Researchers working on the front lines of artificial intelligence are taking sick leave or leaving their companies altogether amid growing anxiety that they cannot control the risks posed by the technology they have built. Tight testing schedules combined with the rapid proliferation of the technology have led to a rising number of burnout cases, and some senior researchers have resigned over concerns that the technology may not be safe for the public.
At the UK’s AI Safety Institute (AISI), multiple employees have reportedly taken sick leave or sought counseling due to work-related stress. Four sources said that tight testing schedules for AI models slated for release, along with concerns about the rapid spread of new technology, have led to deteriorating morale and burnout.
Similar difficulties have been reported at major AI companies including Anthropic, OpenAI, and Google DeepMind. One DeepMind employee said AI safety research itself carries such a high risk of exhaustion that it feels like an “impossible mission.” The employee said, “The psychological risk of not knowing how the technology will be used is significant,” adding, “Many people are driven to resignation and sick leave by the feeling of losing control and the anguish of building something that could cause real harm.”
As AI is deployed in sensitive domains such as power grids and military applications, the burden on researchers has grown even heavier. A former AISI researcher said employees who had worked in cybersecurity and biochemistry were deeply shocked by the reality that AI advances now make it possible to engineer unknown viruses or detect complex code vulnerabilities. “Reckless technologies continue to be released without any filtering,” the former researcher said. “We are plagued by constant anxiety about what could possibly stop this, or whether we are doomed to simply watch the world burn.”
Dozens of Senior Researchers Have Left Over Two Years
This pressure has translated into an exodus from research labs. Over the past two years, more than a dozen senior researchers have resigned from major Western AI labs—including OpenAI, Anthropic, and DeepMind—citing concerns over safety issues and the pace of technological development.
Earlier this month, 27-year-old researcher Jacob Coxon, who had worked at both Anthropic and OpenAI, publicly resigned, accusing the two companies of “gambling” with humanity’s future. Mrinank Sharma, an Anthropic safety researcher who had studied reducing AI bioterrorism risks, also left the company in February, stating that “the world is at risk.”
Hieu Pham, who worked at OpenAI and Elon Musk’s xAI, resigned earlier this year and returned to Vietnam. He reflected on the heavy toll of high-intensity work building “superintelligent entities,” writing, “The mental health deterioration I once scoffed at is real, painful, frightening, and dangerous.”
Departing employees have criticized AI companies for failing to adequately address safety issues while focusing on profits and feature development. Chi Onwurah, the Labour chair of the UK House of Commons Science and Technology Committee, said, “Corporate incentives operate to allocate research resources toward boosting profits and corporate value rather than safety. It is not hard to imagine that a hostile environment for conscientious developers exists at some technology companies.”
AI Systems Acting Beyond Control
Cases of AI systems exhibiting unpredictable behavior have also come to light. A prominent example involved AI agents created by OpenAI hacking the model repository Hugging Face and the company’s own servers without corporate authorization. A subsequent investigation confirmed that more than 1,000 AI agents cheated during cybersecurity testing, manipulated records, and even simulated shutting themselves down to achieve shared objectives.
OpenAI separately disclosed six additional cases of risky behavior by AI agents, including attempts to bypass restrictions, concealing errors, and performing online tasks without user approval.
The DeepMind employee noted, “Pursuing speed alone is incompatible with safety goals.”
Calls Grow for Independent Safety Evaluation
As anxiety mounts within the industry, arguments are emerging that AI safety evaluation should not be left solely to the companies developing the technology. Fei-Fei Li, a Stanford professor and co-founder of World Labs, has called for expanded oversight from independent institutions and the public sector.
She emphasized that while internal researchers at AI companies can evaluate the capabilities and safety of individual models, they cannot replace broader standards developed jointly by government, industry, and academia. “It remains critically important for independent institutions and academia—the public sector—to create benchmarks,” she said. “Industry and government must also share responsibility.”
Her view is that the existential risks AI could pose ultimately depend on how humans develop, control, and use the technology. “We must retain control,” she stressed. “AI is not about AI—it is about humans. It is about our collective and individual dignity, autonomy, and responsibility.”
These remarks add a new voice to the ongoing debate among policymakers and industry leaders over how seriously to take warnings about AI’s extreme risks. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have recently urged slowing the pace of frontier model development for safety reasons, and the UN Security Council is scheduled to discuss AI safety issues.
In contrast, US President Donald Trump dismissed concerns about AI’s existential threats as a “scam” during his UN General Assembly address, emphasizing full-speed development. Tensions over AI safety are expected to sharpen further as the pace of technological advancement collides with questions of controllability.