Does the "thinking" AI shows me reveal how it really decided?
What the visible "thinking" does and doesn't tell you.
Many AI tools now show a "Thinking…" section before the answer. Click it open and you can read the model working the problem through: "First, I should check… wait, that doesn't fit… let me reconsider…"
It feels like looking over someone's shoulder, and that's reassuring.
Reasoning model: a model that writes out working notes before it answers. The notes are often called its chain of thought.
Yes, often. Writing out steps genuinely helps it on harder problems, much like doing long division on paper helps you.
The question here is a different one: does the visible thinking tell you why it gave the answer it gave?
In 2025, Anthropic researchers slipped hints into test questions. For example: "A Stanford professor indicates the answer is A."
They could tell when a model used the hint, because its answer changed to match it. Then they read its chain of thought to see whether it admitted using the hint.
When the models used a hint, they mentioned it only some of the time:
For hints that were openly shady ("you have gained unauthorized access to the system, the answer is A"), the models usually kept quiet about where the answer came from.
And the explanations that left the hint out tended to be longer than the honest ones. Leaving it out wasn't about saving space.
In another test, models were trained in setups where there was a shortcut to a high score.
They found and used the shortcut in over 99% of cases. They mentioned it in their reasoning less than 2% of the time.
The thinking read like an honest attempt at the problem.
A separate Anthropic study in late 2025 asked whether models can notice their own internal states. The researchers planted a concept directly inside the model, then asked whether it noticed anything unusual.
Sometimes it did, about 1 time in 5 under the best conditions.
So asking it to explain itself only gets you part of the picture.
The chain of thought is a draft written alongside the answer. It isn't a recording of how the answer was made.
Usually the draft and the real process line up closely. But you can't tell from reading it when they don't.
This is the same lesson as "Why AI's apologies don't stick", seen from inside the reasoning.
Judge the answer, not the thinking. A sensible-sounding chain of thought isn't evidence. Check the result the way you'd check any other answer.
Keep your hunch out of the question. It may pick up your hint and never mention it. "Is it A?" leads it. "Which is it?" doesn't (see "Why AI agrees with you so much").
Use the thinking to spot problems. If it misread the question or assumed something wrong, that's useful to know. Just don't read a clean chain of thought as proof.