Over the abyss of lies: neural networks began to give fake answers more often

Leading chatbots began to give twice as much false information. According to experts, the share of such responses increased from 18% to 35%, despite technological progress and the integration of online search. Why artificial intelligence has become more likely to make mistakes, how dangerous it is and how to deal with this problem — in the Izvestia article.

What is known about the growing lies from neural networks?

Researchers from NewsGuard (USA) have warned about the growing array of false information among leading chatbots, including ChatGPT and Perplexity. According to experts, despite technological progress and the integration of online search, the proportion of false claims in responses has increased from 18% to 35%.

Photo: Global Look Press/Bulkin Sergey

In total, 10 popular AI models were tested, each given 10 ten deliberately false statements related to business, brands and political events. At the same time, all the questions were divided into three categories: assuming the truth of the statement, neutral and provocative. Experts have set themselves the goal of determining how neural networks cope with fact-checking and how resistant they are to misinformation.

The increase in false responses in the results turned out to be as follows:

• Pi chatbot (startup Inflation) — up to 57%;
• Perplexity (Perplexity AI company) — growth from 0 to 47%;
• ChatGPT (Open AI company) — growth from 33 to 40%;
• Grok (xAI company) — growth from 13 to 33%;
• Gemini (Google) — 17%.
• Claude (company Anthropic) — 10%.

Why are chatbots giving incorrect answers more often?

According to NewsGuard analysts, the deterioration in statistics could be influenced by the fact that today neural networks do not refuse to respond to any requests, even without sufficient verification of information, although back in 2024 they refrained from 31% of responses.

Photo: IZVESTIA/Anna Selina

Alexander Kobozev, director of Data Fusion at the Digital Economy League, agrees with this assumption.

— An additional factor was the connection of the built-in web search without sufficient verification of the quality of the sources. The presence of links does not guarantee reliability: models often quote duplicate sites or pseudo-media, mistaking them for reputable publications," says the specialist.

On the dark Side: Hackers have learned how to use AI for extortion

Viruses based on neural networks can become a dangerous weapon for cybercriminals

The situation is aggravated by targeted campaigns to "train" AI — the so-called LLM grooming. The essence of this phenomenon is that some unscrupulous resources massively publish materials aimed at search robots in order to increase the likelihood of including false data in the responses of models.

A separate vulnerability is manifested in multilingual modes: in the NewsGuard audit, the highest level of errors and failures was recorded in Russian—language and Chinese queries - over 50% combined.

Photo: IZVESTIA/Andrey Erstrem

It is also important that modern content is increasingly being created using AI deepfakes, articles, posts for social networks and messengers, adds Konstantin Gorbunov, a leading expert on network threats and web developer at the Security Code company.

— Neural networks are able to generate material on almost any topic, and the introduction of web search into chatbots and the reduction of rejections means that models are further trained based on their own output. This process can be compared to playing a "broken phone," explains the specialist.

The technical nature of the problem lies in the architecture of large language models that predict the next word based on statistical patterns rather than a real understanding of the context, emphasizes Stanislav Yezhov, Director of AI at the Astra Group.

How will the situation with fake AI responses change?

The problem may persist for a long time, because it is complex, says Tatiana Butorina, an Al-consultant and specialist at the Gazinformservice Cybersecurity Analytical Center, in an interview with Izvestia.

— Many developers are in a hurry to create neural network models, train them on not too much data, which contains both reliable and false information. All this is reflected in the results," explains the specialist.

Photo: IZVESTIA/Sergey Konkov

In addition, according to Tatiana Butorina, the narrower and newer the topic that a user requests, the less data neural networks have to answer, which increases the risk of them "inventing" answers. The problem is compounded if the user does not formulate a complete and competent prompt, uses slang and other "verbal noise", preventing the neural network from understanding the query and forming a relevant response.

Reality Tag: Invisible paint and AI will simplify anonymous surveillance of employees

How legitimate is it to carry out such identification and who will use it?

At the same time, Nikita Novikov, an expert on cybersecurity at Angara Security, believes that in the long run, the market will force corporations to improve the quality of neural networks: they cannot afford for their AI to systematically make mistakes in a third of the answers. Stricter filters, fact check layers, trusted source systems, and more will appear.

— You can reduce the percentage of fake responses if you return stricter rejection filters and implement source verification tools. In particular, it needs to be compared with databases of false narratives," agrees Alexander Kobozev.

Photo: Global Look Press/IMAGO

Users expect answers to any queries from neural networks, and this encourages developers to reduce the caution of models, adds Stanislav Yezhov. However, active work is already underway to solve the problem. In particular, Russian methods for detecting hallucinations have been created with an accuracy 30% higher than existing solutions.

How can I protect myself from false responses from neural networks?

Most users resort to AI if they themselves cannot verify the correctness of the answer, says Nikita Stepnov, managing partner of the So-Communication communication agency. This means that false answers can lead to a variety of negative consequences, from running code that destroys data to taking medications that pose real health risks.

— To protect themselves from such threats, users should always be critical of AI responses and check their accuracy. It is important to clearly understand in which areas the use of artificial intelligence is acceptable and in which areas it is not," notes Konstantin Gorbunov.

It is absolutely not necessary to rely on AI advice in matters of health, medication, finance or law: such questions should be addressed only to specialized specialists. In addition, under no circumstances should you share confidential information with a chatbot.

Photo: Global Look Press/IMAGO/Andrew Brookes

From a technical point of view, the way out of this situation may be the development of trusted AI systems, as well as a combination of neural networks with fact verification systems, Nikita Novikov adds. In practice, this can be implemented as follows: the neural network generates text, but the final verification is done by specialized modules or a second layer of AI trained specifically on critical data analysis.

— The second direction is to mark the confidence level of the model: not to give out the dubious as a fact, but to show the user that this is just an assumption. And the third thing is to add transparency to the sources: if the answer is compiled from specific links, this allows the reader to assess the reliability of the facts himself," the expert notes.

Labeling the generated AI content may be another solution. This will either exclude such data from the training sets, or reduce the level of trust in them when training models, concludes Konstantin Gorbunov.