The popularity of AI is rapidly growing: according to Eurostat data, in 2025, 32.7% of the European Union population aged 16-74 used generative AI, and 15.1% of Europeans used this technology for work purposes. Lithuania exceeded the EU average in the use of generative AI: 36.9% of the population in the country used such tools.
Read more Trade for Dambrauskas – «Sabah» sells the son of the French football star
However, along with the growing use of AI, questions about its reliability are also increasing. Although the technology can generate text, analyze data, or prepare presentations in seconds, its responses still contain errors that can mislead users and pose reputational risks.
One of the most unexpected and hardest to explain phenomena for users is when AI suddenly writes a word in Cyrillic or another language during a conversation in Lithuanian. For example, instead of the word “diplomatinis” (diplomatic), a fragment like “дипломатinis” may appear. This error is not related to user settings.
According to Asta Bagdonavičienė, Head of Artificial Intelligence, Data, and Analytics at Telia, it occurs because large language models are trained on huge multilingual text datasets and predict not words but the most likely next sequence of characters. Sometimes the model “slips” between different language systems and accidentally merges a Lithuanian sentence with Russian, Ukrainian, or another Cyrillic-written language.
“The more we use AI, the better we need to understand its limitations and possible errors. Many people imagine that AI perfectly understands language. In reality, it predicts the statistically most likely continuation of the text. When the model relies on several languages at once, a so-called language switch sometimes occurs, so Cyrillic or foreign words may appear in a Lithuanian sentence,” explains A. Bagdonavičienė.
Errors are inevitable
Another common problem is spelling and grammar mistakes. Although AI usually writes fluently, it is not a professional language editor. Models find it harder to maintain the highest accuracy level in less widely used languages worldwide, such as Lithuanian. For this reason, incorrect cases, confused nasal letters, punctuation errors, or unnaturally sounding sentence constructions occur.
“Sometimes AI creates sentences that seem correct, but Lithuanians do not use such formulations. This happens because models often transfer English language constructions into Lithuanian. For example, AI may suggest ‘realizuoti projektą’ (to realize a project) instead of ‘įgyvendinti’ (to implement it) or ‘turėti pokalbį’ (to have a conversation) instead of ‘pasikalbėti’ (to talk). Such inaccuracies mostly arise because models learn from various internet sources, which contain many incorrect or imprecise formulations,” explains A. Bagdonavičienė.
Read more The new UK government will reduce living costs by cutting energy taxes
She also notes that in longer conversations, AI can lose context: confuse dates, names, or other previously provided information. Moreover, AI app users also face so-called hallucinations. This term describes cases when AI presents fabricated information as fact. The system may create non-existent statistics, assign a person positions they never held, invent a study, or provide a false quote.
This problem remains relevant even for the latest models. A 2025 study published by OpenAI states that large language models often tend to guess when they lack sufficient information because traditional evaluation methods have long encouraged providing an answer rather than admitting ignorance. As a result, the model may choose a confidently sounding but incorrect answer instead of indicating it does not have enough data.
“One of the biggest misconceptions is the belief that the more advanced the model, the fewer mistakes it makes. In practice, AI often encounters situations where information is insufficient or ambiguous. In such cases, it is safest for the user to see the answer ‘I don’t know,’ but due to their programming, models still choose to guess,” explains A. Bagdonavičienė.
According to her, AI is a very fast and productive assistant, but people should not blindly trust the technology. The expert recommends viewing AI-generated text as a first draft, not a final result. Before publishing or sending important information, it is necessary to check facts, dates, statistical data, and quotes, review the text additionally, and ensure there are no foreign language characters or spelling errors.
“Just as we critically evaluate any information on the internet, in books, magazines, or other media, so should we with AI applications. We must understand that AI is not and will not be an infallible expert. The greatest value can be obtained only by combining the speed of technology with our critical thinking and responsibility,” adds A. Bagdonavičienė.
Read more Nicaraguan president shot straight and clear: «There will be no more elections here»
