We fall for how LLMs communicate with us and mistakenly attribute human-like expertise to them.

When using AI, we can easily fall into the same trap as we do with traditional internet browsing: putting too much trust in the information we receive. Because of how “conversations” with large language models are structured, we tend to anthropomorphize them and view them as human experts. We subconsciously fall for the speed and sheer confidence of their responses, automatically assuming a higher level of competence and taking their answers as absolute truth.

These cognitive biases are deeply anchored in us by years of social interaction. We cannot completely erase them from our brains, but we can successfully manage them through deliberate actions.

First: Do not fall for the polite and flattering tone that LLMs use by default. Change its settings as soon as possible so it stops nodding along to everything you say and instead provides constructive criticism.

Second: Operate under the assumption that every sentence generated by AI is false until verified-especially those containing dates and statistics.

Third: Stop thinking of AI as a “colleague,” “assistant,” or “expert,” and start treating it like a highly sophisticated version of your phone’s autocomplete.

People used to say, “Don’t trust open encyclopedias because you don’t know who is writing the articles.” Just yesterday, we were warned about fake authorities on social media. Today, a similar narrative surrounds AI because “we don’t know who or what trained it, and what if it’s hallucinating?”

Over time, many have learned to verify internet rumors and accurately assess their consistency and quality. That is why, despite time pressure and information overload, it is crucial to maintain self-discipline and always make room to cross-check every answer generated by LLMs.

In IT, blindly accepting every response from an AI tool can lead to severe security flaws or performance bottlenecks. Let’s remember that these are still just tools meant to augment a specialist’s capabilities, not a 100% replacement for them.

---

Paweł Nejczew