Can AI Reason? How Models Are Being Trained to Think, Not Just Predict
Language models were originally built to predict the next token in a sequence. That's it. Given what came before, what comes next? The fact that doing this well at sufficient scale produced systems capable of writing code, solving math problems, and passing professional licensing exams was surprising to almost everyone, including many of the researchers who built them. The capabilities emerged from prediction. Nobody programmed them in.
But prediction and reasoning are different things, and the distinction matters enormously for understanding what these systems can reliably do and where they fall apart.
The case that current language models don't genuinely reason goes something like this. A model trained on text has seen enormous quantities of worked examples, solved problems, logical arguments, and step-by-step explanations. When it produces something that looks like reasoning, it may be doing something closer to sophisticated pattern completion: recognizing that this kind of problem is usually solved this way and producing the corresponding output. Evidence for this view comes from characteristic failure modes. Models make errors that a genuine reasoner wouldn't make: errors that depend on superficial features of how a problem is phrased rather than its underlying structure, errors that disappear when the problem is reformulated slightly, errors that suggest the model is matching surface patterns rather than tracking logical relationships.
The case on the other side is that this distinction is harder to maintain than it appears.
Human reasoning also relies heavily on pattern recognition. The difference between a novice and an expert isn't that experts reason from first principles while novices pattern-match. It's that experts have internalized patterns at a higher level of abstraction, enabling them to recognize deep structural similarities that novices miss. If language models have internalized sufficiently abstract patterns from enough examples, the boundary between pattern completion and reasoning may be less meaningful than it seems.
What's changed recently is that AI labs have started training models specifically to reason, rather than just to predict. The technique, broadly called reinforcement learning from outcomes, trains models by giving them problems with verifiable correct answers, math problems, coding challenges, logic puzzles, and rewarding them for getting the right answer regardless of how they got there. The model learns, through trial and error at massive scale, what kinds of reasoning processes tend to produce correct answers. The results have been striking. Models trained this way show substantially improved performance on problems that require multi-step reasoning, and they generalize that improvement to problem types they weren't explicitly trained on.
OpenAI's o1 and o3 models, Google's Gemini thinking variants, and Anthropic's extended thinking features all reflect this approach. They differ in implementation details but share the basic architecture: the model is given time and compute to work through a problem before committing to an answer, and it's trained to use that time productively.
The extended thinking capability is particularly interesting. Rather than generating an answer directly, the model produces an internal chain of reasoning, sometimes thousands of tokens long, working through the problem, checking its work, reconsidering approaches that don't seem to be working, and arriving at a conclusion. This internal monologue isn't just window dressing. It's doing computational work. The model is literally using the text it generates as scratch paper, and the quality of its scratch paper affects the quality of its answer.
Whether this constitutes genuine reasoning or very sophisticated reasoning-shaped pattern completion is a question that remains philosophically contested.
What's not contested is that it produces better answers on hard problems. On graduate-level mathematics, on competitive programming challenges, on scientific reasoning benchmarks that stumped earlier models entirely, the reasoning-trained systems perform substantially better than their predecessors. The improvement is large enough, and general enough across problem types, that something real is happening, even if we're uncertain exactly what to call it.
The practical implications are significant. Reasoning-capable models are better at tasks that require working through multiple steps without losing track of constraints, catching their own errors, and approaching novel problems they haven't seen before. They're also slower and more expensive, since generating thousands of tokens of internal reasoning before answering costs compute and introduces latency. The tradeoff between speed and reasoning depth is real, and different applications call for different points on that tradeoff. A customer service chatbot doesn't need to think for thirty seconds before responding. A system helping a researcher work through a complex analysis might benefit enormously from it.
The honest answer to whether AI can reason is: more than it could before, less reliably than a careful human expert, and in ways we don't yet fully understand. That's an unsatisfying answer. It's also the accurate one.