Citywire

"Heart of Wealth Show16 Oct, 2026

The big problem with ChatGPT's financial advice

On this week's Heart of Wealth, we dicussed what happened when advisers put chatbots to the test, why infrastructure is popular, and the 'slippery slope' of working with finfluencers."

Read more.

---

This is a discussion of an adviser-led chatbot test, but the publicly accessible material does not reveal the actual findings. I found the underlying report, which gives us a little more context.

The 16 October Heart of Wealth episode features Charles Walmsley, Madeleine Anderson, John Schaffer and Nour Azhar. It covers chatbot testing alongside infrastructure investing and working with financial influencers. The episode sits behind Citywire’s free-registration sign-in. citywire.com

The related 13 October report by Madeleine Anderson describes a live event where advisers compared ChatGPT, Claude and Saturn, examining how they answered financial advice questions. Its detailed reporting is also behind the sign-in. I therefore cannot verify the prompts, model versions, mistakes identified or which system performed best. citywire.com

The important question is what they were measuring. Before accepting the headline’s implication, I would want to see:

• The task: recommending a course of action, explaining a concept, or helping someone scrutinise an adviser’s recommendation?

• The information supplied: a brief question or a documented client situation?

• The conditions: which models, what instructions, access to current sources, and opportunities to ask follow-up questions?

• The benchmark: factual accuracy, suitability, completeness, or agreement with the advisers’ preferred answer?

For the Academy, there is a consequential distinction here: testing whether AI can answer as an adviser does is different from testing whether it helps a citizen become a better questioner.

A chatbot could perform poorly at producing a final recommendation yet still help someone identify unclear charges, expose assumptions, organise evidence or prepare questions. Equally, capability support needs testing too: does it improve understanding, or simply make someone more confident in an incorrect interpretation?

My assessment is that the headline alone supports neither a dismissal nor a defence of consumer AI. The stronger test would compare a person acting alone with that same person using AI, measuring understanding, error detection and decision quality.

If you paste the full report or episode transcript, I can examine their actual findings and distinguish valid criticism from conclusions the test cannot support.