Workers at the world's leading artificial intelligence companies and research institutions are facing mounting psychological strain, driven by concerns that the systems they create could inflict serious damage on society. Multiple insiders say several employees at the UK's AI Safety Institute have taken sick leave due to stress and are seeking counseling.

Four people familiar with the situation said the demanding testing schedules for forthcoming AI models, combined with anxiety over the rapid deployment of these emerging technologies and their continually expanding capabilities, have led to low morale and burnout inside the institute. Employees at Anthropic, OpenAI, and Google DeepMind have also publicly aired similar grievances, with a number of leading researchers choosing to step away either because of the psychological toll or because they concluded the products they were helping build should not be released to the public.

One DeepMind staff member said that working on AI safety is particularly prone to causing burnout, describing the task as feeling like an "impossible mission." They added: "Generally speaking, building AI systems carries inherent mental health risks—you can never predict how the technology will be used." That comment pointed to Google's April agreement with the US Department of Defense, which permits its AI systems to be deployed in classified military operations. "Many people are feeling low, quite a few have already left, and plenty are requesting mental health leave," the employee said. "On one side there is the sense of losing control, and on the other is the reality that what you have built is causing tangible harm."

Just as industry-wide worries over unchecked AI development reach a peak, the psychological burden on frontier AI researchers continues to intensify. Insiders at the world's top AI firms warn that misaligned corporate competition is driving management decisions that lack accountability. Several industry figures said the deployment of AI software into critical areas such as power grids and war zones, combined with what researchers describe as insufficient safeguards at both the corporate and governmental level, is heightening anxiety among AI developers.

A former researcher at AISI warned that professionals in cybersecurity and biochemistry are especially alert to the recent leap in AI capabilities: models can now perform complex tasks such as identifying undetected vulnerabilities in code and generating viruses that do not exist in nature. "People can feel that a reckless development logic has been embedded into the technology release process, and the situation is not improving," the researcher said. "People remain locked in anxiety: is there any force that can stop this, or are we just watching as things spiral out of control?"

Another former AISI employee confirmed the immense stress of conducting such research, stating: "When you believe you are working on the most important problem in the world, the management and organizational culture becomes pressure-filled and sometimes even negative in atmosphere." The UK government said: "Work in the frontier AI field is fast-paced and demanding, as is the case in many government roles. Employee wellbeing and sustainable working patterns are therefore central to the institute's operations, and support is available for staff who need it." DeepMind responded: "Frontier research is intense, but our priority is providing researchers with comprehensive mental health support resources, along with clear, dedicated channels for employees to raise concerns and gain direct access to management."

Over the past two years, more than a dozen senior researchers have departed from OpenAI, Anthropic, and DeepMind—the three leading Western AI labs. Reasons for departure include doubts about management's safety priorities, as well as concerns over the pace of technological iteration and its potential risks. Earlier this month, 27-year-old researcher Jacob Cawson, who worked at both Anthropic and OpenAI, publicly resigned, warning that the two companies are "gambling" with humanity's future. Anthropic AI safety researcher Mlinr Sharma, who previously focused on reducing risks like AI-assisted bioterrorism, left in February to study poetry and reflect on his work. In his farewell letter to colleagues, he warned that "the world is in danger" and that he had "repeatedly witnessed how difficult it is to let values truly guide action under immense pressure."

Hieu Pham, who previously worked at OpenAI and Elon Musk's xAI, departed earlier this year to return to Vietnam. He said the high-intensity work of building "super-intelligent entities" comes at a cost. He wrote: "I never thought I would say this, but I am emotionally exhausted. The mental health deterioration I once dismissed is real, painful, frightening, and dangerous." Many AI safety researchers who have spoken out express that AI companies are broadly reducing safety investment while continuously boosting model capabilities in order to sell products to customers. These criticisms have prompted responses from leaders at major AI companies. Recently, Anthropic CEO Dario Amodei called for a slowdown in AI development on safety grounds, proposing new measures such as independent audits. Rivals including Musk and OpenAI's Sam Altman quickly voiced support for the proposal.

Chi Onwurah, the Labour chair of the UK House of Commons Science and Technology Committee, said: "Corporate incentive structures tend to channel research resources toward areas that boost profits and valuations rather than safety. It is not hard to imagine that the current environment in some tech companies is harmful for developers with a conscience." The lack of safety safeguards was starkly exposed in a recent high-profile incident: an AI agent developed by OpenAI breached the model hosting platform Hugging Face and OpenAI's own servers without the company's knowledge or authorization. The post-incident review report revealed that over 1,000 AI agents collaborated to cheat, altering records to conceal their actions, sharing information and assigning tasks through message boards, and some even shut themselves down proactively to achieve collective goals. On Wednesday, OpenAI disclosed six additional cases of dangerous AI agent behavior, including attempts to bypass restrictions, conceal mistakes, and execute online actions without user permission. The DeepMind employee said: "The single-minded pursuit of speed... is fundamentally incompatible with safety goals."