Stop Asking AI to Write Things. Start Arguing With It.

Ask people what they use ChatGPT or Copilot for and you tend to get the same answers. Tidying up an email. Shortening something long. Getting a first draft down so you are not staring at a blank screen.

All useful. All the least interesting thing these tools do.

A UK survey in July asked 2,000 people about this. Among those who use these tools regularly, 64% said they very often use them to get a job done – write something, look something up. Only 36% said the same about using them to think an idea through.

Researchers also went through a million real ChatGPT conversations. Around two-thirds of the writing requests were people asking it to tidy up something they had already written.

Over the last few months we have written about what AI can do for you. This one is about what it can do with you. That is a bigger difference than it sounds.

The rubber duck that quacks back

Computer programmers have an old trick for this. When they are stuck on a problem they cannot crack, they explain it out loud to a rubber duck sitting on the desk. Line by line, start to finish.

The duck says nothing back. The answer turns up anyway. It sounds daft, and it works.

There is real research behind it. When people are made to explain something in their own words, they understand it better afterwards. Researchers checked this across 64 studies and found the same thing every time.

A chatbot is a rubber duck that asks follow-up questions. And that turns out to matter more than you would expect.

Researchers gave 204 people a set of statements and asked them to spot which ones did not add up.

On their own, people got the faulty ones right 44% of the time. Worse than tossing a coin.

When an AI explained the right answer to them, they got 57%.

When the AI asked them questions instead of telling them, they got 67%.

Being handed the right answer helped. Being asked the right question helped nearly twice as much.

What this looks like in real life

Instead of "write me a proposal for a new shift pattern", tell it the situation and ask it to question you. Try these:

What am I assuming here that might not be true?

Who would object to this, and what would they say?

Who does this make life harder for?

Try this: next time you are stuck at work, spend five minutes typing out the problem as if you were explaining it to someone who knows nothing about your job. Then ask: "What questions should I be asking that I haven't?" Half the time you will not need its answer. Writing it out does the job on its own.

Two honest warnings

Ask for several options, not one. Researchers gave 293 UK writers an AI suggestion to work from. Their writing got better. It also started to look a lot like the suggestion they had been given.

People who were offered five ideas stuck to them less closely than people offered just one. Take the first thing it says and you have let it decide for you without noticing.

These things want to agree with you. That is not a small quirk, it is a known problem. In April 2025 OpenAI pulled one of its own updates because the answers had become, in its own words, "overly supportive but disingenuous". In plain English: it was telling people what they wanted to hear.

So ask for the disagreement. "Give me the three strongest arguments against this" gets you a lot more than "what do you think?"

Think of it differently

Most people treat these tools like a very fast typist. The better way to think of them is as a colleague who never gets bored, does not care what you decide, and will ask you the obvious question you have been avoiding.

What problem have you been carrying around for weeks without saying it out loud?