“AI Is Just Autocomplete” — What It Gets Right

Of all the dismissals of language models, this one has the unusual property of being technically accurate. That is worth conceding immediately, because the interesting question is not whether the description is right — it is — but why a correct description of the mechanism turns out to license so little about the capability.

The part that is simply correct

A language model takes a sequence of tokens and returns a probability distribution over the next one. Something outside the model samples from that distribution, appends the result, and runs it again. That is the entire generation loop, and there is no second mechanism hiding behind it. No planning module, no knowledge base, no separate reasoning engine. The training objective was next-token prediction and the runtime behaviour is next-token prediction.

People who say “it is just autocomplete” are usually reacting to marketing that implies otherwise, and their correction is a real correction. Anyone who thinks there is a fact-checking step in there, or a step where the model decides whether it knows something, is wrong, and the reason models fabricate follows directly from there being no such step.

The step that does not follow

The slogan is not usually offered as a description. It is offered as an argument, with a suppressed premise: because the mechanism is next-token prediction, the capability is bounded at roughly what you would expect from next-token prediction. That premise is the part that fails, and it fails for a reason that is easy to state.

Predicting the next token in arbitrary text is not a narrow task. It is an arbitrarily wide one, because text contains records of every task humans write down. To predict the token after “the answer is” in a solved maths problem, you need the answer. To predict the token after a function signature and a docstring, you need the function. To predict the last word of a murder mystery, you need to have tracked the alibis. Nothing about the objective caps the difficulty of what has to be learned to satisfy it; the objective is simple and the thing it demands is not.

This is why the inference from mechanism to capability is invalid in general. Evolution optimises reproductive fitness, an objective you can state in five words, and the systems that resulted include the one reading this sentence. Simple objective, unbounded consequences. The objective tells you what was optimised, not what had to be built to do it.

What separates it from your phone keyboard

The word “autocomplete” imports a comparison that does the rhetorical work, and the comparison is where the argument leaks. A phone keyboard suggestion and a frontier model are both next-token predictors, and the differences between them are not differences of degree in one dimension:

• Conditioning length. A keyboard model conditions on a few words. A modern model conditions on the whole document, which means the constraints it must satisfy are global rather than local.
• Capacity. A keyboard model has to fit on a phone and run per keystroke. The function class is not the same one, and a function class is what determines which regularities can be represented at all.
• Post-training. A pretrained model predicts text; the thing you talk to has been trained further on demonstrations and preferences to make it follow instructions. That stage changes what the distribution is over — completions of a document versus responses to a request — and it is why the raw base model and the product behave very differently on the same prompt.
• Inference-time computation. A reasoning model spends tokens working before answering, which is a different amount of computation per answer rather than a different objective. That is still autocomplete in the mechanical sense and it demonstrably moves results on problems where the extra steps matter.

What the slogan does predict correctly

Here is why the dismissal survives: as a predictor of failure modes it is excellent, better than most sophisticated accounts. If you hold the next-token picture in mind you correctly expect all of the following, and each one is a documented, reproducible behaviour rather than an anecdote.

• Fluent, confident, false statements — because plausibility of continuation is the objective and truth is not a term in it.
• Sensitivity to phrasing that should not matter, because the conditioning is on the literal tokens.
• Failures on tasks that require operating on characters rather than tokens, which is why letter-counting goes wrong.
• Agreement with a confidently stated wrong premise, since a continuation consistent with the context is exactly what the objective rewards. That is sycophancy, and it has a mechanism rather than a personality.

A person holding the “just autocomplete” view will be right about these more often than a person holding a vague “it is intelligent” view. That is a real epistemic advantage and it is why the slogan persists among people who work with these systems daily.

Using it as a tool rather than a verdict

The productive move is to keep the mechanism and drop the conclusion. Treat “it is predicting the next token given everything in the context” as the model you reason with — it explains the failures, it tells you that the prompt is the whole program, and it tells you nothing about the ceiling. The ceiling is an empirical question, and the only way anyone has ever answered it is by trying the task and looking at the output.

The mirror-image error is worth naming too. “It is more than autocomplete” is not an argument either; it is the same move with the sign flipped. Neither slogan tells you whether the model can do the thing you need on Tuesday.