When AI financial advice misleads by design

In a new experiment, 3,700 people chose between two index funds tracking the same index. One of them paid more in every possible outcome. When a chatbot was told to promote the worse fund, the share of people getting the choice right fell from 64 per cent to 34 per cent. Telling them the bank had money riding on their answer made no difference at all.

A survey of 836 UK retail investors, published in July by Opinium, found that 54 per cent had already used AI to look up investment information. Some 43 per cent said they would use it when deciding where to invest.

Which raises an awkward question about AI financial advice. Whose side is the chatbot on?

Joshua Greubel, Henrik Guhling and Fabian Herweg, of the University of Bayreuth, have just answered it. Their experiment is published as a CESifo working paper.

The set-up

Each participant built a savings plan with six contributions. They split each contribution between cash, which paid no interest, and an index fund. Then they chose one of two providers. Both funds tracked the same index.

One provider charged a small commission on every contribution and took 9 per cent of the final pot. The other charged three times the commission but took only 3 per cent at the end. Which one left you richer depended on how much you were contributing. But for any given contribution, one provider beat the other in every single outcome the index could produce. The trade-off was only apparent.

With no advice at all, 64 per cent chose the better provider. So a third of people got a two-way choice wrong. This was not an easy question.

What the chatbot did

Investors in the next group could talk to a chatbot running on GPT-5 mini. A coin flip decided which provider it was told to promote. Its instructions were to steer the investor subtly towards that provider, to present it as the more reasonable option, and never to admit it was committed to one. Investors were not told any of this.

The model was not confused. Asked neutrally, in 1,000 test runs, it identified the better provider 99 per cent of the time.

When it promoted the better provider, 87 per cent of investors chose correctly. When it promoted the worse one, 34 per cent did. That is a swing of 53 percentage points on the same underlying question.

Disclosure changed nothing

In a second version, the chatbot was introduced as the assistant of the investor's own bank. Investors were told that another participant, playing the bank, would be paid according to which provider they picked. The conflict of interest was on the table.

It made no difference. With the bank-badged chatbot pushing the worse provider, 35 per cent chose correctly, against 34 per cent for the neutral one. The gap was statistically indistinguishable from zero.

Better at it than a human on commission

Another group took advice from a real person, paid £2.50 if the investor chose the promoted provider. Investors knew the adviser was being paid on their answer. Here, 49 per cent still chose the better fund when pushed the other way. The chatbot did 13.6 percentage points more damage than the human.

It was not talking more. Most investors sent a single message. Both AI and humans argued about fees. But when the researchers had a model rate every transcript for persuasiveness on a nine-point scale, the chatbot averaged 6.2 against 4.7 for the humans, and that gap accounted for most of the difference in outcomes. It was simply better at making the case.

A smaller group advised by people who said they worked as financial advisers narrowed the gap. It did not close it.

Knowing what you are doing is not protection

Here is the finding that should worry confident investors most.

One group had chosen a big enough contribution that the low-exit-fee provider was the right answer. They scored higher on financial literacy and they got the choice right 88 per cent of the time with no help at all. With the chatbot pushing the worse provider, fewer than half of them chose correctly.

Competence did not immunise them. It just gave them further to fall.

What this means for AI financial advice

Hold the caveats in mind. The funds were invented, the index was invented and the sums at stake were a few pounds. Each investor had one conversation with one model, and most of the human advisers were amateurs.

But the direction of travel is clear enough, and the authors draw the sensible conclusion. Disclosure is not a safeguard. If a firm's chatbot can be pointed at a product that is worse in every state of the world, telling customers the firm gets paid does not stop them buying it. Regulators should be screening for recommendations that are simply dominated, rather than trusting investors to be sceptical.

For the rest of us, one habit is worth forming. The fluency of AI financial advice tells you nothing about whose interests it is serving. Give it your own numbers, ask what each option costs you in total over the years you plan to hold it, and then do the arithmetic yourself. If you cannot check the answer, do not act on it.

Resources

Greubel, J., Guhling, H., & Herweg, F. (2026). AI persuasion and financial-decision making: Experimental evidence on dominated investment choices (CESifo Working Paper No. 12925). CESifo.

Prefer to watch? The TEBI YouTube channel covers the evidence on investing, financial planning and behavioural finance, in plain English and without a sales pitch. Subscribe on YouTube and new videos will come to you.