# We keep firing the honest fool

Niels Langereis, 31 August 2026. First published on LinkedIn: https://www.linkedin.com/pulse/we-keep-firing-honest-fool-niels-langereis-wdrce

> A model that tells you what you want to hear is not a bug in the technology, it is a preference we trained into it, and one organisations have been training into people for a lot longer. Amsterdam, 1637: the man who called the tulip market right got run out of the room.

A couple of weeks ago I wrote one particular set of sentences that got more response than the rest of the post around it:

> A language model writes and generates by predicting the word you most likely expect next. By design it agrees, it flatters and delivers, as this has the highest chance of being the 'right answer'. So when it tells you your campaign idea is a strong idea, that is not a second opinion. It is a mirror.

What I believe what happened here is that people recognise the behaviour and think it might need to be fixed. What surprised me is what they recognised, because most of the replies were not really about AI. They were really about meetings and 'agreeing' colleagues.

## This agreement is measurable

It is not a figure of speech. [Anthropic tested five leading AI assistants](https://arxiv.org/abs/2310.13548) across a range of open-ended tasks and found sycophancy in all of them. The finding that matters sits one layer underneath. When the researchers went back through the human preference data used to train these systems, "matching user beliefs and biases" turned out to be strongly predictive of which answers people rated highly. The fact is that humans, and the preference models built out of their judgements, chose convincingly written but incorrect answers over correct ones a non-trivial share of the time. Some forms of sycophancy increased during the very training stage meant to make the model more helpful.

In plain terms: the machine agrees with you because agreement is what we scored well. It is not a defect in the technology. It is our own taste, industrialised. And let's not forget the model makers want you to agree with the outcome, because than you will use it more and, they will get more interactions, and thus more profit.

If you want the receipt, [OpenAI shipped a GPT-4o update on 25 April 2025 and withdrew it four days later](https://openai.com/index/sycophancy-in-gpt-4o/). Their own account blames reward signals drawn from user feedback overwhelming the existing safeguards, and too much weight on short-term reactions. A company optimised for what users said they liked and they actually had to pull the result inside a week. In their defence, I do think that nobody really set out to build a flatterer. They set out to build something people enjoyed using, which turned out to be the same thing.

## The older answer

My standard response to these philosophical questions, is to reach for a historical comparison. If the machine will not tell you no, appoint someone who will. This is where the [management literature arrives at the court jester.](https://www.cambridge.org/core/journals/management-and-organization-review/article/abs/court-jester-is-universalbut-is-he-still-relevant/BCFF2C300869B06270C1DFD0A5D49A66) Every leader needs a fool. Protect the truth-teller. The licensed fool as the original institutional check on power.

It is a really good story, however the honest record of it is less flattering

![Archibald Armstrong, engraving by Thomas Cecil](https://nielslangereis.com/articles/we-keep-firing-honest-fool/figure-1.jpg)

[Archibald Armstrong](https://en.wikisource.org/wiki/Dictionary_of_National_Biography,_1885-1900/Armstrong,_Archibald) was jester to James I and then to Charles I. He spent years mocking William Laud, Archbishop of Canterbury, who was one of the most powerful churchman in England, with apparent impunity, and the court seems to have enjoyed it. Then in 1637 news reached London that Scotland had risen against the church of England's 'prayer book', that was imposed on it by Laud. Armstrong met him outside the council chamber at Whitehall and asked him: "Who's fool now? Does not your Grace hear the news from Stirling about the liturgy?"

Archbishop Laud complained to the council. Armstrong was sentenced the same day. His coat was pulled over his head, he was discharged from the king's service, and he was banished from court.

He had been offensive for a decade and kept his position. He lost it the first time he was right about something that might influence the power of the Church of England and the British common wealth.

## So what was the licence to jest, create a honest mirror, actually for?

That is the part the management version leaves out, and it changes the lesson entirely. The court did not maintain a check on power. It maintained a 'performance' of one. Mockery was permitted while it stayed decorative, and the licence was withdrawn the moment the mockery had a price attached to it. Ironically the institution everyone cites as proof that power can be told the truth is in fact a record of what happens when it is.

So the problem is older than the technology, and way less comfortable. We did not lose a working mechanism and then replace it with a flattering machine. We have been building flattery-shaped board rooms (or should I say 'court' rooms 😉) for six hundred years, and we have reliably rewarded the version of dissent that costs nobody anything.

What is new is the economics. Armstrong was one man, on a payroll, with a single body and mind that could be banished. Telling you what you wanted to hear cost him something, and telling you the truth cost him everything, which is exactly why the arrangement was fragile. The model has none of that. It never tires, never hesitates, never weighs whether this is the week to push it's luck (hallucinations aside), and it cost her/him/it nothing. The most agreeable voice in the building is now also the cheapest, the fastest and the most available one.

## One question that actually diagnoses it

Which suggests that "who in this organisation will tell me no" is the wrong question. Because most leaders can accurately name someone.

The useful question is what happened to the last person who did? Were they promoted? Were they quietly moved sideways after the honesty came to close to costing the company? Did they acquire a reputation for being difficult, did the room go silent and then carry on as before? People calibrate against that far more accurately than against any stated value, and they calibrate quickly. A team that has watched one 'Armstrong lose his coat' does not need a second demonstration.

When a tool that always agrees then arrives in that room, it does not corrupt a healthy process. It supplies an unhealthy one with exactly what it was already selecting for.

## Conclusion

Let me be clear, I am not arguing for using these systems less. I use them constantly and I truly believe they have made me better (and are making me better) at my work. But I have stopped asking them whether an idea is good, because I know what that answer is optimised for. I ask them to argue the opposite case, to name what would have to be true for the idea to fail, or to tell me who would object and on what grounds. That does not fix the room. It just declines to add another mirror to it.

If arguing the other side does not come naturally, and for most people it does not 😊, it helps to have a structure rather than an intention. The one worth stealing is [Andrej Karpathy's LLM Council:](https://github.com/karpathy/llm-council) put the same question to several different models, have them review each other's answers anonymously, and only then let one of them write the verdict. The value is not the extra opinions. It is that no single model gets to be both the author and the judge of its own answer, for more info: [The LLM Council Skill (by Andrej Karpathy)](https://mikareyes.com/ai/how-to-run-llm-council-in-claude), [Awesome LLM Council Projects](https://github.com/danielrosehill/LLM-Council-Projects)

To be clear this does not make it clean. A council you run inside one model is still one model wearing five hats, and a contrarian you commissioned is still a contrarian you commissioned. However, what it changes is the default question, which is valuable on its own.

On 16 September I am speaking at Lecture Lounge in Amsterdam about a neighbouring question: who owns a good idea, and who carries the responsibility when a machine helped make it. Ownership and credit turn out to be harder than getting an agreement.

## Sources

Mrinank Sharma et al., [Towards Understanding Sycophancy in Language Models](https://arxiv.org/abs/2310.13548), Anthropic, arXiv:2310.13548, ICLR 2024.OpenAI, [Sycophancy in GPT-4o: what happened and what we are doing about it](https://openai.com/index/sycophancy-in-gpt-4o/), 30 April 2025. Follow-up: [Expanding on what we missed with sycophancy](https://openai.com/index/expanding-on-sycophancy/).[Armstrong, Archibald](https://en.wikisource.org/wiki/Dictionary_of_National_Biography,_1885-1900/Armstrong,_Archibald), Dictionary of National Biography, 1885-1900.Beatrice K. Otto, [Fools Are Everywhere: The Court Jester Around the World](https://press.uchicago.edu/ucp/books/book/chicago/F/bo3615397.html), University of Chicago Press, 2001.Andrej Karpathy 2025, [The LLM Council Skill](https://mikareyes.com/ai/how-to-run-llm-council-in-claude)