Myth check: “AI is just autocomplete, a stochastic parrot”
AI is trained to predict the next word, but studies of its insides show it plans ahead and reasons in steps. Its explanations can mislead.
The claim in full “Large language models are just autocomplete: 'stochastic parrots' that stitch text together without internal structure, planning or reasoning.”
What the data shows
The claim: a chatbot only guesses the next word, like phone autocomplete on a bigger scale. It stitches text together with no plan and no reasoning, a “stochastic parrot” repeating patterns at random.
What’s true. Language models are trained to predict the next token (a word or part of a word). The 2021 “Stochastic Parrots” paper rightly argued that fluent text isn’t by itself evidence of understanding. But the training goal describes how a model learns and how it outputs text, not what happens inside it.
What studies of the inside show. Anthropic’s 2025 study of Claude 3.5 Haiku found the model picks a rhyme word (“rabbit”) before writing the line that ends with it. Suppressing that idea made it end the line with “habit” instead. Asked for “the capital of the state containing Dallas”, it worked out “Texas” internally and then “Austin”. Swapping Texas for California inside the model changed the answer to Sacramento. Concepts were shared across languages.
What still backs the critics. The same research found the model’s explanations of its own reasoning can be wrong. Asked how it did arithmetic, it described the school method of carrying digits, which isn’t what it did internally. Given a hint, it sometimes worked backwards to the hinted answer.
What it means
“Just autocomplete” describes how these models are trained, not everything going on inside them. They do some planning and step-by-step working. But they can’t reliably tell you how they reached an answer, so don’t take a chatbot’s account of its own reasoning at face value.
Caveats
- This is one company studying its own models (and this site is partly built with AI models from the same company).
- The methods capture only a fraction of what the model computes.
- Whether these mechanisms amount to “understanding” is a philosophical question the findings don’t settle.