Employees tasked with identifying the risks of advanced AI models and preventing unintended consequences face a serious contradiction: as the technology advances rapidly, they must also assess the potential consequences of that same technology; yet, according to some staff and researchers, there is not always enough time to carry out these tests.
This situation has also been reported at the UK AI Safety Institute (AISI). This government body is charged with examining the capabilities and risks of advanced AI models and testing risk-reduction measures. AISI has said it has more than 100 technical staff and evaluates advanced models before and after public release.
However, according to published reports, some employees at the institute have taken sick leave and used counselling services because of work-related pressure and stress. Four informed sources also claimed that time limits for evaluating models and concern about the rapid increase in the capabilities of new systems have contributed to lowered morale and burnout among some staff.
When safety becomes an "impossible mission"
This pressure is not limited to government bodies. Staff and researchers at companies such as Anthropic, OpenAI and Google DeepMind have also spoken in recent months about the strain of working in AI safety. Some senior researchers have left their companies citing concern about the speed of technology development or the belief that systems being built might be released to the public before protective measures are ready.
One Google DeepMind employee described the situation psychologically as an "impossible mission." According to them, uncertainty about how this technology will be used in the future, especially in areas such as critical infrastructure and military applications, can create significant psychological pressure.
They also said that a feeling of losing control and worry that one may be building technology that could harm people in some circumstances are factors that can lead to resignation or long-term leave.
These concerns have increased as new models are used not only to generate text or images but are expanding capabilities in areas such as programming, cybersecurity, performing multi-step tasks and scientific research. AISI itself has listed areas such as cyber misuse, controlling model behaviour, autonomy and systems' ability to bypass control mechanisms among the main topics of its research programme.
Researchers quitting; a sign of a rift in the AI industry
In the past two years, more than 12 prominent researchers have left three major AI labs - OpenAI, Anthropic and Google DeepMind. Concerns about the prioritisation of safety, the speed of model development and the potential consequences of public deployment of the technology have been among the reasons cited in relation to some of these departures.
For example, Jacob Coxon, a researcher at Anthropic who previously worked at OpenAI, resigned in September 2026 and warned about the pace of development of advanced systems and the inadequacy of protective measures. This resignation followed a wave of internal debate among staff at several major labs about the need for increased controls and oversight of AI development.
Meanwhile, Google DeepMind has emphasised that advanced research in this area requires intense effort, but that researcher welfare and providing clear channels for raising concerns are priorities. The company has also stressed in official positions the responsible and safe development of AI systems.
The UK government has also said that working at the forefront of AI research, like many other government roles, can be fast-paced and gruelling, and therefore employee wellbeing and job sustainability are pillars of AISI's activities.
The Financial Times, in its report based on interviews with staff and researchers in the field, examined this psychological pressure in the context of broader concerns about increasing competition among AI companies. At the same time, other reports also indicate growing researcher concern about the pace at which model capabilities are developing and the difficulty of creating control mechanisms that keep up with that pace.
In such circumstances, the issue is not limited to the question of how powerful AI will become; it has become equally important to ask whether safety, regulatory and organisational mechanisms can advance at the same pace as these systems' capabilities.