"The main risk in using AI is a lack of critical thinking"

The use of artificial intelligence-based models is increasingly penetrating our lives. The banking sector is one of the leaders of the process: the extensive needs in the field of working with big data force banking specialists to develop new AI—based models that help both bank employees and its clients. Maxim Konovalikhin, head of the Department of data analysis and Modeling, Senior Vice president of VTB, spoke in an interview with Izvestia about how his work with AI models began for him and where it will lead fintech in the coming years.

"Artificial intelligence is being overly humanized today"

— To what extent do the current capabilities of digital assistants and artificial intelligence systems correspond to the image that is being formed in the media? Is there a gap between expectations and reality?

— There is a gap, of course. The media image of AI often promises an almost human level of understanding and full automation of the service, while in practice the role of humans in AI processes remains significant. Today, artificial intelligence, in my opinion, is being overly humanized. In fact, these are still programs that the average user doesn't always notice working. Besides, these programs are not as omnipotent as many people think. They have certain limitations related to their technical capabilities, the available capacities, as well as, in fact, the data that the AI works with. This data is not always of high quality, and therefore it often happens that we cannot get what we want from AI. So there is a gap, of course, but it is narrowing over time.

— How long do you think it will take to bridge this gap? And are the technological tools ready for this?

— Yes, the necessary tools exist. And the technology is already advanced enough to bridge the gap. However, there are risks involved. Generative artificial intelligence, large language models are basically non-deterministic. Yes, they are able to maintain a conversation, but it is impossible to guarantee that this conversation will take place in the right direction and at the level of quality that suits us. This is the very example in which our expectations outstrip the available opportunities.

Subconsciously, the user gives them not only competence, but also responsibility, reducing the criticality of the relationship. And this is just the key risk. The danger here lies in false confidence, since confidence, the unbiased and clear style of the AI model formulations does not depend on the real plausibility of the text. She doesn't reflect, and that's the catch. How to correctly define responsibility, how to teach people to use the tool carefully and critically, and how to teach an AI model to explicitly express the degree of confidence in the information provided are the main challenges.

Whether we will be able to overcome these limitations, how quickly and efficiently, depends on many factors, and, first of all, on the quality of training that AI receives, on who will teach this language model and how.

On the other hand, optimism is inspired by the fact that today bots — voice assistants, for example — are already working quite adequately compared to what they were three or four years ago. Therefore, there are really many components that primarily depend on learning ability, on learning. But in general, I think that in the very near future we will be able to get the quality that we really like.

— Where does artificial intelligence clearly act as a human assistant today, and in which areas does it already create real risks and threats?

— As an unambiguous assistant, AI acts where it is necessary to quickly aggregate a large amount of information, structure it and present it to the user in a readable form. This is how modern search engines using AI functions work, for example: when we want to get information about the effects of drugs, about the features of a particular product, etc. Such operations can significantly save time, so people actively use them. At the same time, you definitely should not forget about the risks of receiving distorted or incomplete information. To mitigate these risks, the user must critically look at the results of the AI search and double-check the sources of information.

The real risks at this stage of technological development may include, for example, autopiloting a car using AI, which can lead to real human casualties. But it is necessary to take into account the opposite cases, when the car's autopilot, based on radar data and a collision avoidance system in fog, prevented serious accidents. We often don't know about such cases, and there may be more. Therefore, it is necessary to calculate and draw conclusions very carefully.

"A national AI tool is necessary, and it should be freely available"

— The so-called universal large language models are developing in the world today — in English, as well as national models — in Chinese, in Russian, etc. How important is it, in your opinion, to develop large language models specifically in national languages, rather than relying on universal ones?

— It's important. There is such a thing as national identity, which should not be forgotten. A national instrument is needed. It allows you to preserve national identity, strengthen cultural ties within a particular region, country, etc. And already on this basis, universal models should be formed. They contribute to the mutual comprehensive enrichment of representatives of various social and national groups, ethnic groups, social, cultural, linguistic, etc. Due to their versatile approach and the involvement of a large number of users, they are constantly improving and easy to learn. Thus, universal models will appear exactly where it is necessary and useful for people, they will not be artificial, which means they will be in demand.

— Do you think the future belongs to national models that take into account the cultural context, or to global universal solutions?

— I believe that they should develop in parallel. The development of one AI model leads to the development of another. Now, in our difficult times, friendly countries, no matter what, strive to unite. And it's fine if AI tools contribute to this. On the other hand, it is necessary to carefully analyze the geopolitical risks: if their level is high, I would be careful not to invest in such projects. Nevertheless, AI is a very powerful tool that can influence both human communication and culture in general. Therefore, the development of certain AI models should be approached individually and very carefully.

— What are the key opportunities and main risks of the spread of AI for the younger generation?

— The main risk is a lack of critical thinking. Artificial intelligence—based models, on the one hand, are not perfect today and make mistakes, but on the other hand, they are able to impose their opinions qualitatively. But if you don't think critically about their answers, there is a risk of making a big mistake in choosing a solution — in any area. In addition, AI is so convenient and accessible that it is easy to succumb to temptation and not think at all, but simply use the information offered. As a result, you may lose the ability to think critically altogether — as they often say now, young people will stop thinking and will only push buttons.

But if we approach information from AI critically, if we comprehend and verify it, and use this tool where it is really needed — to select hypotheses, to find the necessary facts — then AI will be a powerful help. We have a lot of young people working for us, and I see from their example how much such use of it can give.

— How necessary is it to teach older people the basics of interacting with AI, for example, to protect themselves from fraud and just to stay in the context of modern life?

— Here's a question for the people themselves: you can't teach anyone anything by force. But as far as I know, most people at any age are interested in this topic. And I think it is necessary for everyone to master AI tools at the level of mastering digital literacy today. My uncle is 90 years old, he has successfully mastered instant messengers, we correspond with him. So when there is interest, why not?

At the same time, AI services should not become some kind of complex tool that requires painstaking training to interact with. It should not be like programming, which needs to be specially trained for a person. AI services should be native and easy to use.

Speaking about their use by older people, it is also difficult to give any individual recommendations, because they are universal for young people, pensioners, and all other age groups - maintaining a level of critical thinking and careful perception of new information.

"We will come to the conclusion that a segment is not a group of customers, but a single customer"

— In what ways has AI fundamentally changed the banking industry over the past 10-15 years? What was the most significant thing?

— The changes are really very noticeable. Twenty years ago, in 2005, we launched one of the country's first credit pipelines at VTB-24. And in a sense, he became the main point for us, the basic model on the basis of which we were engaged in the creation of new AI models - advisors and others. And this technology began to develop very quickly: already in 2015, we were engaged in behavioral models, behavioral scoring, and began developing recommendation AI systems.

Moreover, in the West, this topic began to develop around the same time, but, tellingly, not in the banking sector, but in social networks. And in Russia, large banks are dealing with really large amounts of data, so we really need AI technologies. And now, starting with scoring, twenty years later we see that hundreds of banking processes are based on AI, right up to computer vision, which until recently was a fantasy.

The scope of AI applications is constantly expanding. Those processes that it has not yet reached due to high model risks are likely to be covered by AI technologies in the future.

The development of artificial intelligence and data analysis technologies allows banks to create more and more personalized services. We are observing this around us today. Banks may not be guided by any "cliches" in the needs of potential customers, as it was before. It's about analyzing patterns among hundreds of different parameters. Ultimately, this gives the bank the opportunity to offer the client the services they really need, on the terms that really suit them. This is a fundamental change in the banking industry in recent years.

— How do you see the bank under the influence of AI in 5-10 years, what will change dramatically for the client?

— Hyperpersonalization is already becoming the main trend and will become even more so. Now we divide clients into segments based on sociodemographic data, but in the end, we will come to the conclusion that a segment is not a group of clients, but each client individually. And if he agrees to the processing of personal data, we have the opportunity to make unique personalized offers. Therefore, personalization, in my opinion, will become the main trend, which will become more and more evident over time. And besides, the experience of China will probably be used, where today bank branches serviced by robots are operating 24/7. It's convenient, profitable, and technically feasible, so I think we'll get to that soon.

— When did you realize for yourself that AI is not just an academic discipline, but a technology that can greatly change the lives of millions of people?

— More than ten years ago, using the example of underwriting. That was the first time we used the "Underwriter" AI model - it was more complex than the scoring model and radically changed the current state of affairs at that time, qualitatively improved the product. And from the reaction of the underwriters, we realized that our model was just a salvation for them. That's when I realized, this is the future.