Discoverable Doubt

If a senior executive turns to an LLM for help with a decision and, in prompting the AI, expresses doubts about their position, gaps in their comprehension, or general confusion, are they on the record or off the record?

The assumption must be that they are on the record. But has any leader asked? Should senior leaders’ prompt and generation archives be classified as permanent business records? Treated like confidential working papers? Held to the standard of a private conversation?

Current technical and legal architectures place a leader's LLM use on the record. The closed-door office, the executive session, the attorney-client conversation were architectures of privacy that expanded with rank.

A CEO's prompt history carries enough fiduciary risk that the enterprise has its reasons to step in and govern it. Most LLM rollouts, perhaps for this reason, target operations: efficiency, shortcuts, automation.

But pointing LLMs at senior leaders' decision making to help explore uncertainties, challenge consensus and sharpen judgement would likely be the single highest-value application of generative AI in the enterprise. Thinking with AI, however, leaves a permanent, auditable trail of unfinished thinking. A log that could expose mind-changing, testing, backtracking, could make executive judgement look less like the product of experience and intuition and more like someone working through a problem by trial and error.

A leader is presented with a tool that could sharpen their judgement and will log every doubt they express along the way.

Consider a board that approves an acquisition which later fails. Traditionally, plaintiffs rely on hindsight to prove negligence. What if they subpoenaed the CEO's prompt history and the record shows the CEO typed, "I'm not sure I understand why this works," and then approved the deal anyway.

Institutions have always recognised that some thinking serves everyone better when it cannot later be produced. Legal privilege, the seal of the confessional, and clinical confidentiality all operate on the same principle. They rely on a human intermediary bound by rules that forbid disclosure: bar ethics for lawyers, sacramental doctrine for clergy, and professional standards for therapists. Yet for AI, no such protected space exists. An LLM is a third-party vendor's server. You cannot easily grant "privilege" to a cloud service. Until enterprises build models that sit outside the reach of subpoena, the "AI confessional" remains a fiction.

Thinking with AI means thinking in a glass house.

Leaders need personal protection from their own prompt histories, not just corporate protection. Until then, they remain safe by avoiding the tool. Given the pace of improvement, that avoidance is not free. The irony is in plain sight: AI will do its finest work everywhere, except at the very top.