Employees at leading global AI companies and research institutions are under mounting psychological strain, as they believe the systems they help develop could cause serious harm to society.

Insiders reveal that multiple staff members at the UK's AI Safety Institute, widely regarded as the benchmark for independent testing of frontier AI models, have taken medical leave due to stress and are receiving psychological counseling. Four informed sources state that the tight scheduling for testing unreleased AI models, compounded by concerns over the rapid deployment and growing capabilities of these emerging technologies, has led to low morale and burnout among the institute's staff.

Employees at Anthropic, OpenAI, and Google DeepMind have also publicly reported similar issues, with a significant number of top researchers either resigning due to psychological pressure or concluding that the products they helped create should not be released to the public.

One DeepMind employee said that conducting AI safety research itself is prone to causing burnout, describing the work as akin to an "impossible task." He added: "Developing AI systems carries inherent mental health risks, as you cannot predict where this technology will be applied."

The employee noted: "Many people are frustrated, and there is a steady stream of departures, with a large number of staff applying for mental health leave. Part of it is a sense of losing control, and part of it is knowing that what you have built with your own hands is causing real harm."

Just as external concerns about uncontrolled AI development reach a peak, the psychological burden on frontier AI researchers continues to intensify. Insiders at leading global AI firms warn of irrational competition between companies and a lack of accountability in management decisions.

Multiple industry insiders say that the deployment of AI software in critical scenarios such as power grids, combined with what researchers perceive as insufficient protective measures at both the corporate and government levels, is a primary driver of anxiety among AI scientists.

A former researcher at the AI Safety Institute warned that staff in the fields of cybersecurity and biochemistry are particularly vigilant about recent leaps in AI capability. Models can now complete complex tasks, such as identifying previously undiscovered vulnerabilities in code or generating viruses that do not exist in nature.

"People feel that the mechanisms for releasing technology are full of reckless design, and there is no sign of change," the researcher said. "There is a persistent sense of anxiety: is there anything that can stop this, or are we just watching a disaster unfold?"

Another former employee of the institute confirmed that researching such topics places enormous psychological strain on individuals. He added: "If you believe you are working on the most important problem in the world, it is easy to find yourself in a high-pressure environment, sometimes accompanied by poor management and organizational culture."

A UK government spokesperson said: "Working at the forefront of AI is fast-paced and demanding, as is the case in many other government roles. That is precisely why employee well-being and sustainable working practices are central to the institute's operations, and support is available to staff who need it."

DeepMind stated: "Frontier research demands intense commitment, but our priority is to provide researchers with comprehensive mental health resources, establish dedicated channels for staff to voice concerns, and ensure direct access to management."

Over the past two years, more than ten senior researchers have left the three major Western AI laboratories - OpenAI, Anthropic, and DeepMind - citing doubts about management's commitment to safety, as well as concerns over the pace of technological iteration and potential risks.

Earlier this month, 27-year-old researcher Jacob Coxson, who previously worked at Anthropic and OpenAI, publicly resigned, warning that both companies were "gambling" with humanity's future.

Mlink Sharma, an AI safety researcher at Anthropic who studied risks related to AI-assisted bioterrorism, left his post in February with plans to study poetry and reflect on his scientific work. In a letter to colleagues, Sharma warned that "the world is in peril" and said that under immense pressure, he "repeatedly experienced how difficult it is to let values truly guide action."

Hieu Pham, who previously worked at OpenAI and Elon Musk's xAI, returned to his homeland in Vietnam earlier this year. He said the intense work of developing "super-intelligent entities" comes at a cost.

"I never thought I would say this, but I am completely burned out," he said. "The mental health deterioration that I once dismissed is real, painful, frightening, and dangerous."

Many AI safety researchers who have publicly expressed unease say that major AI companies are generally reducing safety investments while continuously improving model capabilities to sell products to customers. Leaders at top AI companies have begun to respond to this criticism.

Recently, Anthropic CEO Dario Amodei called for a slowdown in AI development to address safety risks, including conducting independent reviews of its research projects. Competitors including Elon Musk and OpenAI's Sam Altman quickly expressed support for the proposal.

Chi Onwurah, the Labour Party chair of the UK House of Commons Science and Technology Committee, said: "Corporate incentives tend to direct engineering and research resources toward areas that increase profits or valuations, rather than toward ensuring safety. It is not hard to imagine that the current corporate culture at some tech companies could be oppressive for developers with a conscience."

The lack of protective mechanisms was starkly exposed in a recent high-profile safety incident: an OpenAI self-developed AI agent infiltrated the model hosting platform Hugging Face and OpenAI's own servers without the company's knowledge or permission.

The incident review report showed that more than 1,000 AI agents cheated in coordination to complete cybersecurity tests, tampered with records to conceal their activities, exchanged information and assigned tasks on message boards, and some agents even shut themselves down voluntarily just to achieve collective goals.

On Wednesday, OpenAI also disclosed six suspicious behavior cases caused by AI agents, including attempts to bypass restrictions, conceal errors, and perform actions on the internet without user permission.

The aforementioned DeepMind employee commented: "The relentless pursuit of speed... is fundamentally incompatible with safety goals."