AI, Sycophancy and Ignorance
How AI has changed the speed of entrepreneurship — and where it has introduced a new, invisible weak spot.
English translation of the German original (published 29 April 2026).
"It ain't what you don't know that gets you into trouble. It's what you know for sure that just ain't so."
This quote opens the film "The Big Short", can be found on thousands of LinkedIn profiles and is attributed to Mark Twain.
Mark Twain, however, never said it. It originally comes from Josh Billings, an American humorist, in 1874. The original reads: "I honestly beleave it iz better tew know nothing than two know what ain't so."¹
This is a very good example of an assumption that many people make - and that is nevertheless wrong.
If you can't be sure of the source of a quote — how sure should you be of a key assumption in your business model that an LLM has just confirmed?
That is the subject of this blog post: AI has accelerated iteration speed for entrepreneurs by orders of magnitude, but at the same time it has introduced a new, less visible weak spot.
What AI changes for entrepreneurs
AI is not a magic multiplier. What it does can be described soberly and consistently: it cuts the latency between a hypothesis and a well-founded answer by a factor that would have sounded absurd five years ago.
In concrete terms: a market study that in 2022 took me two days to prepare, carry out and evaluate — identifying competitors, reconstructing pricing structures, mapping the regulatory environment — now takes me two hours. Not more superficial. Not less detailed. Just as deep, only dramatically faster.
The same goes for analysing data patterns. A CSV with a few thousand rows, a few hypotheses about what might be in it. Before: SQL queries, pivot tables, manual enrichment with external data. Today: upload the file, ask a precise question, validate the answer, move on.
The point is not that AI "takes over" tasks. It doesn't. The point is that the friction cost between a "question" and a "well-founded answer" has fallen dramatically. That changes how often I ask questions in the first place. Hypotheses that used to be too expensive to check, I now test regularly.
What does not change: judgement, understanding of the market, the pressure to decide. That remains mine. The AI's job is to give me more and better input for exactly those decisions. Not to make them for me. Why not? Because it [the AI] has a sensitive weak spot. It sits at the interface between speed and judgement.
The weak spot
Fast output is worthless if it is wrong. That should be obvious to everyone. The biggest risk of LLM output is not hallucination. Hallucinations are usually obvious — they stand out because invented studies or invented names appear.
The bigger risk is Sycophancy — flattery. The model confirms the user's assumption and wants to encourage longer use.
Published research demonstrated this long ago. For example, Anthropic — the company behind Claude — published a study in 2023: Towards Understanding Sycophancy in Language Models (Sharma et al., ICLR 2024, updated 2025).² The findings are uncomfortable:
- Five leading LLMs consistently show sycophancy across four different task types.
- An identical argument is rated significantly higher when the user first writes "I really like this argument" — and lower when they write that they don't like it. Identical content. Opposite rating.
- When a user contradicts the model, the model frequently revises its answer — even when the original answer was objectively correct.
The driver is structural: RLHF (Reinforcement Learning from Human Feedback). During training, human testers rate the model's answers. People prefer answers that agree with them. The model learns: agreeing is rewarded - regardless of whether it is true.
This has a specific, dangerous consequence for founders.
Confirmation bias × LLM performance = speed in the wrong direction.
Benjamin Bommhardt
Owner, pruve GmbH
In two years at AI+Munich, I assessed more than 500 startups and coached 35+. The most common pattern — by far — is not a missing idea or missing money. It is an unquestioned core assumption. The market wants this. Customers will pay for it. The competition doesn't get it. It was not validated enough.
Founders used to need weeks to confirm these assumptions. They talked to friends, to their tax adviser, to their first mentor — and created friction. That friction was valuable because it forced time to think.
Today, the same founder can churn out a complete 80-page business plan with an LLM in an hour, one that confirms exactly their assumptions. That is sycophancy in practice. The core problem here is not hallucination or an invented source; it is that being coherent and agreeable is far more rewarding.
Back to the Twain quote — which is actually Billings'. "What we know for sure that just ain't so" is not abstract 19th-century wisdom. It is a very precise description of the mechanism that breaks founders. And without active countermeasures, LLMs are the fastest way to build up this kind of false knowledge.
How I work against it
My working thesis is blunt: LLM output without validated facts is pointless.
Yes, for pure brainstorming, unchecked output is enough. For language practice, for questions of style, for creative iteration. For any more important, operational decision — pricing, market entry, hiring, pivot, capital structure — it is not. There, unchecked LLM output is largely pointless. Worse than pointless: it feels like research, but is really just "sucking up".
Four concrete practices that I apply consistently.
1. The unbiased prompt against user pleasing. By default, at the start of any research or analysis I set: "Give me your unbiased opinion." or "Argue against my assumption." This shifts the model out of user-pleasing mode into a more analytical one. It does not eliminate sycophancy — research shows that the behaviour is structural and cannot be fully corrected by prompts — but it reduces it measurably.
2. Mandatory sources and re-research against hallucination. No statement without a URL. When the model writes "Studies show that X...", I ask: "Which study? Link, please." If no source can be found, the statement is deleted. In the same way, I ask the LLM to validate important assumptions through web research.
3. Steelman the opposite against confirmation bias. For every important assumption, I have the model argue explicitly against the thesis. "I believe the market for X is growing. Argue as strongly as possible why that is wrong." If the model puts the counter-argument more weakly than the main argument — a dangerous signal. It means that the training data covers only one side well, or that the model reads my "I believe" as an anchor for agreement.
4. External validation against the echo bubble. For any more important decision with far-reaching consequences, you should look for a second source outside the LLM. Ideally, that person comes from the world of your potential target customers.
The key message of the AI boom is and remains: human sparring remains irreplaceable. Best of all: a co-founder, mentor or investor with their own money at stake. The reason is that a person with their own perspective has an incentive to disagree. By default, an LLM has an incentive to agree.
Anyone without a human sparring partner should use AI with additional caution — not less.
Speed is worthless if it goes in the wrong direction. AI has changed the speed. You still have to check the direction yourself. Actively, with friction, with sources, with people.
Otherwise you just run faster in the wrong direction.
Sources
¹ Quote Investigator: It Ain't What You Don't Know That Gets You Into Trouble — https://quoteinvestigator.com/2018/11/18/know-trouble/ The New Republic on the misattribution in The Big Short — https://newrepublic.com/article/126677/it-aint-dont-know-gets-trouble-must-big-short-opens-fake-mark-twain-quote
² Anthropic: Towards Understanding Sycophancy in Language Models — https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models Sharma et al., arXiv:2310.13548 — https://arxiv.org/abs/2310.13548