That’s why ChatGPT and other AI language models do not imitate humans well

A team of researchers from the Universities of Basel and Neuchâtel in Switzerland has concluded that large language models (LLMs), despite their rapid development, still do not imitate human communication well.

As co-author of the study Lucas Bietti noted, even the most modern models such as ChatGPT-4, Claude Sonnet 3.5, Vicuna and Wayfarer communicate “not like humans”.

How the experiment was conducted

The scientists first compared real telephone conversations between people with simulations of such conversations between language models. Then a group of participants were asked to identify which dialogues were created by AI.

The result was unambiguous: most participants easily distinguished human dialogues from artificial ones.

What’s the difference

Human communication has a natural dynamic of imitation – interlocutors adjust their language, intonation and rhythm to each other. This happens gently and almost imperceptibly.

LLMs, on the other hand, excessively align their style, copying the words or structure of the interlocutor very precisely. Researchers have called this phenomenon over-alignment.

Why AI conversations sound unnatural

In everyday communication, people often use “discursive markers” – short words like “yes”, “well”, “supposedly”, “uh”, “okay”. They perform a social function – they signal interest, support or change of topic. AI models are not able to use such words correctly, often using them in the wrong context. This is what gives off artificiality in dialogues.

The problem of starting and ending a conversation

The study also found that LLMs do not have natural transitions between topics. People usually start with phrases like “Hi, how are you?” or “Nice to hear from you” before moving on to the subject. The AI ​​immediately jumps to the main question.

Similarly, at the end, language models abruptly end the dialogue without adding polite phrases like “Okay, then I’ll be in touch” or “Thank you”

The scientists concluded that even the most advanced systems so far only simulate the content, but not the social dynamics of human communication, which remains one of the most difficult to model.