Language Models and the Evolution of Human-Machine Interaction

This study examines the transformative impact of large language models on human-machine interaction through a mixed-methods design with 120 participants across three experimental conditions: traditional command-based systems, LLM-integrated systems, and transparency-enhanced LLM systems. Quantitative measures included task completion time, error rates, NASA-TLX workload scores, and galvanic skin response monitoring, while qualitative data were collected through semi-structured interviews. Results revealed that LLM integration reduced task completion time by 32% and subjective mental demand by 41%, yet simultaneously increased physiological arousal, suggesting heightened engagement rather than anxiety. Transparency-enhanced systems generated significantly higher user trust and confidence in reliability. However, a 12% hallucination rate in multimodal contexts underscores reliability concerns for high-stakes applications. The findings indicate that sustainable human-LLM collaboration depends not solely on technical efficiency but on transparency mechanisms that enable appropriate trust calibration. This study contributes to reconceptualizing LLM-mediated interaction as a distributed cognitive system where agency is negotiated between human and machine.