15.09.2025, Aleksandrovskoye.
Artificial intelligence is useful provided that a person has some competence and is able to evaluate the facts and conclusions it produces
Artificial intelligence (AI), its capabilities, and its prospects are increasingly becoming a subject of debate, both among experts and among people not working in the IT sphere. It is perceived as having both useful opportunities for humans as well as negative, and sometimes even dangerous, ones. But despite the ambiguity of this technology, Russian government has placed a major bet on the development of artificial intelligence. It is being actively introduced into all spheres of activity, even into those that would seem to require human presence the most — education and medicine.
What should we expect from this new turn in technological development, this time digital? Will AI replace humans, and if so, where and how? How soon will this happen? Finally, should we prepare for a “machine uprising,” or are such fears greatly exaggerated? We discussed this with AI research engineer Andrey Zhonin.
Interviewer: Andrey, tell us what you do.
Andrey Zhonin: I am a research engineer in the field of speech technology. This is one of the areas of artificial intelligence. More accurately, it should be called machine learning. We work on speech recognition, speech synthesis, natural language processing, including the very large language models that have now caused a big stir.
Interviewer: The very term “artificial intelligence” sparks great debate; it has many definitions, sometimes contradictory.
Andrey Zhonin: This term has been used since the 1960s. Most specialists today believe that AI is simply something that replicates human intellectual abilities: vision, hearing, logical thinking, the ability to understand text, to compose, to create, and so on. In other words, one might assume that if a machine were to gain free will, it would in some sense become equivalent to a human. Such a notion exists in certain circles, one could even call it an ideological premise. But, of course, that is not the case now; machines currently do not possess free will or consciousness.
Interviewer: So we can still treat artificial intelligence as a tool, not as some kind of sub-personality?
Andrey Zhonin: For now, yes, as a tool. But this raises many questions about the future of humanity, because this tool will develop, and sooner or later its intellectual capabilities will surpass those of humans. That means we will have a tool disproportionate to a human being. And in connection with this, we must think about human development and humanity’s place in this world.
Interviewer: In which fields can AI replace humans, and in which can it not? Which professions will soon become obsolete for humans?
Andrey Zhonin: First to be replaced will likely be professions that are somewhat routine and easily automated. For example, salespeople and drivers. We already see the emergence of supermarkets without cashiers. There will still need to be a person in the store to resolve complex issues, but the number of sales staff can certainly be reduced. The same with drivers: one can imagine a situation where a single person controls not one vehicle but several, thus reducing the need for workers. This is bound to happen. Perhaps in time, support staff and sales managers in mass markets will be replaced. That is probably what we should expect.
Interviewer: Russian authorities are actively promoting AI technologies in medicine and education. Will AI be able to replace people in such professions as doctors and teachers?
Andrey Zhonin: I strongly doubt it. Last fall, there was news about the creation of an AI system for decoding tomography results, but later the Ministry of Health decided to discontinue its use. It turned out the system made major errors if the tomography data were to certain extent complex, if the equipment was different, or if the situation simply went beyond the system’s training dataset.
A human, however, even in an unfamiliar situation, begins to think logically and at least understands that this is a different situation. A machine cannot do this yet. Even for narrow medical tasks, such as analyzing tomography, for example, substantial refinement is required. Doctors talk about nosology, that is, classification of diseases. And they have tens of thousands of cases in this nosology. For an AI therapist to be useful, it would need to recognize those tens of thousands of cases. Then there are combinations of symptoms, special cases, and in general everything is individual to each patient.
At the AI Journey 2024 conference last fall, Sberbank presented a system that, as they said, could recognize 400 cases in nosology. Compare that: 400 versus tens of thousands. It is hardly possible for an AI therapist to replace a doctor.
Interviewer: But how can AI help doctors?
Andrey Zhonin: Our unfortunate reality is that staff positions are being reduced, especially in sparsely populated areas, villages, district centers, and so on. As far as I know, clinics are being centralized, and there may be a temptation to replace some doctors with an AI system that makes diagnoses, because “if there are no doctors anyway, then let it be this way.” But in reality, replacing humans today is only feasible in analyzing technical information. For example, tomography, 3D tomography, or AI-based MRI. These are built on similar technologies.
Additionally, doctors need quick access to precedents, similar cases of prescriptions, and so on. They need a search system to represent patient medical histories in a structured format. This could be a database where similar cases could be found and recommendations obtained. In the USA, where doctors’ labor is very expensive, and recording results from tests or examinations takes much time, the process was automated: a doctor dictates, the system recognizes speech and fills out the patient’s records. That really saves time and is useful. We are also working on this.
Interviewer: Of course, as a tool to assist doctors, AI is undoubtedly needed. But now there is an attempt to implement AI in schools, to check homework, and even to select an individual educational path for a child, building a database for this purpose.
Andrey Zhonin: I am doubtful about this approach, because one must understand how AI learns. It requires a training dataset in order to make the right decisions. In this particular case, we do not know which intermediate steps were correct; we only know the final result: for example, “a successful professional who loves their job.” And we have to somehow figure out, based on the final result, which intermediate steps were correct. The specificity of such systems is that in order to train them correctly, we need millions, and in some cases billions, of datasets.
I can hardly imagine this being realistic. Moreover, historical situations change, society’s needs change, psychology of children, population changes. How does the domain, i.e. the field of applicability, which are constantly talking about, change? But AI systems require stability in the domain; otherwise, models stop working accurately. How then can such a system be trained reliably? It will in any case perform worse than the worst teacher. Only slightly better than random choice.
As for homework checking, I don’t understand where this demand comes from. Is the teacher so overloaded they have no time to check, or is homework structured so that no human is needed? If answers can be checked formally, that’s essentially a test. And it’s unclear why artificial intelligence is needed here, because the test requires to select several options or wrote the answer in the boxes. There are already systems for this, they work fine, and there’s no need to call them artificial intelligence.
If we’re talking about essay checking, last fall Yandex reported they had created a system that could write essays passing the Unified State Exam (EGE) in Russia. I looked at how it did this, and I had serious concerns about the requirements for such essays. If the point is just to meet formal requirements, including key words and phrases, then what is the actual purpose of such essays? Here, it isn’t AI being raised to the level of a teacher; it’s students being lowered to a level AI can check. This is, in my view, a negative phenomenon.
Interviewer: So the temptation is to deliberately lower students’ education level so that AI can cope with the work?
Andrey Zhonin: If you look at how essays are structured, their formal requirements aren’t bad in themselves. Formal requirements exist for dissertations too. But a dissertation that only fulfills formal requirements and nothing more is worthless. The same should be true for essays, yet for some reason it isn’t. For some reasons, it is assumed that formal requirements are enough. Naturally, this is degradation. Training a person, who actually should be a creative individual, for something that is checked by a non-creative system is a bad approach.
Interviewer: Does this mean that developing creativity in children is incompatible with AI? Since AI cannot, in principle, comprehend or recognize creativity.
Andrey Zhonin: Yes. One cannot say that AI can evaluate creativity. It may serve as a training tool, but even here there are problems. One can expect that AI can “hallucinate”, in other words output false or erroneus information. For teaching history, for example, that is unacceptable. Experience shows that when a student turns to AI, there is no guarantee the information is reliable. It is simply impossible. AI still cannot replace classical textbooks.
Interviewer: Maybe humanity just isn’t ready to build a universal or powerful enough AI yet? Maybe in the future, technologies will change, and it will become possible?
Andrey Zhonin: Technologies will change, of course. But as for current large language models, which are what people usually mean by AI today, they inevitably hallucinate, and nothing can be done about it.
Interviewer: What does “hallucinate” mean?
Andrey Zhonin: They invent facts. They invent facts or relationships between facts. They say things that don’t exist. They can even invent fake references. Developers try to reduce such errors, but completely eliminating them is impossible. I’ll explain why it is like that. At today’s level, all these systems are trained on human texts. Literally, all internet texts are collected (and now also videos, i.e., texts are collected from YouTube and all possible sources of video recordings), and systems are trained on them.
But the companies doing this, even huge ones like OpenAI, Microsoft, Google, Facebook (organization banned in Russia), Yandex, Meta (organization banned in Russia), cannot verify all those texts. They cannot even read them all. It is physically impossible. They cannot fully check the datasets used to fine-tune these models. In that sense, we are training models on human intellectual output, which is not always of good quality.
Yes, language models gain certain knowledge in some areas, but it’s not quite the same thing. And there’s another issue: the more time passes since these language models were created, the more texts generated by them appear, and these texts end up being used to train these very language models, which negatively affects their performance.
As for the future, we can only speculate. First and foremost, what is a human, really? How does a human differ from a machine? If we say a human is a creative being, then unlike a machine, a human can generate truly new information.
All neural networks and information-processing algorithms are designed in such a way that they cannot generate new information, they only recombine what they already know. We may consider an AI-generated Kandinsky-style painting something “new,” but in fact it’s a subtle recombination of things already known to the model. A human, on the other hand, is capable of creating entirely new things.
Looking at human history: in the Stone Age there was nothing, and now we have a vast civilization full of things not found in nature. That means humans invented something new. Therefore, I believe we should not expect machines to ever match humans in this sense, but perhaps they will become something else for humans. Perhaps they will become some kind of intellectual amplifier or even, in a sense, a partner. There are many very interesting, mostly fantastical ideas on this subject. But they relate to the future, not the current state of affairs.
Interviewer: How can AI be used in the defense industry, where human lives literally depend on its work?
Andrey Zhonin: Let’s recall a recent case involving Israel’s Lavender system, which was used to identify Hamas members. They used this system to determine strike targets in Gaza. According to the Russian International Affairs Council, the usage of Lavender system resulted in the death of more than 30,000 civilians, including many women and children. People died because Israeli forces used an AI system. Of course, this was primarily human error. The system was trained on something.
It had 0.9 reliability, meaning 90% accuracy in identifying Hamas members. But they began using it more broadly, in different conditions. Hamas fighters don’t stay in one place for long, a member might visit a house, do something there, and then leave. Hours later, the house would be struck, but other people lived there. Israel decided: collateral damage doesn’t matter if there’s a chance to kill a Hamas member. That’s military logic but also misanthropic logic, I would say.
We have entered a new era, where robots are beginning to fight each other. We already face drone warfare. It is likely that victory will go to those who are able to develop the most advanced systems, including autonomous ones based on the use of artificial intelligence. Last fall, the US launched the Replicator initiative, aimed at mass deployment of unmanned systems. They finally admitted that costly piloted platforms like the F-35 are unsustainable, since their losses are too costly. The US is investing in “attrition warfare” — cheap, modular drones, produced in large numbers.
They believe modularity will provide flexibility. Autonomous control systems are also being built: humans only assign tasks, while drones fly, drive, or sail, completing them. Losing them isn’t a problem, since they cost much less than, say, an F-35. The F-35 then stays at a distance, controlling everything. That’s how it will likely work.
Naturally, many groups are developing autonomous drone management and counter-drone systems. Since drones fly very fast and humans cannot react to them quickly enough, we naturally need a system that can do this faster than humans. It will undoubtedly be autonomous and AI-based. Imagine a turret monitoring the air, small, fast-reacting, able to identify drones, distinguish friendly from enemy, and shoot down the hostile ones. Swarm systems will also emerge. They haven’t succeeded yet, partly due to computational and energy limitations.
Military experts warn that AI tech is dangerous because, unlike nuclear or missile technology, it’s easy to implement — even by small actors, political or military. You can download, use, apply. Eventually, these technologies will be in everyone’s hands. What kind of world will that be?
In essence, it will no longer be clear who is doing what, because anonymous drones will, for example, carry out terrorist attacks, and it will be unclear who organized them, why, and what to do about it. That is not the case yet. Despite bold claims, these systems still work poorly. In some narrow areas, problems will be gradually overcome. But working solutions will eventually appear.
Interviewer: So AI becomes a powerful universal tool, usable for both good and harm?
Andrey Zhonin: Yes, certainly. But most frightening, in my opinion, is the possibility of using AI to assemble proteins, molecules, and perhaps eventually even living organisms with specific properties.
For example, one could imagine assembling viruses. I think this would still be difficult, since our understanding of life is, to put it mildly, very rough. The whole story of vaccines showed that. It is complex, and we likely lack the knowledge to create something apocalyptic. But some problems will certainly arise. And in general, creating a virus no longer requires national-level institutes or industries. A medium-sized lab with specialized biotech machines (which already exist) is enough. Any moderately wealthy actor could build such a thing. Naturally, some will want to for various purposes.
Interviewer: And such purposes may be extremely dangerous. Against the backdrop of AI’s development, more and more questions arise about human development and human’s role in this new world.
Andrey Zhonin: Yes, certainly. Military experts are discussing another scenario related to the fact that decisions made by these artificial intelligence systems are made very quickly. It may well be that a world war will start unintentionally because one drone attacks another, which then attacks in response, thus triggering a process of escalation. A situation could arise in which we learn about a global nuclear war literally a minute before it begins.
AI requires human oversight. And with our education system, the future is frightening. Who will oversee AI systems if education is reduced to serving AI. That is, adapted to AI’s capabilities?
Yes, AI is objectively useful. But only if the human has competence and is able to evaluate the facts and conclusions AI produces.
Interviewer: Thank you, Andrey, for a very interesting conversation.
This is a translation of the interview first published in The Essence of Time newspaper, issue 603.
