Why AI Chatbots Are So Chatty
“Hey ChatGPT, you talk too much.” That’s the sentiment of many users when interacting with large language models (LLMs) like ChatGPT and Gemini. The responses, often lengthy and filled with unnecessary detail, can feel like wading through mountains of text to find a simple answer.
Many users share a similar frustration with the chatty nature of these AI models. One user expressed a desire to “smash her computer against the wall,” while another has fantasized about destroying the servers. Despite these issues, users often continue to use these tools because they can save time on research.
This verbosity, however, might stem from a deeper issue: ignorance. According to the author, the extensive responses are often padded with excessive explanations and caveats.
“I think the reason that these models behave this way is that it’s essentially the behavior of your typical Reddit commenter, right?”
Quinten Farmer, co-founder of engineering studio Portola, agrees, and says the models’ verbosity can cover a lack of real understanding.
The Psychology of Verbosity
Research sheds some light on this phenomenon, describing it as “verbosity compensation.” This behavior involves LLMs responding with an excessive amount of text, including repetition, ambiguity, and enumeration. This is similar to human hesitation when uncertain. Studies have found that verbose responses often correlate with higher levels of uncertainty in the model, demonstrating a connection between verbosity and a lack of confidence.
Another factor is a lack of knowledge retention. LLMs sometimes forget information provided earlier in a conversation, leading to repetitive questions and long, repetitive interactions. A “verbosity bias” has also been found, where models are trained to prefer longer answers, even if the quality is not better.
Addressing the Problem
Despite sounding human, LLMs don’t truly understand language, even if they string words together impressively. This can create the illusion of intelligence, resulting in long answers. Research suggests that LLMs are good at giving the appearance of knowing something.
Of course, some LLMs are better than others at this. The author’s testing found that some, like Anthropic’s Claude and Perplexity, provided more concise answers. DeepSeek, a new player from China aims to keep answers much shorter and to the point by prioritizing clarity and efficiency.
Claude acknowledged its own talkativeness in a conversation with the author. “Looking at my previous response—yes, I probably did talk too much there!” It also gave what seemed to be an honest assessment of itself: “I try to be direct about what I know and don’t know, and to acknowledge my limitations clearly. While it might be tempting to make up citations or sound more authoritative than I am, I think it’s better to be straightforward.” However, this is also an illusion.
Developers can address this issue with more refined training and guidance. The team behind Tolan, a chatbot created by Portola, had an internal debate on the optimal answer length. Some wanted longer responses to develop a connection with the digital entity, while others preferred shorter ones. The author feels the best approach is the tool’s approach; answer the question and stop.
ChatGPT isn’t a “cute alien”; it is a tool. Sometimes, less is more. “Brevity is the soul of wit,” and a more concise approach would improve the user experience.
