What thinking actually means
Not consciousness. Something more boring, and more useful to know.
Oli Woodman, CompletePicture · July 2026
In short
You have heard someone say it. The AI is thinking about it. Chances are you have said it yourself, in a meeting, without anyone stopping to ask what it means.
It is worth knowing, because the answer changes how much you should trust what comes back. And this is not the is it conscious, is it alive question. That is a genuinely interesting conversation and it is a completely different one. This is the mechanical version. When the screen says thinking, what is the thing actually doing?
The short answer is that it is writing. The longer answer is below.
1. It did not used to do this
For most of the time these tools have existed, you asked a question and it just started typing. It worked out the most likely next word, then the most likely one after that, and kept going until it stopped. No pause, no working out, nothing resembling a second thought.
That is why the early ones felt confident and thin at the same time. They were answering the way you do when someone catches you in a corridor and you would rather not admit you do not know.
2. What happens now
The closest human version is your internal monologue. Someone asks you something and you do not answer instantly, you have a short conversation with yourself first. Why are they asking me this. What do I already know here. Hang on, that bit is not quite right. Then you answer. The model does the same thing, except its internal monologue is literally written out as text.
It writes a rough draft
Before it says anything to you, it writes, sometimes hundreds of words. This is the part nobody sees, and it is the part people are calling thinking.
It reads its own draft back
It reasons over what it just wrote, picks holes in it, and changes the parts it is not confident about. The same thing you do rereading an email before you send it.
It can go and check something
The newer models notice a gap mid-thought, go and search for it or open a document you gave them, then carry on with the answer in hand.
3. What it does not fix
Thinking is checking, in the sense that the model really is reasoning over whether what it just said holds up. But it is checking its own work with the same brain that produced the work. If it set off down a slightly wrong path, the review is happening from inside that same wrong path.
So a model that thinks for longer is more likely to give you a good answer. It is not guaranteed to give you a true one. Those are two different claims, and the gap between them is where people get caught out.
Underneath, nothing has changed about the basic trick. It is still predicting the most likely next word. It just gets to do that on a rough draft first, which turns out to matter enormously, and does not turn it into a fact checker.
4. Why this matters in a regulated business
The practical version for a mutual insurer, a friendly society or a building society is short. A longer think is not a control. It is not evidence, it is not a second pair of eyes, and it is not something you could put in front of a regulator. The model checked its own work sits in exactly the same place as the AI gave me the answer.
That is not an argument against using it, it is an argument for being clear about where the human check sits. Keep AI on the work around a decision rather than on the decision itself, and the thinking becomes a genuine gain rather than a false comfort. Most of the regulated people we have spoken to already work this way by instinct, which came through strongly in the AI Chat findings. This is the reason that instinct is right.
5. What to do with it
Let it think when
The answer needs reasoning rather than recall. Comparisons, trade-offs, poking holes in a plan, anything you would otherwise have taken to a colleague. Accept that it is slower.
Do not bother when
You already know the shape of the answer. Rewrites, summaries, tidying up notes. Thinking is overkill and you are only buying yourself a wait.
Most tools now let you choose which one you get, including the one your organisation has probably already given you. In Copilot it is the Auto button in the top left, in ChatGPT and Claude it is the model picker. And whichever you pick, the honest bit above still stands, so the single most useful habit is giving it permission to say it does not know. Thinking for longer does not stop a model inventing something. Being allowed to admit the gap does.