Girls and Gents,
I’m very happy to open this panel on synthetic intelligence (AI). AI is a breakthrough innovation that may result in actual financial disruption. That is very true within the monetary sector, the place AI has been an essential predominant driver of transformation in recent times. The arrival of generative AI is anticipated to additional speed up this pattern, not solely by growing customers’ adoption of AI instruments, but in addition by structurally accelerating the tempo of innovation (let’s suppose, for instance, of the brand new capability to generate pc code primarily based on natural-language queries).
Nevertheless, these vital developments elevate plenty of questions, together with from the central banker and monetary supervisor perspective with which I’m taking a look at them. I want to share a few of these questions with you, earlier than expressing my views about how we must always sort out the problem.
*
1/ First, regardless of latest progress, the underlying AI expertise doesn’t but look like totally mature, significantly as regards generative AI (GenAI). There are nonetheless plenty of unanswered questions on this topic. I’ll contact on two of them.
First, the query of general-purpose AI (GPAI) fashions: how will they carry out in an entire vary of duties related to the monetary sector? This query actually arises on two ranges. Will GPAI fashions turn into the usual for all makes use of, to the detriment of specialised fashions? Will smaller, well-trained – that’s, extra specialised – fashions have the power to dwell as much as bigger, extra general-purpose fashions? These efficiency points have many potential penalties, not least by way of competitors: if massive GPAI fashions are launched in all areas, we run a excessive danger of ending up in a pure monopoly or oligopoly, including to the already largely oligopolistic nature of the cloud market.
Second, the problem of the vulnerabilities of AI methods: whereas we’re beginning to acquire a clearer image of the state of affairs, analysis on this topic is way from full. That is significantly true within the subject of cyber safety for GenAI fashions, with the latest discovery of the hazards of “oblique immediate injection”. Whereas this race between the ‘sword’ (improvement of latest assault strategies) and the ‘defend’ (improvement of efficient countermeasures) is conventional within the safety subject, our capability to adequately safe AI methods can have a significant affect on the power of various actors to make in depth use of this expertise.
*
2/ Despite the fact that AI applied sciences aren’t but totally mature, it appears to me that central banks and monetary supervisors ought to embrace them directly, for at the very least three causes.
First, to proceed to hold out our missions successfully, by doing extra and doing it higher. AI can after all assist us turn into extra environment friendly, by growing the extent of automation. However we additionally need to supply new capabilities to brokers. For instance, our LUCIA instrument, an AI-based system with the capability to investigate massive volumes of banking transactions, permits us to evaluate the efficiency and relevance of banks’ AML/CFT fashions throughout our on-site inspections
Second, to develop important experience in AI. Utilizing AI for our personal functions permits us to regularly purchase an excellent command of the expertise, and is a really efficient approach of correctly understanding its advantages and dangers. The virtues of studying by doing clarify why inside makes use of of AI are very complementary to the supervision of AI methods deployed by the monetary sector. For instance, very not too long ago, the ACPR, with the assistance of Banque de France’s innovation middle, Le Lab, organized a “Suptech Tech Dash”, a hackathon supposed to discover what generative AI can carry to the varied supervisory capabilities. In three days, this occasion revealed the potential of huge language fashions (LLM) for supervision.
Lastly, to drive the monetary ecosystem, by sending a sign to the market that it can also – or should – make the leap. For instance, the cutting-edge work being carried out on the Banque de France on post-quantum cryptography is elevating consciousness amongst non-public stakeholders about the necessity to deal with this menace.
So, whereas it’s clear to me that central banks and supervisors should seize the alternatives provided by AI, the query is: How can we do this?
*
3/ It appears to me that we should first lay down a basic precept of governance: AI have to be on the service of humanity and society, and not the opposite approach spherical. From this angle, even when it doesn’t clear up all the issues, the continued adoption of the European AI Act, the world’s first binding textual content laying down the rules of “reliable AI”, is a welcome step. Specifically, this article will enhance client confidence, whereas offering authorized certainty for financial operators.
This governance precept may be supplemented by three operational rules.
First, utilizing AI proportionately and progressively. With a easy rule: the extra important the use case for our activty, the extra now we have to do it ourselves. For establishments like ours, this goes to the basic difficulty of information: among the information held by central banks and monetary supervisors are too confidential to be saved on a third-party cloud infrastructure.
Second, experimenting directly, even with easy use instances, to search out the correct approach of integrating AI into our exercise, resulting in an “augmented agent” fairly than a “substituted agent”. Certainly, we will anticipate AI to considerably reshape the patterns of human-machine interactions. Discovering the correct combos will encourage the adoption of the brand new instruments, by profitable the buy-in of customers, which is an important difficulty.
Third, collaborate with others, to share greatest operational practices and to construct a coherent AI supervision framework. In fact, I’m pondering first of worldwide cooperation, as a result of AI-related points are by their very nature international. On this space, whereas there could also be nuances by way of methods to proceed, I word above all that many jurisdictions are expressing comparable considerations, which ought to allow worldwide cooperation to maneuver ahead. However we additionally must cooperate with authorities in different sectors, particularly competitors, cyber safety, basic rights and even the inexperienced transition, as AI-related considerations are largely interconnected. For my part, these completely different types of cooperation are an important situation if we’re to contribute to the emergence of probably the most related and resilient AI fashions, in different phrases, if we’re to affect the event of the expertise within the path of the final curiosity.
Thanks in your consideration.