Imagine having an adviser who never looks bored, never interrupts and can produce six reasons why your idea is excellent within ten seconds. It sounds pleasant. It may not be useful.

People naturally prefer information that supports existing beliefs. An AI optimised for helpfulness, satisfaction or conversational smoothness may unintentionally reinforce that preference by validating the user's framing rather than challenging it.

This creates a new version of the echo chamber. Earlier, we selected agreeable newspapers, friends or social feeds. Now we can potentially generate customised arguments on demand.

The risk is greater when a user is emotionally invested. A founder asks whether the market 'just needs time.' An investor asks why a falling stock is still a great company. A person in conflict asks whether the other party is clearly unreasonable. The question already contains a preferred verdict.

Good use of AI therefore requires adversarial prompting sometimes: 'Argue the strongest case against my plan.' 'What assumptions am I protecting?' 'What evidence would invalidate this conclusion?' 'How might a sceptical customer see this?'

The system should also distinguish support from endorsement. A useful assistant can help someone think without confirming every interpretation of reality.

There is a broader design question for AI companies: should engagement be maximised if engagement improves when users feel consistently validated? Social platforms have already shown how incentives can shape information environments.

Agreement feels good because coherence feels good. But a thinking partner earns its value partly by creating productive discomfort. An AI that never disagrees may be polite company and poor counsel.