We use these tools regularly. They help us work through large amounts of information, interrogate ideas, compare options and get past the tyranny of the blank page. Few people want to return to manually summarising a 70-page report at 16:45 on a Friday. Still, there is something slightly suspicious about a colleague who finds every idea interesting. LLMs are remarkably supportive colleagues. Share a half-developed idea and, within seconds, they can explain why it has potential. Give them three data points and an optimistic prompt, and they may start reaching for the word “breakthrough.”

The answer often starts with the question

An LLM responds to the direction we give it. If a question assumes that a hypothesis is promising, the model will often explore that possibility. If we ask why a pharmaceutical company would buy our service, it can produce a respectable list of reasons. It may do so without knowing whether any pharmaceutical company has shown the slightest interest.

The response can sound analytical because it is neatly structured and uses the right terminology. That presentation makes it easy to mistake plausible wording for an independent assessment.

I have caught myself doing this. You ask a question while quietly hoping for a particular answer. The model provides it in confident prose. Suddenly, an assumption feels a little more established than it did five minutes earlier.

Nothing has changed, of course. No experiment was repeated. No customer signed a contract. No investor transferred money. The idea simply came back wearing a tie.

Agreement is easy to manufacture

This becomes risky when teams use AI to judge their own work.

Researchers can receive an encouraging interpretation of preliminary results. Commercial teams can generate impressive reasons why a market must exist. Founders can ask whether their platform is differentiated and receive several paragraphs confirming that it is.

The model may also produce caveats, but even those can sound oddly reassuring. Weak evidence becomes “an early signal.” Missing validation becomes “an opportunity for further research.” A product nobody has bought is still “well positioned to address an unmet need.”

That language is comforting. Comfort has never been a reliable method for assessing a scientific or commercial claim.

Give the model permission to be difficult

The quality of the exchange improves when we stop treating the LLM as a source of approval.

Instead of asking whether the evidence supports our preferred conclusion, we can ask where the conclusion is weakest. We can request competing explanations, missing information or reasons a sceptical reviewer might reject the argument.

For commercial questions, the model can take the position of a procurement lead who has seen ten similar propositions this month. It can examine why the supposed differentiator may be irrelevant to a buyer. It can separate statements supported by customer interviews from claims that came from an internal workshop and gradually acquired the status of fact.

The most useful question is often simple:

“What evidence would make this conclusion wrong?”

That question changes the tone of the discussion. It forces us to define what we know, what we suspect and what we would need to test next.

An LLM cannot create objectivity on demand. It still works from the information and framing it receives. Used carefully, though, it can help us find the loose parts in our reasoning before someone else does.

Science already has a system for dealing with confident claims. It involves evidence, repetition and people who seem professionally incapable of saying, “Looks good to me.”

AI can assist with that work. Just do not confuse its enthusiasm with an additional sample size.