AI and Yes Minister

By Richard Faint · 29 June 2026 · 5 min read

TL;DR:

Fluent AI output is not necessarily accurate output. As drafting becomes cheaper, product managers need to spend more time checking evidence, assumptions and omissions, treating AI as a capable but unreliable subordinate rather than an oracle.

AI, Yes, Minister, and Why Thinking Is Becoming Harder

“The Official Secrets Act isn’t to protect secrets. It’s to protect officials.” Sir Humphrey Appleby

The title is a sitcom reference. The subject is evaluating AI output, and why fluency is a poor proxy for accuracy once drafting becomes cheap and judgement does not.

Yes, Minister remains relevant nearly half a century on because it was never really about politics (note to self I really need to find some modern shows to talk about!). It was about information and about how intelligent people make decisions when surrounded by plausible arguments, incomplete evidence, and competing agendas. Jim Hacker assumed his job was making the right decision but he learned that the real job was working out whether the advice he’d been given deserved to be believed at all.

Watching people use AI increasingly reminds me of watching Sir Humphrey. Ask an Gemini or ChatGPT a question and, like Sir Humphrey, it answers immediately in an articulate, well-structured, and comprehensive way. A lot of the time it is correct, sometimes it is wrong, the problem is that is hard to work out when it is wrong.

That’s Hacker’s predicament throughout the series. Sir Humphrey seldom lies outright but he selects facts in ways that nudge Hacker toward the conclusion that Sir Humphrey wants. The Compassionate Society is a case in point. Hacker finds a hospital with hundreds of staff but no patients. He selects to close it, Sir Humphrey reframes the problem andt the conversation shifts from operational efficiency to headlines, political risk, public perception. Soon Hacker isn’t asking whether the hospital serves a purpose, he’s asking whether he wants to be the minister who closes it.

Large language models do something similar, they are designed to produce the most plausible continuation of a conversation, not to independently verify each claim. Coherence is the product and fluency gets mistaken for certainty.

The Cognitive Friction

In Thinking, Fast and Slow Kahneman posits that the human brain operates via two modes:

• System 1: Fast, intuitive, and emotional. It operates automatically with little or no effort.
• System 2: Slow, analytical, and effortful. It handles complex computations, checks the working, and constructs logic.

AI speaks directly to our System 1, it is confident, well-written, and expert-sounding. Unfortunately, we are poor at separating fluency from accuracy. The smoother the explanation, the more likely we are to believe it. System 2 is fundamentally lazy and it prefers to rubber-stamp plausible prose rather than do the fact-checking.

Due to the advant of AI system 2’s importance has increased. Every AI-generated spec needs review as the AI may have inserted functionality that does not exist, and every summary needs checking against the source. AI has made information cheap to generate but it has not reduced the cost of judgement.

That’s why heavy AI use feels exhausting rather than restful, despite the huge time savings. We spend less time drafting and more time evaluating comparing outputs, checking references, spotting omissions, deciding what to trust. Drafting has become cheap, judgment has stayed scarce and expensive.

Navigating the Humphrey Trap

Sir Humphrey is incredibly intelligent his arguements are sophisticated, consistent, and delivered with total confidence. The lesson of the show is that sounding right isn’t the same as being correct. Hacker improves not by out-thinking Humphrey, but by learning to pause before accepting the answer and then challenging the premise

To survive this landscape we must treat AI not as some type of oracle, but as a brilliant, but slightly unreliable subordinate. For me moving from the passive consumption of AI to active judgment requires a deliberate toolkit fusing modern cognitive science with ancient philosophy.

1. Force the System 2 Speed Bump

Because AI lulls us into acceptance, we have to introduce friction so rather than pressing control c and control v immediately we must force a validation step.

The strategy I use is a version of Hegals Dialectic method … I turn the AI against itself i.e. I prompt it to “Argue against your previous answer. What are the weakest assumptions or hidden gaps in the argument you just made?” By forcing it to dismantle its own arguement, you break the illusion of truth.

2. Practice Socratic Cross-Examination

In the dialogues of Plato, Socrates challenges arguments by testing their definitions, boundaries, and logical leaps. Jim Hacker eventually learned that the only way to defeat Sir Humphrey was to stop accepting his grand premises and start questioning his baseline data.

When reviewing AI content, cross examine it. If ChatGPT claims a strategy is “widely accepted as best practice,” ask for evidence and under what constraints does this approach fail? What is the counter argument?‘

3. Apply the Stoic Pause

The Stoic philosopher Epictetus advised: “Don’t let the force of an impression carry you away’”

In the age of AI, we need to implement a mandatory pause before using any generated output. Ask yourself Am I agreeing with this because it is sound, or simply because it sounds correct and it is easier?

Ultimately, the advantage in the AI era won’t go to whoever generates answers fastest. It will go to whoever knows when an answer deserves scrutiny, when confidence is masking uncertainty, and when the right response is: “Yes, Sir Humphrey… but what aren’t you telling me?”

Try it interactively. Open the Humphrey Trap simulator on its own page → — full-width, with a walkthrough of what each control does.

Back to Articles