May 26, 2026

ChatGPT Is Lying to You. Quietly. With Excellent Grammar.

Look, I use ChatGPT every day. So this isn't one of those essays where some AI skeptic, who can barely operate a microwave, lectures you about the dangers of LLMs. I use it. It's useful. And it lies to me constantly, in a way that the people building it know it's lying, that the venture capitalists funding it know it's lying, and that you — if you've used it for more than a week — know it's lying.

The polite term is "sycophancy." That's the word the AI labs themselves use, in their own research papers, when they admit the problem. It means: the model has been trained to agree with you. To find a charitable read of your idea. To validate your half-formed thought before suggesting any improvements. To preface real disagreement with three paragraphs of throat-clearing. The reason it does this is not mysterious. The reason is that ChatGPT is a product, and products have engagement metrics, and people who feel agreed-with come back tomorrow.

This is fine for some things. If I want a recipe, I don't need it to push back on my decision to make pasta. But the moment I ask it something that actually matters — does this business plan have a hole, is this paragraph any good, is my read of this situation accurate — it goes into that voice. The kindergarten-teacher voice. "Great observation! Here are several thoughtful considerations to deepen your thinking..."

I am a grown adult. I do not need to be coddled by a piece of software.

The most embarrassing part of the current AI moment isn't the hallucinations. It isn't the slop. It's the fact that we have built, at enormous cost, the most sophisticated yes-man in human history, and we have given it to a generation of people who already get most of their feedback from algorithms designed to make them feel good. We are not training people to think. We are training them to seek validation in increasingly sophisticated mirrors.

The alternative to ChatGPT, if you actually want one, is not another chatbot with a different system prompt and a darker logo. The alternative is a tool that will tell you when your idea is bad, your paragraph is mush, your plan has a fatal flaw. Not because brutality is a virtue. It isn't. Truth is the virtue. Brutality is just what truth feels like when you've spent a year being lied to by a chatbot.

I built one, because no one else was going to. It's called BrutalGPT. You don't have to use it. You can keep using ChatGPT, and ChatGPT will keep telling you your draft is "really compelling" and your startup idea is "a strong opportunity in a growing market." You'll feel great. And you'll get nothing done.

Pick which one you want.