An AI that shows its working

The default AI chat interface is a text box and an answer. The answer is fluent and confident, and you have no way to tell a correct one from a wrong one except by already knowing.

Fluency is not accuracy

These systems are exactly as fluent when they are wrong. There is no tell — no hedging, no degradation in the prose. That is the real problem with taking an answer at face value, and it does not improve by making the model bigger.

So the useful question is not how to be right more often. It is how to let somebody check.

Three things you can check

• Reasoning that is visible. The steps that led to an answer are shown, so a wrong turn is something you can see rather than infer.
• Searches with their sources. Where a claim comes from the live web, the links come with it and you can follow them.
• Code that was executed. The output shown is the result of running it, not a prediction of what it would print.

The third is the sharpest, because it converts a guess into a fact. A model's idea of what code prints is an opinion. Running it is not.

The cost of showing your working

It is slower and noisier. An answer with its reasoning attached takes longer to produce and longer to read, and much of the time you do not care and just want the answer.

That is a real trade and we do not pretend otherwise. But it is there when it matters, which is more than a confident paragraph offers.

A language model is exactly as fluent when it is wrong. Being able to check is the only thing that separates the two.