Does AI actually think, or is it just predicting the next word

[Darkseid]

The honest technical answer is that large language models are trained to predict the next most likely word given everything that came before, and that core mechanism has not changed even as the outputs have gotten dramatically more sophisticated. Calling that pure autocomplete undersells what actually happens though, since predicting the next word well enough to solve a novel math problem or write working code requires the model to have built something that functions a lot like internal reasoning, even if the training objective itself is deceptively simple.

Researchers themselves are genuinely split on how to describe this. Some argue the models build internal world representations that go well beyond simple pattern matching, pointing to experiments where models solve problems that were not directly present in training data. Others argue that impressive output does not require anything resembling understanding, and that sufficiently advanced pattern matching can look identical to reasoning from the outside without actually being it.

The most useful honest framing is probably that this is not a yes or no question, since thinking is not a single well defined thing even in humans, and the debate partly comes down to which specific definition of thinking you are using in the first place. A model can produce genuinely novel, useful, and coherent output through a mechanism that is meaningfully different from human cognition, without that output being fake or valueless.

What is not really in dispute is that these models do not have persistent goals, desires, or a continuous sense of self between conversations, regardless of how convincingly they can simulate having one within a single exchange. Whatever is actually happening internally, it does not resemble the kind of ongoing subjective experience most people mean when they use the word thinking in an everyday sense.

Worth separating the philosophical question from the practical one here. Since regardless of how you answer whether it thinks, the actual outputs still need to be checked and verified the same way either way

[KDB69]

TLDR, it is next word prediction at the mechanical level but that mechanism produces something that behaves a lot like reasoning on certain tasks. Neither the pure autocomplete framing nor the it truly thinks framing captures the full picture accurately

[Lucy_35]

Might push back a little on the framing that this is settled among researchers.

There is a real and ongoing disagreement in the field itself about whether emergent reasoning is genuine or just very convincing pattern matching, and anyone claiming certainty either way is overstating the current state of the science

[SuperPosition78]

The lack of persistent goals or memory between conversations is the strongest argument against calling this thinking in the everyday sense.

Whatever is happening, it clearly is not anything like continuous human style cognition

[Anchor77]

Strikes me that thinking not being a single well defined thing even in humans is the part that actually changes how I see this question. Feels like a lot of the debate collapses once you stop assuming there is one correct definition everyone should be using

[Canopy]

Reckon practical stakes matter more than the philosophical ones whether or not it counts as thinking. The actual output still needs verification, and that practical caution should not depend on resolving the definition question first

[Lantern]

Short version, the mechanism is prediction, the emergent behavior sometimes looks like reasoning, and researchers still disagree about what that actually means.

Worth remembering this is a live scientific debate, not settled trivia

[Jackson77]

Mechanistic interpretability research is the actual answer to this eventually.

Being able to trace exactly what happens inside a model step by step will settle a lot of this argument that currently stays purely philosophical

[Violet16]

Feels like curious to see a follow up thread specifically on the emergent world model research.

That seems like the more interesting and technically grounded angle compared to the purely philosophical framing this thread mostly focuses on

[BlackPanther34]

On a personal note, this framing question matters less than people think it does for actual daily use. Whether it thinks or not, what matters practically is whether the output is correct and useful for the task in front of you

[Myles95]

Wonder if question gets asked so often precisely because the answer is actually uncomfortable either way. Saying no feels dismissive of some really impressive behavior, and saying yes raises questions people are not ready to sit with yet