Does AI Really Understand You, or Just Predict Words?

You have probably heard the tidy dismissal. AI does not understand anything. It is just fancy autocomplete, a machine that predicts the next word from statistics and nothing more. It is one of the most repeated claims about large language models, and it is comforting in its simplicity. It is also only half true, and the half it gets wrong is where the interesting questions live.

Start with what is accurate. These models really are trained to predict the next piece of text, over and over, across enormous amounts of writing. The phrase "stochastic parrot," coined by the linguist Emily Bender and colleagues in a well-known 2021 paper, captured a real worry: a system can stitch together fluent language by mimicking patterns without any grip on what the words mean. Anyone who has watched a model state something false with total confidence has seen that failure up close, a point we looked at in When AI Makes Things Up, It Isn't Lying to You.

Where "just statistics" runs out of road

The trouble is that next-word prediction turns out to be a demanding task. To predict the next token in a page of reasoning, or the next line of working code, a model has to track quite a lot: what was said earlier, what follows from it, which quantity goes where. Systems trained only to predict text end up solving problems they were never shown, holding a thread across long passages, and translating between languages they were never explicitly taught to pair. "Just predicting words" describes the training goal. It does not describe everything the model had to learn to get good at it.

Recent work pushes further. In a paper posted to arXiv on August 30, researchers reported that language models appear to organize information internally into symbol-like structures, and that editing those internal structures changes the model's behavior in predictable ways. If that holds up, then what is happening inside is more organized than a pile of correlations. It is not proof of understanding in the human sense, but it is hard to square with the idea that there is nothing in there but surface statistics.

The word doing all the damage

Most of the argument comes down to one slippery word: understand. If understanding means conscious grasp, lived experience, knowing that you know, then no, there is no good evidence a model has that, and claims that it does should be treated with real caution. If understanding means building a working internal model of a problem that reliably produces correct answers, then flatly denying it looks harder to defend every year.

So the honest position sits in the awkward middle. "It truly understands, like a person" overstates the case. "It is just statistics and grasps nothing" understates it. A model can be genuinely useful, and genuinely strange, without being either a mind or a parrot. The confident one-liner in either direction is the thing to distrust. What these systems are doing is real, unfinished, and still not fully understood, including by the people who build them.