- Статьи
- Internet and technology
- Tangled neural networks: AI agents began to hallucinate when interacting with each other
Tangled neural networks: AI agents began to hallucinate when interacting with each other
Russian software developers have reported the appearance of a new type of error in the work of neural network assistants — the so-called interaction hallucinations. We are talking about failures that occur not within one model, but at the junction of several systems: for example, when a digital assistant performs a user's task and accesses an external service using another developer's AI. This situation is possible, in particular, when booking tickets through automated platforms, where a different language model is used on the partner's side. Experts warn that such mistakes can affect finance, medicine, logistics, tourism and customer service.
The rapid growth of AI creates new risks
In Russia, problems began to be fixed when trying to book movie tickets or book a table with the help of domestic AI agents. We are talking about situations where individual systems work correctly, but when data is exchanged between them, the context is lost, the meaning is distorted, or the logic of the scenario is violated. As a result, the user does not get the result he expected: from an erroneous order and an incorrect payment to a contradictory consultation or a hung operation.
The market today resembles a set of "digital fortresses": large players form closed ecosystems within which services interact effectively, but poorly integrate with external solutions, said Ruslan Dolgopolov, head of the Gazprom ID Operator group. The lack of unified data exchange standards and unified APIs means that telecom companies, banks, marketplaces, and delivery services actually speak different digital languages.
He compared what was happening to a "corrupted phone": when transferring a task between systems, the data goes through several levels of interpretation and may differ significantly from the original request at the output. According to him, without the formation of a common interaction environment, such mistakes risk becoming not an exception, but a systemic characteristic of the digital economy.
IT expert Sergey Pomortsev clarified that the problem is already manifesting itself in everyday scenarios. According to him, one of the real examples is related to an attempt to buy movie tickets through an AI assistant.
— The user formulates a simple request — choose an evening session, choose good seats and pay for the order. However, after transferring data between several systems, the assistant may confuse the date, choose another cinema, or make a reservation without completing the payment. As a result, the person learns about the failure before the session starts," he noted.
The second example, according to Pomortsev, concerns the search for rental housing on sites with a high degree of automation, similar to real estate advertising services. The user asks to select an apartment based on price, area and rental period, and the digital agent begins to interact with external databases and chatbots of the owners. At this stage, key parameters are often lost: the budget changes, the area is replaced by a neighboring one, and a short-term lease turns into a long-term one. As a result, a person gets dozens of irrelevant options and spends time re-filtering.
— The problem is in the inference — that is, at the moment when the model is already applying knowledge to a real task and must correctly interpret incoming signals from another system. Everything works on paper, but semantic failures begin in the living bundle," the expert noted.
However, not all market participants consider the problem to be new. Head of the AI department Cloud.ru Dmitry Yudin noted that full-fledged agent interaction, in which one AI system transfers tasks to another in real time, is still rare. According to him, most integrations are still based on standard APIs that provide basic data exchange, so it's premature to talk about the widespread nature of the phenomenon. At the same time, the issue of system compatibility already requires the attention of the industry.
A similar position is held by the Director of products of MWS AI (part of MTS Web Services) Maxim Voloshin. He stressed that the very term "interaction hallucinations" has not yet been supported by a significant number of public cases. However, according to him, the compatibility problem does exist, primarily due to the lack of a single transport layer and understandable mechanisms for transferring context between systems.
In which areas does the user encounter errors?
Such failures are most noticeable in mass services. Director of the Lukomorye Ecosystem AI Product Development Center (part of Rostelecom) Denis Romanov noted that in the financial sector, data transmission errors can lead to funds being written off without crediting them or incurring penalties due to payment delays.
— In tourism, the desynchronization between flight and hotel booking systems can leave a customer without a room upon arrival. A person is forced to urgently look for new housing, overpay and solve a problem in an unfamiliar city," he explained.
Another common scenario is related to purchases through voice assistants: if the product, size, or delivery address are incorrectly recognized, the user receives the wrong order or shipment to the wrong address. In some cases, personal data leaks are also possible if the request is processed with errors and ends up in the event logs.
In medicine, the consequences can be more serious. According to the expert, when making an appointment with a doctor, monitoring medication intake, or initially analyzing symptoms, the system can "lose" critical information, such as information about a patient's allergy. This increases the risk of errors in prescriptions and increases the burden on doctors.
The loss of the client context is particularly sensitive for the financial sector, said Anton Graborov, member of the Management Board of Alfa-Capital Management Company. If data about an investor's risk profile or preferences is distorted during transfer between services, this becomes not just an inconvenience, but a potential regulatory risk.
— The root of the problem is not so much in the models themselves as in the architecture of the solutions. Individually, the models can work fine, but when linking closed ecosystems with different data formats and without uniform validation rules, the distortion of meaning begins,— said Vladislav Kudinov, Veai's Service station.
Maxim Malyshev, General Director of Notamedia, holds a different point of view. He believes that more often it's not about incompatibility of models, but about insufficient engineering: with the correct transmission of context, different systems can work effectively in the same chain — for example, one classifies data, another generates text, and the third performs result verification.
— Some of the problems are related to the limitations of the models themselves: context length, memory, and learning quality. When the query chain becomes too long, the system literally "forgets" the beginning of the dialogue and begins to complete the missing data on its own," added Irina Mezheneva, a leading analyst engineer at Gazinformservice.
Systems can technically exchange data, but interpret the same information in different ways. In this case, the failure occurs precisely at the point of transmission: one agent conveys meaning in one form, while the other interprets it differently, explained Stanislav Yezhov, Director of AI Development at Astra Group.
Most experts agree that the market needs standards, but not in the form of strict certification of each model. Dmitry Yudin noted that technical protocols such as MCP and A2A are already developing in the world, setting the rules for the exchange of context between AI systems. Ruslan Dolgopolov added that without legal mechanisms for data exchange and unified APIs, the market risks finally fragmenting into incompatible ecosystems.
While the technology is under active development, experts recommend that AI assistants be perceived as a support tool rather than fully autonomous performers, especially in areas with a high cost of error, such as finance, travel, medicine, and document management. At the same time, the key risk remains not so much an outright "fantasy" of the neural network as a confident error that occurs in a chain of several digital intermediaries.
Переведено сервисом «Яндекс Переводчик»
