Tim Harford posed a more interesting question than most AI commentary this week: not whether chatbots hallucinate, but why we believe them when they do.

His answer draws on cognitive science. Fluent, confident language is a powerful social signal of competence. We evolved to treat it as a reliable proxy for expertise. AI systems, meanwhile, are explicitly optimised to produce fluent, confident language. The result is that the cues we normally use to calibrate trust can become systematically misleading.

This is not simply a user-education problem. Almost 1 billion people worldwide use AI tools weekly. You cannot realistically train users at that scale to maintain the right level of scepticism for every AI-generated response. It is fundamentally a design and deployment problem.

Organisations deploying AI in client-facing or decision-critical contexts have a responsibility to build verification into the workflow - not as a compliance exercise, but as a structural feature of how work gets done.

A related Harvard Business Review article this month explored the psychological costs of AI adoption. One finding stood out: employees using AI in high-stakes settings often report elevated anxiety because they struggle to judge when the output can be trusted. AI removes one form of cognitive load while introducing another.

The fluency problem is likely to intensify as models improve. Better outputs will make errors harder to detect, not easier.

The real question is whether organisational design, governance, and decision-making processes can evolve quickly enough to keep pace.

#AI #CognitiveBias #TrustInAI

https://lnkd.in/gA3h7zp4