Prof. Wojciech Cellary's video "Can you trust what AI says?" has sparked an important discussion on how to use language models safely. The expert explains why even the most convincing artificial intelligence responses can be completely fabricated – and offers tips on how not to fall for digital illusions.
In an era where artificial intelligence helps us write emails, code, or search for information, it is easy to fall into the illusion that there is a thinking entity on the other side of the screen. Prof. Wojciech Cellary's video "Can you trust what AI says?" brutally dismantles this belief. The expert reminds us of a fundamental truth: AI does not think, understand, or analyze facts. It merely generates text based on statistics. This is the key to understanding why the reliability of artificial intelligence depends solely on our vigilance.
Why does AI "lie"?
When we ask a language model a question, the answer is not the result of logical reasoning or searching a knowledge base to find the truth. It is the result of a statistical puzzle. The algorithm predicts which word should come next, based on the massive datasets it was trained on.
In his video, Professor Cellary provides a vivid example of fabricating scientific literature. If we ask a model to prepare a bibliography on a niche topic, AI can generate a list of articles that look extremely professional. We will find names of real researchers, titles of renowned journals, and sensible-sounding topics. The problem is that these publications never existed. The algorithm simply connected elements that statistically often appear next to each other.
This phenomenon is called hallucination. Although technology developers are constantly working to improve it, the problem still persists. Scientific research confirms that even advanced language models regularly generate completely fabricated facts, especially when we ask about detailed or specialized topics.
Where do AI errors come from?
The mechanisms behind artificial intelligence errors stem directly from its architecture. We can identify several main reasons why AI deviates from the truth:
1. Statistics instead of truth
Models learn from texts created by humans – from scientific articles to discussions on internet forums. AI has no built-in moral compass or truth detector. If the training data contained errors, myths, or contradictory information, the model will simply replicate them. For it, what matters is the probability of words appearing next to each other, not their factual accuracy. As the authors of analyses on language models note, these systems can replicate and amplify disinformation present on the web.
2. Knowledge Cutoff
Most language models have a specific point at which their training process ended. This means they do not know about events that occurred after that date. If we ask a model about the latest political events, changes in law, or fresh scientific discoveries, it may either honestly admit its ignorance or try to make something up based on older data.
3. Susceptibility to user suggestion
The way we phrase questions (so-called prompts) has a huge impact on the answer. If we ask a leading question, e.g., by asking for arguments supporting a false theory, AI will easily generate a smooth and convincing argument, even if it contradicts facts. Security analyses of AI systems show that properly formulated queries can easily bypass model safeguards and induce it to generate nonsense or harmful content.
When can you trust AI?
Does this mean we should abandon these tools entirely? Of course not. In his material, Professor Cellary emphasizes that AI is an excellent assistant, provided we know how to use it. The key is to divide tasks into those where this technology shines and those where it becomes downright dangerous.
Applications where AI excels
- Creative support – brainstorming, generating ideas, writing outlines, or alternative versions of headlines.
- Working with text – correcting language errors, changing the style of expression, summarizing long articles, or translations.
- Preliminary analysis – organizing unstructured data and looking for recurring patterns within it.
Situations where AI can harm us
- Health issues – diagnosing diseases based on symptoms without consulting a doctor is a huge risk.
- Legal and financial advice – law and finance require precision and knowledge of current regulations, which a statistical model may not know or may interpret incorrectly.
- Searching for hard facts and sources – without independent verification, the risk of citing non-existent research is very high.
How to verify AI responses?
The most important lesson from Prof. Cellary's video is: responsibility for the word always rests with the human. AI has no legal personality; it is not held accountable for its mistakes in court or before your boss or client.
Here is how to use these tools safely:
- Principle of limited trust – treat every AI response as a draft that requires checking.
- Verify sources at the root – if the model cites a specific law, name, or study, find them in independent databases such as Wikidata or official government and scientific institution websites.
- Use fact-checking tools – in case of questionable information, it is worth using search engines like Google Fact Check Explorer or services like Snopes.
- Compare results – ask the same question to different language models. If their answers differ significantly, it is a clear signal that the topic requires deeper investigation.
- Don't be fooled by the tone – just because AI writes in an extremely confident, professional language does not mean it is right. Style is not proof of factual accuracy.
The future of AI: will it ever be fully reliable?
Will technological development eliminate the problem of hallucinations? Experts are skeptical – as long as models rely on statistical word prediction rather than understanding concepts, the risk of error will always exist.
What can improve the situation?
- Integration with external knowledge bases – systems that search verified databases or the internet in real-time before providing an answer are less likely to fabricate facts.
- Legal regulations – the EU's AI Act imposes greater responsibility on technology developers for transparency and system safety, which forces better control over the quality of training data.
- User education – social awareness is key. We must learn critical thinking when faced with algorithms, just as we once learned not to believe everything written on the internet.
Summary: AI is a tool, not an oracle
Prof. Wojciech Cellary's presentation is a sober voice in a discussion dominated by extreme emotions – from delight to panic. Artificial intelligence is a powerful tool that can incredibly facilitate our lives and work. However, the key to success is treating it like an extremely efficient, but sometimes absent-minded assistant, whose work must always be carefully checked.
If you want to learn more about how AI affects our daily lives, we recommend our previous posts:
- Is AI making us stupid? 2026 research on the impact of artificial intelligence on our brains
- How is artificial intelligence changing science in 2026? A review of the film "Briefing: AI in the service of science"
- What is really hidden in Claude's "mind"? Architecture, limitations, and secrets of the Anthropic model in 2026
Comments