Former Google DeepMind Security Researcher Alex Turner Supports Jacob Coxon
Former Google DeepMind security researcher Alex Turner publicly supported Jacob Coxon, stating that many researchers believe they are building systems that could potentially kill everyone on Earth, and that thinking about how to prevent this was part of his daily work.
Turner left DeepMind in June without taking a new position, subsequently turning to independent alignment and safety research. He worked at the company for over two years, focusing on assessing and attempting to mitigate the catastrophic risks posed by cutting-edge systems.
He previously explained that the main reason for his departure was not an abstract slogan, but rather Google's signing of an AI collaboration with the Pentagon, allowing its use for classified operations without any written constraints against lethal autonomous weapons or mass surveillance. He claimed to have fought internally for months, witnessing ethical commitments fail under pressure, and was unable to continue working after the contract was signed.
In 2018, DeepMind and relevant officials promised not to participate in or support the development or use of lethal autonomous weapons. Turner believes that after the deal was finalized, the only honest options left were commitments, explanations, or resignations, with silence not being one of them. He also mentioned that the defense department had pressured other labs to hand over unrestricted models due to economic consequences.
He quantified the probability of mass death separately: defining P(Doom) as the probability of AI causing at least one billion deaths before 2050, estimating it to be around 25% to 30%, with about 10% attributed to alignment failures and the rest more from misuse. He acknowledged that progress in alignment research has accelerated beyond what he expected during his PhD, but the timeline has been significantly compressed.
From a market mechanism perspective, this is a talent and credibility issue rather than a product release: after the internal veto power for safety positions failed, researchers priced risk by resigning; military contracts and cloud business were absorbed into government budgets, weakening the safety brand premium by internal witnesses. The beneficiaries are defense procurement and computing power contracts, while the pressured side is the lab governance structure that relies on "principled constraints" for recruitment.
After Coxon's departure, the remaining alignment lead at Anthropic also publicly acknowledged the existence of extinction risk estimates. Turner's support expanded the narrative of "both sides losing pre-trained personnel" into "the safety position itself also considers global extinction as a work object," extending frontline testimony from capability competition to governance commitment breakdown.
Source: Public Information
ABAB AI Insight
Turner's trajectory reflects the transition from the alignment academic circle to a major tech company's safety role and then back out: a PhD in alignment from Oregon State University, a postdoc at Berkeley's Center for Human-Compatible AI, and then to DeepMind's safety research. He was not a core operator in the pre-training competition but was hired to envision how systems could go out of control. He left in June due to the Pentagon contract and endorsed Coxon in September, indicating that the two crises are the same in his view: the lab acknowledges the potential to kill everyone while handing models over for military and capability competition.
The capital path is clear. Cloud and model access to defense contracts translates to scaled orders and "legitimate combat use" status; safety research receives internal budgets and principled texts. When contracts conflict with the 2018 commitment against mobile robot lethality, resources flow toward contracts. After attempting to block the deal internally, Turner could only take his veto power outside the company. Independent research cannot match the same computing power, so resigning changes the testimony, not the training schedule.
A similar structure is seen in the 2018 Google employee opposition to Project Maven, leading to adjustments in the company's principled texts, and the contrast of whether Anthropic can hold the red line under defense pressure. The industry phase has shifted from ethical charters to binding national demands: whoever can deliver models into classified networks can secure the next round of procurement.
This represents a transfer of pricing power before regulatory changes: risk pricing power has shifted from safety teams to procurement contracts. The mechanism is that principles lack enforceable clauses; once a national client bids, internal safety roles can only choose to stay and execute or leave and testify.
ABAB News · Cognitive Laws
The daily work of safety roles often involves addressing extinction risks.
When principles meet procurement, what remains are mostly explanations.
Resignation testimonies can price risks but cannot determine training schedules.