People tend to use less formal language when interacting with AI chatbots compared to humans, which can reduce the accuracy of the AI's understanding of user intent. Researchers at Amazon found that training AI on human-like conversations highlights this gap. Adapting language styles may be key to better chatbot performance.
A study by Fulei Zhang and Zhou Yu at Amazon reveals that users often communicate with AI chatbots in a more casual manner than with human agents. Analyzing conversations, they used the Claude 3.5 Sonnet model to score interactions and discovered that human-to-human exchanges were 14.5 per cent more polite and formal, 5.3 per cent more fluent, and 1.4 per cent more lexically diverse than those with chatbots.
"Users adapt their linguistic style in human-LLM conversations, producing messages that are shorter, more direct, less formal, and grammatically simpler," the authors write in their paper. They attribute this to users viewing large language model (LLM) chatbots as less socially sensitive or capable of nuanced interpretation.
To explore the impact, Zhang and Yu trained the Mistral 7B model on 13,000 real-world human-to-human conversations and tested it on 1,357 messages sent to AI chatbots. The model, annotated with user intents from a limited list, struggled to accurately label intents in chatbot interactions due to the informal style.
Attempts to bridge this using Claude for rewrites yielded mixed results. Rewriting terse messages into human-like prose reduced accuracy by 1.9 per cent, minimal blunt rewrites by 2.6 per cent, and enriched formal versions by 1.8 per cent. However, training Mistral on both minimal and enriched rewrites improved performance by 2.9 per cent.
Noah Giansiracusa at Bentley University in Massachusetts offers a balanced view. "The finding that people communicate differently with chatbots than with other humans is temptingly framed as a shortcoming of the chatbot – but I’d argue that it’s not, that it’s good when people know they are talking with bots and adapt their behaviour accordingly," he says. "I think that’s healthier than obsessively trying to eliminate the gap between human and bot."
The research, detailed in a paper on arXiv (DOI: 10.48550/arXiv.2510.02645), suggests that either users should adopt more formal language or AI training must better accommodate informality to enhance chatbot effectiveness.