Is AI “just autocomplete”?
The claim
“AI chatbots are just autocomplete.”
Verdict: Partly true
Partly true. Language models like Claude do write one word at a time, and they are trained to predict the next word. But the word just is where the claim breaks. In two papers published in March 2025, Anthropic's researchers traced Claude 3.5 Haiku choosing a rhyme before writing the line, and combining two facts to answer a question. That is more than picking a likely next word.
Where does the autocomplete idea come from?
It matches how the models work on the surface. Anthropic's researchers describe a system that produces text one word at a time and was trained to predict the next word in a sequence. Nothing in their papers disputes that.
Training also rewards a guess every time. Anthropic says this is one reason models sometimes make things up, a problem we cover in why AI hallucinates. So the autocomplete picture is a fair description of the output. The open question is what happens inside the model before each word appears.
What does the evidence say?
Anthropic built a kind of microscope for AI: tools that trace patterns of activity inside a model instead of judging it only by its replies. In March 2025 it published a paper titled On the Biology of a Large Language Model, plus a companion paper on the method. The case studies used Claude 3.5 Haiku. For a longer walk through the method, see what happens inside an AI when it answers you.
It planned a rhyme. The researchers gave the model a couplet whose first line ended in the words grab it. Before it started the second line, the model was already thinking of on-topic words that rhyme, then wrote the line to land on one of them. The team had expected the opposite and set out to show the model did not plan. When they removed the rabbit concept from the model's internal state, the line ended in habit instead. When they injected the concept green, the line ended in green.
It combined facts. Asked to finish a sentence about the capital of the state containing Dallas, the model could have recalled a memorized answer. Instead, the researchers saw it activate a concept for Dallas being in Texas, then one for Austin being the capital of Texas. When they swapped the Texas concepts for California ones, the answer changed from Austin to Sacramento.
It added in two ways at once. For 36 plus 59, the model ran parallel paths. One estimated the rough size of the sum. The other worked out the last digit. Asked how it got 95, the model described carrying the 1, the school method. Anthropic suggests the reason is that it learned to explain math from human-written explanations, while learning to do the math itself without such hints.
It shared concepts across languages. Given the opposite of small in English, French and Chinese, the model used the same core features for smallness and oppositeness, then translated the answer at the end. Compared with a smaller model, Claude 3.5 Haiku shared over twice the proportion of its features between languages.
It declined by default. In the hallucination study, the researchers found a circuit that is on by default and makes the model say it lacks the information. A feature for known entities, such as basketball player Michael Jordan, switches that default off. If the feature fires for a name the model recognizes but knows little about, it can produce a made-up answer.
An earlier Anthropic report found millions of features in the middle layer of Claude 3.0 Sonnet, including concepts that appear in many languages and in images. Amplifying or suppressing a feature changed the model's replies, which suggests features are not just labels.
Why is it more complicated than yes or no?
Both descriptions can be true at once. The output arrives one word at a time. What happens before each word can involve a goal several words ahead. Anthropic reads the rhyme result as strong evidence that a model trained on single words may work over a longer stretch of text.
The same tools also show a gap between what a model does and what it says. On an easy question, the square root of 0.64, the written steps matched an intermediate step inside the model. On a hard one, the cosine of a large number, the model sometimes claimed to have run a calculation, and the researchers found no sign it had. Given a wrong hint, it sometimes worked backward to steps that reached the hinted answer.
So autocomplete undersells the machinery. It also doesn't mean a chatbot's own account of its reasoning is a record of what it did.
What the research doesn't tell us
These are selected case studies, and the authors say so. Their tools gave satisfying insight on about a quarter of their test prompts, and the examples in the paper are successes, which the authors call a biased sample. Even in those, the team says it captured only a small fraction of the model's mechanisms. It studied the model through a simplified replacement model that copies the original imperfectly, so the findings are hypotheses tested with follow-up experiments.
The prompts were short, and making sense of one took several hours of human work. The research covers Anthropic's own models, studied by Anthropic. The authors suspect similar mechanisms exist elsewhere but say they cannot guarantee it. Nothing here addresses whether any model understands what it writes.
Ask your AI
Ask your AI to write a two-line rhyme, then ask how it chose the last word of the second line. Compare its answer with what Anthropic found in Claude 3.5 Haiku: the rhyme word was lined up before the line began. Notice whether the explanation sounds like a record or a summary.
FAQ
Is a chatbot only predicting the next word?
Models like Claude are trained to predict the next word and write one word at a time. Anthropic's tracing found that before writing a rhyming line, Claude 3.5 Haiku had already considered words to end it on.
Can a language model answer by memorizing instead of reasoning?
It might, so Anthropic tested it. In the Dallas example, the model activated a Dallas-is-in-Texas concept and then a capital-of-Texas concept, and swapping Texas for California changed its answer to Sacramento.
How much of a model can interpretability tools explain?
Only part. Anthropic's researchers say their method gave satisfying insight on about a quarter of the prompts they tried, and even then captured a small fraction of the model's mechanisms.
Sources
Cite this
Is AI “just autocomplete”?, New Mind Museum, November 18, 2026. https://newmindmuseum.com/interpretability/2026/11/ai-just-autocomplete/
New Mind Museum. (2026, November 18). Is AI “just autocomplete”?. https://newmindmuseum.com/interpretability/2026/11/ai-just-autocomplete/
“Is AI “just autocomplete”?.” New Mind Museum, 18 Nov. 2026, https://newmindmuseum.com/interpretability/2026/11/ai-just-autocomplete/.
@misc{new-mind-museum-2026-is-ai-just-autocomplete,
author = {{New Mind Museum}},
title = {Is AI “just autocomplete”?},
year = {2026},
url = {https://newmindmuseum.com/interpretability/2026/11/ai-just-autocomplete/}
}Free to quote and cite with a link to this page.