Don't trust and check. Why Artificial Intelligence Lies and What to Do About It

AI is well established in many areas, from medicine and law to finance and marketing. Behind brilliant opportunities hides an unpleasant feature: often neural networks pass lies for the truth and do it with such confidence that even professionals do not notice the catch. The statistics are impressive: according to HalluHard, a benchmark that tests AI’s tendency to hallucinate in multi-stage dialogues, even the most powerful models with Internet access are wrong 30 percent of the time. And if you turn off web search, the amount of fiction jumps to 60 percent.

In artificial intelligence, this phenomenon is arranged almost like a natural one: information that does not exist in reality is depicted as plausible and quite logical. The root of the problem lies in the very architecture of neural networks, explained at the Department of Intelligent Information Technologies BSUIR. AIs are trained on huge arrays of texts, catching statistical patterns. The model simply predicts which word or phrase is most likely to appear next. Unlike humans, an algorithm has no internal picture of the world. If the database does not contain information on a given topic, the system can come up with them from similar facts, names or language patterns. Completely wean the car from fantasies, alas, it is impossible. But reducing the number of errors to an acceptable level is quite possible. The specialist advises to formulate clear and strict restrictions for the neural network: specify the sources of information, demand to separate facts from assumptions, ask for references. Here is an example of a working prom: “Respond only on the basis of attached documents. For each important statement, state the source. If there is not enough data, write about it and do not add guesswork to the answer.” But hallucinations are not limited to the list of dangers. There is a more disturbing phenomenon called model collapse. Neural networks draw knowledge from the Internet and inevitably encounter the content of their predecessors. The professional community is sounding the alarm about the “pollution of science”: a recent study of the electronic archive arXiv found almost 150,000 fictitious citations on the largest scientific repositories. So, some algorithms take as a proven scientific fact information invented by other algorithms. In recent years, neural networks have improved: databases are constantly updated with fresh news, and the function of automatic search on the Internet has become the standard. The problem of errors is less acute, but it has not disappeared. Instead of gross factual blunders, modern models often stumble on subtleties. For example, they can offer a dubious site, mix fresh data with old ones, make a curved conclusion, and attach a fictional link for greater persuasiveness. In general, rely on AI, but do not be a fool: even the most advanced algorithms do not relieve the user of the responsibility for fact-checking.

Special vigilance when working with artificial intelligence should be shown in critical areas, emphasizes the head of the Department of information protection BSUIR Olga Boyprav. Mistakes in medicine, law or finance can result in real losses and legal consequences. One of the most high-profile precedents occurred in 2023: an American lawyer entrusted ChatGPT with the preparation of a lawsuit. Without thinking (or thinking at all), a 30-year-old lawyer has inserted neuron fantasies into an official document. The case almost ended with a disciplinary action.

Before pressing the “copy” button, the expert urges you to carefully check the sources on which the bot relies: – First of all, click on the link and make sure that the source resource is really the same numbers, dates and facts as in the answer. Pay attention to whether the connection to the web server is protected: a lock icon should burn next to the address bar. If the domain name of the resource is registered recently (you can check this through special services like whois.ru), be skeptical of the information. Why it is impossible to feed AI their data Deepfakes are even more dangerous. The first videos, where neural networks were forced to dance kittens or talk tomatoes, caused only a smile. Now, highly generated content with live faces, natural speech and believable emotions has become a real tool for manipulation. How not to get caught? Olga Boyprav recommends to look at the details: whether the lips move synchronously, whether the clothes of the characters are suitable for the environment, whether the shadows naturally fall, whether there are strange artifacts.

- To establish whether the file was processed, will help, for example, the service FotoForensics, says the expert. But the threat lies not only in what the neural network creates, but also in the fact that we trust it. Why not give the bot a routine: say, upload passport details to draft a contract or corporate reports for analytics. Quick, convenient and... fraught with consequences. All the information fed by AI is deposited on the servers of the developer company, and it is far from the fact that there is due attention to security issues. With a successful hacker attack, your secrets may well go to the black market. If your data remains in the database of the service - your full name, place of work, phone number - it will not be difficult for fraudsters to use them for their own purposes. In addition, many developers use correspondence with users to train their models. It is possible that a trade secret or personal data will one day surface in a response that the AI will form for another person, concludes Olga Boyprav. Effective and safe how to work with neural networks

Do not feed sensitive data to AI services. If a neural network is still necessary to work with large amounts of personal information, pre-depersonalize it: remove names, addresses, phone numbers and other identifiers. Always check the sources referenced by the neural network. Even if the answer looks flawless, follow the links, check the numbers, dates, and facts. If you publish information created by a neural network, honestly indicate this.