Thinking and Reasoning in AI

What is Thinking and Reasoning in AI?

Ask an AI to directly compute “23 times 17.” It might blurt out “391”—but that’s because it memorized similar arithmetic from its training. Give it an unfamiliar logic puzzle, like “A is taller than B, B is shorter than C, D is taller than A. Who is the tallest?” If it answers on the spot, it’ll probably mess up.

But if you tell it to “write down your thinking process first”—list the conditions, derive step by step, and then conclude—its accuracy jumps dramatically. That process of “scribbling mental arithmetic on scratch paper” is what the AI field calls thinking and reasoning.

Don’t get carried away, though. AI doesn’t have “aha” moments. It’s still playing that word‑chain game underneath, but this time, it treats the solving steps as part of the chain. It feeds its own intermediate conclusions back into its context as references for the next step. It’s like how you don't write the final answer directly in a math exam—you write “shifting terms gives... combining like terms...” and inch toward the result.

The mainstream technique today is called Chain of Thought (CoT)—in plain terms, “forcing the AI to draft.” A fancier version is Tree of Thoughts, which lets the AI try several reasoning paths simultaneously; if one leads nowhere, it backtracks and tries another, like navigating a maze.

But you must stay clear‑headed: this is all simulated reasoning, not genuine understanding. AI has no clue what “taller” means. It just knows from patterns in its training data that “A > B, B > C” is usually followed by “A > C.” If you throw in a tricky twist or a common‑sense trap (like “an elephant that fits in a fridge”), its reasoning chain collapses immediately—yet it will still sound rock‑solid.