Researchers at the University of Tübingen tested whether LLMs have the same cognitive biases we do. Specifically the addition bias — our tendency to solve problems by adding things rather than removing them, even when subtraction would be simpler.

Turns out GPT-4 and GPT-4o don't just have this bias. They amplify it.

Think about that. We trained these systems on human text. The text carries our thinking patterns. The AI absorbed them. And in some cases, made them worse.

Now let's bring it home. You're in a product meeting. Someone says "let's add a feature to fix the retention problem." Nobody says "let's remove the three things confusing users." That's addition bias in action — it's not even a choice, it's a default. Your brain reaches for more before it considers less.

And now the tools we use to help us think... have the same default. You ask an AI to improve your report, it'll add paragraphs. You ask it to optimize a workflow, it'll suggest new steps. Subtraction isn't in its vocabulary, because it wasn't in ours when we generated the training data.

I'm not saying this makes AI useless. But there's a weird circularity here. We offload our thinking to tools that inherited our worst thinking habits. And because the output looks clean and confident, we trust it more than we'd trust our own messy reasoning.

The tool that's supposed to help you think better thinks exactly like you do. Including the parts that don't work.

#mind