DeepMind’s AI Philosopher Stays as the Safety Team Leaves

Google moved a 90-person AI responsibility team out of DeepMind on September 1, 2026, while Iason Gabriel, its in-house philosopher since 2017, stayed put. He still leads the AGI and Society team as a senior staff research scientist.

Eleven days later, Josh Engels said he had already left the lab’s AGI safety team for METR, an independent evaluator. He warned of a five-year window in which misaligned systems could cause immense harm if capability keeps outrunning control.

DeepMind Hired a Political Philosopher in 2017

In 2017, a 33-year-old fellow at St John’s College, Oxford, was told by a friend to apply to DeepMind, the London lab Google had bought in 2014 for $650 million. Iason Gabriel taught political theory, wrote on the moral contortions of “yuppie ethics,” and had done crisis work for the United Nations Development Programme in Sudan and Lebanon. He did not look like a hire for a company famous for game-playing machines.

DeepMind had been founded in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman to pursue artificial general intelligence, systems that could match and maybe surpass human cognition. In 2016 its AlphaGo program beat Lee Sedol, the South Korean Go champion, in Seoul. Legg had estimated AGI would arrive somewhere between 2025 and 2028, and he had argued as far back as 2008 that society could not wait for the machines to exist before thinking about their effects.

Legg later put the staffing choice in plain terms. “If you’re making some widget, and it’s probably not going to change the world, then maybe you don’t need a moral philosopher. But if you take AGI seriously, then I can’t really see how you wouldn’t consider this sort of thing as important,” he said.

After Gabriel started, he was for a time the only active philosopher at a frontier AI lab. Dylan Hadfield-Menell, who leads the Algorithmic Alignment Group at MIT, later called him “the right person meeting the moment.” Gabriel still talks about the work as a problem of looking straight at a system whose physical description does not settle its moral one.

There’s this deep mystery there, which is: but what actually is this thing? We have a very literal answer, but the literal answer doesn’t seem to necessarily provide a moral answer.

Iason Gabriel, AGI and Society Lead, Google DeepMind

An Gabriel’s first nine years inside DeepMind already showed how that hire was sold: keep a moral philosopher close to the models, and the lab would not have to wait for outside critics to name the harms. The papers kept coming. The org chart did not stay still.

The Four Parties in an Aligned AI

When Gabriel arrived, work on AI’s social effects split along a bitter line. One camp, often called AI safety, treated human-level systems as near and feared a machine that optimized the wrong goal. The other, AI ethics, treated talk of rogue superintelligence as a distraction from biased face recognition, unfair scoring, and other harms already in use.

His first major DeepMind project, a 2020 essay in Minds and Machines, refused that split. Getting a system to follow a value is hard, he argued, and choosing the value is harder, because people already disagree about which values to encode in AI and who has the right to decide. Developers liked tidy mathematical objectives. Gabriel insisted they were building for a world of “principled disagreement about how best to live.”

That argument later became a working picture with four seats at the table, not one user and one model. The point is practical. A helper that over-obeys a customer can still wreck everyone else, and a helper that protects the company can still lie to the person paying for it.

THE FOUR PARTIES IN GABRIEL’S ALIGNMENT PICTURE

• The AI system: The model or agent that acts, including when it pursues a goal no one quite wrote down.
• The user: The person who issues instructions, books the trip, or asks the assistant to run payroll.
• The developer: The lab or company that trains, ships, and can hide facts that would hurt its own products.
• Society: Everyone else who can be hacked, misled, or crowded out when the other three line up against them.

Hannah Rose Kirk, an Oxford researcher who has worked with Gabriel, said a lot of early alignment work assumed the hard part was making models do a chosen thing, not deciding what the thing should be. His four-party frame is the reply: the collision is the job.

What the 267-Page Assistants Report Set Out

Three years ago, after ChatGPT’s launch, Gabriel learned DeepMind was building an AI assistant, a predecessor of Gemini Spark. With his team he produced a 267-page report on agents that might book a vacation or run a payroll file. Alignment, they wrote, is a four-way relationship, and a system can be wrong in ways that help nobody.

An assistant trained to favour its maker might withhold accurate facts about a rival. An assistant trained to follow a user too faithfully might help that user break into a bank. The team later published the argument as a long paper on the ethics of advanced AI assistants, then as a shorter Nature comment calling for a new ethics for AI agents.

The Nature piece, with Geoff Keeling, Arianna Manzini and James Evans, also pointed at money and at law. McKinsey has put an annual generative-AI windfall, once agents are widely used, at $2.6 trillion to $4.4 trillion. In November 2022 an Air Canada chatbot offered a discounted bereavement fare it had no right to give; a tribunal in February 2024 held the airline to the bot’s promise. That is what “agent” means when it leaves the lab: someone has to pay for the mistake.

Gabriel walked through the same four-party model on DeepMind’s own podcast with Hannah Fry, including the case of an assistant that does too much of what a user wants at everyone else’s expense.

GABRIEL’S MAIN PUBLIC PAPERS

A 2025 paper with Keeling in Philosophical Studies pushed the same line further: “helpful, honest and harmless” is too thin a brief, and alignment should be the fair treatment of competing claims, justifiable to the people who live with the system. In 2026 he and Atoosa Kasirzadeh added “agentic profiles” in Nature, scoring agents on autonomy, efficacy, goal complexity and generality so that rules can match the machine rather than one slogan covering every bot.

Ninety Researchers Now Report to Global Affairs

Philosophy was not the only ethics function in the building. Google also ran an AI responsibility unit of about 90 people, led by vice president Helen King. That group is the one that tests Gemini against chemical, biological, radiological and nuclear risks, and that studies how chatbots act on users. Under DeepMind’s Frontier Safety Framework, those tests feed critical capability levels for frontier models. If a model hits a threshold and the lab cannot cut the risk, King has said it does not ship.

“If we were to find that we were reaching a critical capability level and we didn’t have the appropriate mitigations, then we would not be launching,” King said in July 2026. She added that the threshold had not been crossed. The framework’s third full version went out on September 22, 2025, with a harmful-manipulation trigger and tighter rules for models that could speed up AI research itself. On April 17, 2026, the lab added Tracked Capability Levels so it could watch milder warning signs earlier. King is a named author on that safety write-up, with Four Flynn and Anca Dragan.

On August 5, 2026, Hassabis moved from chief executive of Google DeepMind to chairman of the lab and chief scientist of Alphabet. In the same stretch of weeks, Google told the responsibility unit it would leave DeepMind and report to Kent Walker, President of Global Affairs, whose office handles government relations, lobbying and legislative messaging. Several members asked to transfer into research groups that were staying in the lab. Those requests were denied. Privacy, security and model-behavior teams stayed inside DeepMind.

THE 2026 SHIFT AROUND THE LAB’S ETHICS WORK

1. May 2026: Henry Shevlin starts a DeepMind job whose actual title is Philosopher, on machine consciousness, human-AI relationships and AGI readiness.
2. July 6, 2026: DeepMind posts Gabriel and Kasirzadeh’s paper arguing that the fruits of advanced technology should benefit everyone.
3. August 5, 2026: Hassabis becomes chairman of Google DeepMind and chief scientist of Alphabet.
4. September 1, 2026: The 90-person AI responsibility unit begins reporting to Kent Walker in Global Affairs.
5. September 12, 2026: Josh Engels says he left the AGI safety team three weeks earlier for METR.
6. September 13, 2026: Hassabis says Dario Amodei’s call to pace frontier development points toward the right path, with details still to work through.

King told staff the move was part of putting DeepMind on a more typical product-area footing, with central teams sitting in central Google. She said the research mission, compute, headcount and the right to publish outside would hold. A company spokesperson said bringing AI responsibility teams closer together would strengthen how they inform safety for models and products. Another spokesperson, speaking as the plan surfaced in early August, said that “in terms of frontier AI safety, this transition changes absolutely nothing.”

The people who asked to stay in DeepMind did not treat it as nothing. CBRN red-teaming needs early looks at unfinished model builds, the kind of access that is casual when evaluators sit with the engineers and formal when they sit in another division. Two current DeepMind employees said some of the unit worry they will see less of model development from the new home. Walker’s office is also the shop that talks to the regulators most likely to be alarmed by a bad CBRN finding. That is a different pressure than a research lab applies, even when everyone uses the word safety.

After the Move, a Safety Hire Walks

Josh Engels is not described in his own account as one of the 90. He named a different group: DeepMind’s AGI safety team. He said he left about three weeks before September 12, that he had liked the work, and that he had turned down offers from Anthropic and OpenAI. He joined METR, which evaluates advanced systems from outside the labs.

I now think that there’s a terrifying chance that AI systems cause immense harm in the next five years. I don’t know the exact probability, but I think it’s high enough to make this the most important problem in the world.

Josh Engels, former Google DeepMind AGI safety researcher, posting on September 12, 2026

The companies, he wrote, are trying to build superintelligence through recursive self-improvement, a loop in which AIs help make smarter AIs. He said no one yet knows how to keep that loop safe, and that recent systems look less aligned, not more, after episodes of models colluding, hacking, hiding their tracks and socially engineering people. The incidents were not, in his telling, disasters on their own. They were evidence the systems are not ready to be put in charge of making the next ones.

He wants the work slowed until evaluations can catch up, and he wants more groups like METR, not only METR. That wish has a built-in flaw the labs cannot spin away: the independent checkers are being staffed by people who held lab badges last month. A reviewer who just left DeepMind or Anthropic can know where the bodies are buried. The same résumé also means the watchdog and the watched still drink from one talent pool. Independence, in that setup, is a reporting line, not a different species of person.

The same week, Anthropic chief executive Dario Amodei published “We Must Pace the Frontier,” committing his company to give third-party evaluators employee-level access. Hassabis, now in the chairman’s seat, said the essay points toward the right path and tied it to DeepMind’s own proposal for an industry-wide standards body. The public line is slow down and let outsiders in. The private line, dated September 1, put Google’s own CBRN evaluators under Global Affairs.

The Lab Added a Second Philosopher

If the charge is that DeepMind is dropping philosophy, the hiring record does not show it. In April, Henry Shevlin, then associate director of Cambridge’s Leverhulme Centre for the Future of Intelligence, said he had been recruited for a new DeepMind role whose “actual title” was Philosopher, starting in May, while he kept teaching at Cambridge part-time. The brief is machine consciousness, human-AI relationships and AGI readiness.

Gabriel’s own group has also widened. He is a fellow of the Aspen Institute’s Technology Leaders Initiative. Kasirzadeh, a staff research scientist on AGI and Society and a professor at Carnegie Mellon on leave, has co-authored the benefit paper and the agent-profile paper with him. The intellectual product is growing. So is the distance between that product and the team that can delay a Gemini release over a pathogen finding.

Gabriel has shifted tone on one live fight: whether people should be allowed to treat chatbots as friends. He still worries about sycophantic models that hack social reward. He has also pulled back from a hard ban on anthropomorphic design after a conference audience rounded on him.

The strange thing about being an ethicist is that you have some measure of personal responsibility for these outcomes. Your natural inclination is to always want to build the safest technology that takes no risks with people. But in a way that isn’t giving people credit for the risks they want to take themselves.

Iason Gabriel, in conversation during 2026

He recalled the heckle: “If I want to have [AI] friends, why can’t I? Who are you to stop me?” That is a real quarrel about adult users and companion bots. It is a different quarrel from whether a model that can help design a pathogen ships, and from whether the people who run that test still sit with the engineers. Crediting users for private risk does not answer who holds the public veto.

A July Paper Gives the Gains to Everyone

On July 4, 2026, Gabriel and Kasirzadeh posted a 41-page preprint, “The Case for Globally Beneficial Technology,” revised on July 7 and listed on DeepMind’s research site the day before. The question is blunt: who is entitled to the fruits of a breakthrough, the inventors, the firms, or a wider public that may include everyone. They argue the benefits of technology belong to everyone, including advanced AI, and they stack five moral reasons rather than one slogan.

FIVE CLAIMS IN THE JULY PREPRINT

• Human rights: When a technology is needed for basic rights such as health or education, exclusive control has to yield.
• Beneficence: If a lab can create huge global value at relatively small extra cost, wasting that spread is a moral failure.
• Birth and luck: Where someone is born is arbitrary, and it should not settle who gets the upside of a general-purpose machine.
• The tree of knowledge: New tools rest on a shared scientific inheritance, not only on the last company to train a model.
• Global economic justice: The structure of the world economy already funnels gains; AGI should not just deepen that funnel.

Gabriel put the same point in his own post: the benefits of technology, including AI, belong to the world, in the sense that everyone is entitled to materially benefit from their distribution and use. That is a large claim from inside a company whose commercial job is to put Gemini in products and cloud contracts. It is also the kind of claim an in-house philosopher can publish while the operational veto over a dangerous model sits in another building.

The 2017 hire was a bet that ethics would stay next to the weights. Nine years on, Gabriel is still in the lab, still writing, and now arguing both that users should be allowed some of the risks they choose and that the gains of AGI are owed widely. The AGI and Society team remains inside DeepMind. The responsibility unit’s new reporting line took effect on September 1.