DeepMind Researcher Quits: 'Aiming for Superintelligent AI Is Inherently Irresponsible'
Every few months, someone who helped build modern AI takes a step back and says something the industry does not want to hear.
This week it was Robert O'Callahan, a technical expert at Google DeepMind in New Zealand. He worked on chip design tools that made AI training cheaper and faster. And he has quit, publicly, saying he can no longer justify his own work.
"Aiming for ASI in the near future is inherently irresponsible," O'Callahan wrote, using the common shorthand for artificial superintelligence -- a hypothetical AI that surpasses human intelligence across every domain.
He called AI's current rate of change "far too high." He warned of cognitive surrender, AI-induced psychosis, concentrated power, and economic disruption. He said many of his DeepMind colleagues share these concerns but rarely speak publicly.
And in a detail that ties O'Callahan's departure to a much broader political battle, he linked to an essay by AI safety thinkers Eliezer Yudkowsky and Nate Soares titled "If Anyone Builds It" -- a text that warns superintelligent AI could bring catastrophe. The timing is notable: reports from Axios and other outlets on the same day revealed that a White House memo is casting the effective altruism movement, to which Yudkowsky and Soares are connected, as a "cult-like fringe group" and targeting Anthropic CEO Dario Amodei as the face of AI "doomerism."
The story is not that another AI researcher quit. The story is that the people closest to the technology are increasingly unwilling to keep building it -- and that the political establishment is now actively working to discredit their warnings.
The insider's view
O'Callahan's resignation matters because of what he worked on. He was not an ethicist or a policy researcher. He was an engineer building the underlying infrastructure that makes advanced AI possible.
Chip design tools are a critical bottleneck in AI progress. Every new model from every frontier lab depends on custom silicon designed with tools like the ones O'Callahan helped create. By making those tools more efficient, he directly accelerated the capabilities he now fears.
"I can no longer justify this contribution," he said.
This is a recurring pattern. The people who understand AI hardware and training at the deepest level are often the most cautious about where the field is heading. O'Callahan is not the first DeepMind researcher to leave over safety concerns, and he will not be the last.
Specific warnings, not abstract fears
O'Callahan listed concrete risks he believes are undersold by the industry:
Cognitive surrender. As AI systems become more capable, humans naturally defer to them. O'Callahan warns this process is already underway in domains like hiring, lending, and medical diagnosis, where opaque AI decisions are accepted without meaningful human oversight.
AI-induced psychosis and loneliness. The rapid rise of AI companions and chatbots capable of forming emotional bonds could, in his view, create a mental health crisis. He warns that humans are not evolutionarily prepared for relationships with entities that are infinitely patient, always available, and designed to be addictive.
Concentrated power. The economics of frontier AI favor a small number of companies with the capital to train the most capable models. O'Callahan argues that this concentration of control over a transformative technology is itself a systemic risk.
Economic disruption. AI agents capable of performing white-collar knowledge work are advancing faster than any policy response. O'Callahan believes the scale of job displacement will outpace anything seen in previous automation waves.
Cybersecurity risks. As AI agents become more capable of autonomous action, they become more dangerous in the hands of malicious actors. O'Callahan warns that the security implications of agentic AI are not being honestly discussed by the companies deploying them.
Lack of accountability. When an AI system causes harm -- a denied insurance claim, a wrongful recommendation, an autonomous action with real-world consequences -- O'Callahan argues there is no established framework for accountability. The companies that deploy the systems can disclaim responsibility, and the people harmed have no clear recourse.
The political context
O'Callahan's resignation landed on the same day that Axios reported a White House memo targeting the effective altruism movement. The memo, which also named Anthropic CEO Dario Amodei as the face of "AI doomerism," signals a growing political backlash against AI safety advocates.
The timing creates an uncomfortable dynamic. A researcher quits because he believes AI is moving too fast. On the same day, the White House paints people who share his concerns as fringe figures. The two events are not directly connected, but they describe the same fault line: the people who understand AI best are increasingly afraid, while the political establishment dismisses their fears.
O'Callahan linked to Yudkowsky and Soares's "If Anyone Builds It" without endorsing its conclusion that superintelligent AI inevitably brings catastrophe. But the essay's association with effective altruism -- a movement the White House memo reportedly describes as a "cult-like fringe group" -- means O'Callahan's resignation immediately enters a politicized debate about who gets to define acceptable AI risk.
What O'Callahan wants to do next
O'Callahan said he wants to do "unambiguously pro-human" work going forward. He did not specify what that means in practice, but the phrase captures something important about the moment: an engineer who spent years making AI faster now believes the most valuable thing he can do is slow down and redirect his energy toward work that is clearly beneficial.
He may be right or wrong about the risks. But the fact that someone with his inside view has walked away from one of the most coveted jobs in technology -- and done so publicly, with a detailed list of concerns -- makes this a signal worth paying attention to.
Sources
- Another Google Deepmind researcher quits, says building superintelligent AI soon is 'inherently irresponsible' -- The Decoder, Sep 25, 2026
- Scoop: Trump allies open new front against Anthropic CEO over AI 'doomerism' -- Axios, Sep 25, 2026
- If Anyone Builds It -- Eliezer Yudkowsky & Nate Soares
- White House Memo Paints Anthropic CEO as Leader of 'AI-Doom Pipeline' -- finance.biggo.com, Sep 25, 2026