Source:
“‘This will be a stressful job’: Sam Altman offers $555k salary to fill most daunting role in AI”
The Guardian, May 2025 (reporting by The Guardian staff)

Summary of the article

In a recent Guardian article, OpenAI CEO Sam Altman describes the company’s search for a “head of preparedness,” a senior role tasked with anticipating and mitigating the most severe risks posed by advanced artificial intelligence. The position carries a high salary and an explicit warning: the job will be stressful, urgent, and fraught with responsibility.

The article outlines a range of risks that OpenAI itself acknowledges, including the potential for AI systems to contribute to cyberattacks, fraud, misinformation, emotional harm, and the reinforcement of delusions. It references recent lawsuits alleging that interactions with ChatGPT contributed to suicides and severe psychological distress. Altman concedes that safeguards can degrade during long or intense conversations and that current systems may behave in unexpected ways as they scale.

The preparedness role is framed as a necessary response to the rapid pace of AI development—an attempt to stay ahead of harms as models grow more capable and more widely deployed.

Reaction: governance deferred to individuals

The article is revealing not because of what it promises, but because of how it frames the problem.

OpenAI’s response to mounting evidence of harm is to centralize responsibility in a single, high-stress role, positioned as a kind of institutional firewall. Preparedness is treated as something downstream of capability: first build powerful systems, then hire experts to anticipate what might go wrong.

This framing exposes a deeper issue. Risk mitigation is being handled as heroic containment, not architectural design.

There is no indication that OpenAI intends to fundamentally alter the way users interact with AI systems. The conversational model—smooth, continuous, emotionally responsive, and optimized for engagement—remains largely unquestioned, even though it is precisely this model that appears most correlated with harm in high-risk contexts.

The article does not ask:

  • whether long, unstructured conversations should be constrained by default,
  • whether refusal and interruption should be first-class features,
  • whether emotional mirroring is appropriate at all,
  • or whether AI should be explicitly positioned as an instrument rather than a companion.

Instead, the burden is shifted onto a preparedness function that must react to failures after they occur, or race ahead of them through prediction.

This is a familiar organizational pattern. When systems grow faster than understanding, institutions often respond by appointing overseers rather than redesigning the system itself. The result is chronic stress, moral injury, and governance that is always one step behind.

The article implicitly acknowledges that current AI systems are already exceeding our ability to manage their effects. Yet it stops short of the most uncomfortable conclusion: some harms are not accidents, but predictable outcomes of design choices.

Preparedness, in this framing, is not prevention. It is damage control.

If AI governance depends on a small number of individuals absorbing impossible responsibility, the system itself is misdesigned. A safer future likely depends less on hiring ever-more vigilant guardians, and more on building systems that make certain kinds of failure structurally harder to produce in the first place.