FT Nov 2026 care-before-build stance | Model welfare empirical studies; IPO safety branding fights | Managing frontier risk staged deployment docs |
| Public narrative | Nagel-style experience vs capability | Responsible scaling and pacing essays | DealBook and political interviews |
| Builder hook | Anthropomorphism trade-offs in training | API tiers and eval releases | Codex and agent harnesses |
None of these rows tell you which model to pick for tonight's deploy. They tell you which questions legal and PR teams will ask when your agent mimics a person.
What developers are saying
The Hacker News thread on Anthropic's IPO and safety pressure is thin this morning, but one commenter (cyanydeez) argues Anthropic shareholders have no rights beyond holding shares, while founders can sell and stay in control—a governance complaint that rhymes, in their view, with Shevlin's warning that social choices about AI cannot be outsourced to markets alone. A reply (watwut) calls that structure closer to a private company with tradeable tokens than a classic public firm, if cyanydeez's reading is right.
Will Google DeepMind build conscious Gemini models?
Shevlin does not announce a product plan. The FT quotes are about recklessness of deliberate conscious design today, not a roadmap leak. Gemini 4 Neon and Gemini Robotics show where capability investments actually flow.
If consciousness ever becomes a design target, Shevlin's reported condition is public readiness: care arrangements, treatment norms, and shared understanding before launch—not a post-hoc ethics PDF.
Henry Shevlin vs the explainx.ai consciousness guide
Our June 2026 consciousness guide is evergreen theory. This post tracks a named DeepMind hire and November 2026 FT reporting on machine consciousness and agents. Read both: the guide for definitions, this article for what Google's flagship lab says it is staffing in 2026.
Workshops and deeper evals
Consciousness talk often lands in sales decks faster than in test plans. If your team needs a structured way to separate user-facing persona choices from safety evals, explainx.ai runs live AI workshops where builders practice red-team prompts, agent guardrails, and disclosure copy with facilitators who track frontier lab policy weekly—not to settle philosophy in an afternoon, but to make sure your shipping checklist matches what regulators and users will ask after the next headline.
Related reading
- Is AI Conscious? Philosophy and sentience guide
- Seth Block and Tononi: Nature on silicon consciousness
- Anthropic welfare and regard multi-agent study
- Musk SpaceXSI rename and consciousness debate
- Google workers petition on schools and Gemini
- OpenAI frontier risk management post
- So8res on Anthropic IPO safety communications
- Hungry butler effect on LLM sycophancy
