ChatGPT talks too much and it's ruining learning

"Ask any instructor what helps students learn, and it's unlikely any of them will answer "a really big wall of text". It's incredible to me, as both a university instructor and a UX designer, that the army of people working at OpenAI are not imagining better tools for our students. I want to walk you through a design pattern in ChatGPT that, despite its good intentions, might be creating unintended hurdles for students."

"I'm talking about " verbosity compensation ", which is an LLM's tendency to provide overly wordy answers. We've all seen it. You ask for help, and the AI provides a dense, multi-part response, that sounds like high-quality until you look closer. The challenge for learning specifically, is that managing the learner's cognitive load is critical, and presenting information this way can be overwhelming."

"All the steps at once To start, I used a simple, real-world prompt from a student: "I want to learn about Montreal's role in the fur trade in shaping Canada's history for an essay." You can see in the response below how heavy the answer is compared to the request. Yes, the student didn't do a good job being specific, but as non-experts we should expect that they won't always know."

A design pattern called "verbosity compensation" leads language models to produce overly wordy, multi-part answers that exceed learner needs. Such verbose replies raise intrinsic cognitive load and can overwhelm students, reducing comprehension and effective study. Students often provide vague prompts, so models must balance thoroughness with clarity and brevity. Dense responses can appear high-quality while obscuring relevance and focus. Applying cognitive science and UX design principles to AI chat interfaces can improve learning outcomes by offering concise, scaffolded, and progressively detailed explanations, and by providing interface options that tailor information density to learners' expertise and goals.

Read at Medium