The interview was conducted by Meike Schreiber and Markus Zydra.

Ms Köhler-Geib, have you already asked the AI agent on your smartphone a question today?

This morning, not yet. I generally use AI on a very regular basis – mainly if I want to gain a better understanding of complex topics, if I need quick summaries, or, for example, if I am preparing for an interview. Then I take a look at what journalists have been writing about recently. At the Bundesbank, too, we work with our own interesting AI tool: our staff can use large language models without the input data being sent externally to the providers or being processed further. This creates trust and, at the same time, allows for a wide range of practical applications in their day-to-day work.

Are you concerned about the rapid development of artificial intelligence?

Artificial intelligence is a disruptive technology that is causing profound change to the economy and society. In Europe, we need to take greater advantage of the opportunities arising from this. This will then also allow us to better mitigate the risks. In terms of large AI models, Europe is clearly lagging behind at present – especially with regard to the United States and China. This is a structural challenge that we need to overcome. A recent example is the case surrounding the US company Anthropic and its Mythos 5 AI model: the US administration ordered access to the latest versions to be blocked for the rest of the world. This very clearly shows how strategically significant these technologies have become.

What does that mean for Europe exactly?

Models such as Mythos are powerful tools for identifying security vulnerabilities in complex IT systems. If European governments and enterprises do not have access to such capabilities, they will be competing at a clear disadvantage from the outset. As was the case with earlier groundbreaking innovations such as the spinning frame or GPS satellite navigation, if people are prohibited from using a particular technology, they usually do not just give up on that technology, but develop their own alternatives instead. This now needs to happen for AI. Figuratively speaking, the United States has a ferocious tiger in a cage, and Europe is allowed to look if it pays the price. Europe really needs to be in a position to say: “We have a tiger of our own and it is at least as ferocious as yours”. We do not have this option at the moment. It is clear that we need to advance our digital sovereignty.

How dangerous is the use of AI models in the equity markets, for example? What does it mean for the stability of the international financial markets?

That depends on a number of factors. First, a role is played by the nature of the AI model in question – for instance, whether you are working with reinforcement learning models, large language models, or other approaches. These are very different systems that can also behave differently. Second, a crucial factor is the heterogeneity of the data used to train each of these models. If all models are based on the same datasets, this quickly leads to them all moving in the same direction, so to speak: in human terms, it would be as if every equity investor had the exact same education, the exact same models, and would make the exact same decisions. This can considerably amplify herd behaviour, or even cause it to emerge in the first place.

How can herd behaviour in the equity markets be prevented?

In principle, we have been aware of phenomena such as herd behaviour in the capital markets for a long time now. Accordingly, there is also regulation aimed precisely at limiting excessively similar responses. With AI, it is taking on a new technological dimension. My impression is that both central banks and supervisory authorities are being very proactive in this regard: they are trying to obtain a picture at an early stage so that they can understand the potential risks and – where necessary – put guardrails in place. One example of this is the BIS Innovation Hub’s Project Logos: in this project, central banks can observe how AI agents process information, allocate capital, and respond to market changes within a simulated test environment.

High-frequency algorithmic trading has existed for more than 20 years now – what would actually be new or potentially dangerous if traders increasingly use AI?

It is true that algorithmic trading without artificial intelligence has been around for a long time. In the United States, around 60 % of equity trading is settled in an automated manner on the basis of mathematical models; in Europe, it is around 50 %. What is new, however, is that AI is shifting the boundary at which we humans have direct influence over trading decisions. With what is known as “agentic” AI, which is also becoming increasingly powerful, systems are making more and more decisions autonomously without human intervention in each individual case.

Do public authorities know where and which AI models are being used for equity trading?

That is precisely the challenge. For instance, the European Banking Authority is currently conducting a survey among supervised banks in order to obtain a better picture of where AI is being used and which models are being employed. This goes beyond specific trading models and is broader in scope – it is about better exploring the landscape as a whole. The risks arising from these models do not necessarily need to be greater than those that we already know from the past 15 years of algorithmic trading without AI.

What kinds of AI models can be used for equity trading?

Here, the models are reinforcement models, also known as “Q-learning” models. They do not form an “opinion” about whether markets are more likely to rise or fall. Instead, they are strongly orientated around transactions that worked in the past, i.e. how their own yields were – and how other market participants behaved. They respond to patterns from the past without contextualising them in broader lines of reasoning. This is precisely what can lead to existing trends or herd behaviour in the markets becoming more pronounced. Reinforcement learning models do not “think”; they are very mechanical and translate past yields and observed market behaviour directly into actions.

How do large language models work?

They also do not think in the way that humans do. Large language models are neural networks, i.e. a type of statistical model that is trained with large amounts of text data and that can create links autonomously. For this reason, they also always raise the question: What data actually feed into these models – and how reliable are they? Large language models do not “read” data in the traditional sense and have difficulties in directly processing structured content such as tables. This is a real challenge, particularly from a statistical perspective, because they generate answers on the basis of probabilities – i.e. they produce the result that seems most plausible.

Do you have an example of this lack of precision?

If, for example, I ask “How high is Germany’s gross domestic product?”, a model like this will typically draw on a large number of texts – from academic texts to reports – and, on this basis, generate a number that sounds likely. However, it does not specifically draw on the official, accurate tables from the Federal Statistical Office or the Deutsche Bundesbank in order to provide the exact current figure. Of course, it is also possible that the AI could spit out an official figure, but perhaps from the wrong year. Exactly this is a key limitation that we have to bear in mind when using these models.

So AI models are not necessarily the better investors?

Interestingly, similar biases to those of human investors can creep into AI models. For example, they act in an overly optimistic way or ignore risks. The fascinating thing here is that these models are often very good at identifying such biases. However, they have not yet been able to reliably address these biases. This is due, in part, to the fact that these models lack consciousness and self-awareness in the human sense. This may see further development in the coming years – but, as things stand today, it remains a key constraint for the way in which AI is used in the financial markets.

© Süddeutsche Zeitung GmbH, Munich. 
Courtesy of Süddeutsche Zeitung Content