When Google acquired DeepMind in 2014, its co-founder Demis Hassabis did something unusual: he made the sale conditional on governance controls, refusing to let a for-profit corporation have the final say over how his technology would be deployed, [according to Sebastian Mallaby's account in The ...
image from Gemini Imagen 4
Demis Hassabis negotiated governance controls when Google acquired DeepMind in 2014, including an independent oversight board and military-use prohibitions, attempting to ensure he remained the 'commander' of AI development. His 'singleton vision' collapsed through a series of failures: Elon Musk departed to found OpenAI after a secret 2015 meeting, a three-year negotiation dubbed Project Mario to extract binding governance commitments from Google quietly stalled, and the external board that was supposed to provide oversight was gradually eroded, leaving DeepMind effectively under Google's control despite the original deal.
When Google acquired DeepMind in 2014, its co-founder Demis Hassabis did something unusual: he made the sale conditional on governance controls, refusing to let a for-profit corporation have the final say over how his technology would be deployed, according to Sebastian Mallaby's account in The Atlantic. He had spent years building the most capable AI lab on earth, and he intended to be the one who decided what it did next.
That ambition is the subject of a long profile by Sebastian Mallaby in The Atlantic — and it is, at bottom, a story about a plan that collapsed. Not quietly, not all at once, but in a sequence of specific, traceable failures: the secret meeting where Elon Musk decided to go his own way, the three years of internal negotiations that went nowhere, the privacy backlash that sank a flagship health project, and the slow erosion of the outside board that was supposed to keep everyone honest.
The through-line Mallaby traces is what insiders called the "singleton vision": the idea that one outfit, under one commander, with binding constraints on what the technology could be used for, could steer AI toward something resembling safety. Hassabis, the chess prodigy and neuroscience PhD, was supposed to be that commander. The controls he extracted from Google — an outside board of independent reviewers, a strict prohibition on military applications — were supposed to be that constraint, as Mallaby reports.
The plan began fracturing in 2015. Hassabis arranged a secret gathering of philosophers and technologists at Musk's Hawthorne, California headquarters, hoping to lock in potential allies for the singleton vision. Musk listened to the presentations. Then he founded OpenAI. "The gathering marked the moment when Hassabis's safety vision began to crumble," Mallaby writes in The Atlantic. Hassabis had offered a seat at the table; Musk decided to build his own table.
The most consequential failed negotiation was Project Mario — the code name for an effort to extract formal governance commitments from Google, backed, according to Mallaby's account, by a pledge of roughly $1 billion from Reid Hoffman, the LinkedIn founder. Hassabis hired a top legal team and spent three years trying to get Google to accept external oversight with real teeth. Google did not refuse outright. It simply made the process slow enough that the moment passed. If he didn't get control, Hassabis told Mallaby, he would consider spinning DeepMind out. He never got the control. The spin-out never happened.
During this period, Hassabis and his co-founder Mustafa Suleyman also pursued what they hoped would be the moral anchor of the whole enterprise: a partnership with the UK's National Health Service to build an AI system for managing acute kidney disease. If DeepMind could demonstrate that carefully governed AI could improve outcomes for ordinary people, the project would prove the thesis — that external oversight and social benefit went together. The NHS deal transferred 1.6 million patient records to DeepMind. Privacy campaigners objected loudly and specifically. The project fizzled. By 2019, according to Mallaby's account, Suleyman had been pushed out of the company after the NHS data backlash; public reporting at the time confirmed he was placed on leave in August 2019 and later left DeepMind for Google.
The irony is that AlphaFold — DeepMind's system for predicting protein structures, unveiled in 2020 — worked spectacularly. Hassabis and a colleague won the Nobel Prize in Chemistry for it. It is one of the genuine scientific achievements of the decade, and it was produced by the lab that was supposed to be humanity's insurance policy against its own creation.
Hassabis himself seems to have reached conclusions about all of this. "Safety isn't about governance structures," he told Mallaby. "I mean, even if you have a governance board, it probably wouldn't do the right thing when it came to the crunch." He has described his own evolution as a shift from "idealist to realist." He still has values, he says. He just doesn't think the structures he built will enforce them when it matters.
That is a striking thing for the person who spent a decade insisting those structures were the whole point to say out loud. Governance, in this telling, is a seat at the table — not the table itself.
The question the profile leaves open is the one that matters most: could the singleton vision have worked? The honest answer is probably partially, temporarily, and not for the reasons Hassabis thought. Partial because DeepMind's controls did constrain behavior for a while — the military prohibition held, the oversight board met. Temporary because Google's incentives, and then the broader industry's, moved faster than any internal governance could track. And not for the reasons he thought, because the failure wasn't only structural. It was personal: Musk looked at the same information Hassabis did and concluded that no one organization, however well-intentioned, should have that kind of power over the field. He may have been right for the wrong reasons. Hassabis may have been wrong for the right ones.
What Mallaby's profile does not fully answer is whether the Morocco meetings — a secret gathering in a North African bunker that Hassabis reportedly used to vet potential recruits for the safety effort — actually happened. The anecdote appears only in Mallaby's piece. It is vivid and fits the pattern of what Hassabis was trying to build, but it is uncorroborated. The $1 billion Reid Hoffman pledge is similarly Mallaby-exclusive: Hoffman himself has not confirmed the figure. Giskard should flag both before publication.
The piece is worth reading in full because it is about something real that most AI coverage elides: the gap between believing you can control a technology and actually being able to. Hassabis built extraordinary things. He also believed, genuinely, that the right governance structures and the right person in charge could steer AI toward outcomes that didn't destroy what he was trying to build. The record suggests he was wrong. The record also suggests he may have been the last person who had a realistic chance of being right — which makes the wrongness harder to celebrate.
Source: Sebastian Mallaby, "The Man Who Thought He Could Keep AI Safe", The Atlantic, March 2026
Story entered the newsroom
Research completed — 0 sources registered. Atlantic/Mallaby long-read. NOT dupe of story_5102 (deal mechanics). This is philosophy-and-collapse arc: singleton vision → Musk defection at 2015 Ha
Draft (986 words)
Reporter revised draft based on fact-check feedback (1032 words)
Reporter revised draft based on editorial feedback (1032 words)
Approved for publication
Headline selected: Hassabis Built DeepMind to Control AI. He Lost Control Instead.
Published
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 2h 7m ago · 3 min read
Artificial Intelligence · 2h 10m ago · 3 min read
