AI still struggles to mimic natural human conversation
Researchers found AI can be grammatically perfect yet still sound artificial due to poor handling of subtle social cues.
A recent study reveals that large language models such as ChatGPT-4, Claude, Vicuna, and Wayfarer still struggle to replicate natural human conversation. Researchers found AI over-imitates, misuses filler words, and struggles with natural openings and closings, revealing its artificial nature.
The research, led by Eric Mayor with contributions from Lucas Bietti and Adrian Bangerter, compared transcripts of human phone conversations with AI-generated ones. AI can speak correctly, but subtle social cues like timing, phrasing, and discourse markers remain hard to mimic.
Misplaced words such as ‘so’ or ‘well’ and awkward conversation transitions make AI dialogue recognisably non-human. Openings and endings also pose a challenge. Humans naturally engage in small talk or closing phrases such as ‘see you soon’ or ‘alright, then,’ which AI systems often fail to reproduce convincingly.
These gaps in social nuance, researchers argue, prevent large language models from consistently fooling people in conversation tests.
Despite rapid progress, experts caution that AI may never fully capture all elements of human interaction, such as empathy and social timing. Advances may narrow the gap, but key differences will likely remain, keeping AI speech subtly distinguishable from real human dialogue.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
