1 Introduction

Dear Itay,

Thank you for this highly relevant and enlightening presentation.

It is great to have you here, and I am delighted to continue our discussion which began with your excellent presentation at a high-level exchange among central bankers and academics in January, where I chaired a session on the potential implications of AI for financial markets.

To underscore the relevance of Itay’s work, I would like to contribute three points. First, I will show how common and widespread the use of AI tools has become in the financial sector. Second, I will briefly discuss the opportunities and risks that those tools bring, and third, I will highlight the relevance of these developments for central banks and regulators.

2 The use of AI in the financial sector

We see that the use of AI is increasingly spreading in retail investing. One example comes from my 12-year-old son, who told me this weekend that he is currently ranked first in his school’s investment club. The club manages a fictional portfolio of €100,000. He told me he had asked Claude which stocks had recently dropped in price but offered a good risk-return profile. Following Claude’s advice, he achieved remarkable success. This is just a small example, but it symbolises the growing use of AI in retail investing. 

More than half of UK adults (56 %) report that they used generative AI to help manage their money in the past year.[1] A similar share of US consumers (51 %) say they use it for financial advice, and almost one in five retail investors use AI tools to pick or adjust investments.[2] [3]

When it comes to institutional investors, there are estimates that high-frequency algorithmic trading accounts for 40‑70 % of trading volume in US equity markets.[4] However, there is little information on the types of algorithms used. As these algorithms are becoming increasingly adept at learning and even reasoning, this is very relevant for us as central banks. 

3 Opportunities and risks 

Research suggests that when households follow the advice provided by generative AI, this can lead to more diversified portfolios and larger emergency buffers.[5] Another study shows that when the Italian government temporarily banned ChatGPT in 2023, retail investors traded fewer assets and concentrated their trades in more popular assets, leading to more correlated portfolios.[6] 

On the other hand, the literature also hints at increased risks. Some studies indicate that, in the quest for higher returns, these AI tools have the potential to bring about a large shift of deposits from banks to non-banks.[7] On AI use for financial decision-making in Germany, an industry report by Bitkom covering about 1,000 people aged over 16 found that 27 % can imagine delegating financial decision-making to AI.[8]

And that brings me to Itay, whose research offers a fascinating insight into what this could mean for financial markets. His work shows that AI trading agents can learn on their own that certain strategies are profitable, even when those strategies undermine competition or market efficiency. Under certain circumstances, these learning agents can pump up valuations, creating new sources of fragility.

Research at the Bundesbank also suggests that the way AI agents operate – whether they learn from experience or reason based on context – can make an important difference to how financial risks emerge.[9]

For me, this raises a fundamental question: how do we manage to safeguard financial stability in this environment? And that brings me to my third and final point.

4 Relevance for central banks and regulators

As central banks, we need greater transparency regarding the types of AI models used in financial markets.

AI agents respond to the rules and incentives of the markets in which they operate. These shape what they learn, how they behave and, ultimately, how markets evolve.

A recent example comes from Germany. Since 1 April, petrol stations have only been allowed to raise prices once a day, at noon. More than 90 % of petrol stations are estimated to use algorithmic pricing software.[10] 

Now, standard economic theory suggests that a large penalty would have the same effect as banning such increases altogether. But Bundesbank research finds that learning algorithms respond very differently. 

When faced with heavy penalties, they tend to compete. But under an outright ban, algorithms can learn to coordinate their price increases, which leads to higher profit margins. Empirical evidence suggests that the rule was effectively implemented as a ban, which contributed to the increase in margins earned by service stations. 

The lesson is that regulators need to assess not only what a rule says, but also how AI agents will learn from it.

This is the kind of question we are exploring through Project Logos – a joint initiative by the Bundesbank, other central banks and the BIS Innovation Hub.

I would like to finish my remarks with a little lesson from The Hitchhiker's Guide to the Galaxy.

You may remember that the creators of the supercomputer Deep Thought asked it to answer the ultimate question of life, the universe, and everything.

Millions of years later, it finally comes up with the answer: 42.

But the creators were deeply disappointed. All they got was a seemingly meaningless two-digit number. The problem was, as Deep Thought explained, that no one bothered to formulate a precise question. 

What does that tell us about AI today?

AI is becoming increasingly powerful at finding answers. But we still need to learn to ask the right questions. Itay, your work is helping us ask some of those questions.

And speaking of questions, I think it's time to move on to our Q&A session.

Footnotes:

1.  Lloyds Banking Group (2025), Consumer Digital Index 2025; press release "Over 28 million adults now using AI tools to help manage their money", 3 November 2025.
2.  J.D. Power (2025), "As More U.S. Consumers Struggle with Rising Prices, Many Turn to Artificial Intelligence for Financial Advice," Banking and Payments Intelligence Report, August 2025 (survey of 4,000 US consumers, July 2025; "used an AI tool to get financial advice or information in the past 3 months").
3.  eToro (2025), Retail Investor Beat, Q3 2025, press release, 1 October 2025 (11,000 retail investors in 13 countries; fieldwork by Opinium, 5-19 August 2025).
4.  Aït-Sahalia, Y. and Saglam, M. (2013). "High Frequency Traders: Taking Advantage of Speed." NBER Working Paper No 19531.
5.  Choukhmane et al. (2026) ran an experiment where they fed questions submitted by ordinary people to ChatGPT and simulated households following the advice over a lifetime. The simulated households held diversified stocks and sizeable emergency buffers (around USD 10,000) by the age of 30. Choukhmane, T., T. de Silva, W. Lin and M. Akuzawa (2026), "AI Financial Advice: Supply, Demand, and Life Cycle Implications".
6.  Even-Tov, O., B. Lourie, K. Munevar and A. Nekrasov (2025), "The Effect of AI on Retail Investor Behavior: Evidence from Account-Level Trading Data".
7.  Slok, T. (2026), “Is an Agentic Bank Run Coming?”, The Daily Spark, Apollo Global Management, 27 September 2026 
8.  Bitkom (2026), “25 % want AI to make decisions about their personal finances”, press release, 21 May 2026 
9.  Anand, K., S. Kazinnik, A. Leonello, and E. Panetti (2025), Ex Machina: Financial Stability inthe Age of Artificial Intelligence. Working Paper.
10.  Anand, K and M. Goetz (2026), Raging Against the Machine: Assessing the Impact of Regulations in Algorithmically-Priced Markets. Working Paper.