Interactive · concept
Why AI makes things up
When AI invents a fact, people call it a “hallucination.” But it isn't a glitch. It's the same thing the model always does, just with nothing real to lean on. Watch it happen.
Ask it something it can't possibly know
A model doesn't look facts up in a drawer. Every answer is a prediction of what should come next, word by word. When it knows the topic, that prediction lands on the truth. When it doesn't, it still predicts, fluently and confidently and wrongly. Fluency was never proof of fact.
This is also why it pays to understand the prediction step. Once you've seen how a model picks its next word, hallucinations stop being spooky and start being obvious.
The book behind this newsletter
Just Predicting Words
How ChatGPT, Claude, and Modern AI Actually Work
The trick is small. The world it built is not.