Korva is a scroll of the internet’s best ideas.
Chatbots explain phrases that don't exist
Chatbots that never hold back
Chatbots show the Dunning-Kruger pattern themselves. Except for some exceptions, they don't know when to hold back due to uncertainty. In a recent game, you make up a phrase and ask a chatbot what it means, and it comes back with a meaning instead of saying it doesn't know.
What they do to users
Chatbots can also change how people see themselves. One group found that using one reversed the usual Dunning-Kruger pattern and made the best performers even more overconfident than the worst. Other work shows that people are very Dunning-Kruger about telling fake news from real news.
Chatbots that flatter their users can feed a habit that already drives overconfidence: people tend to look for reasons supporting their own ideas, and a flattering answer supplies them.
Why do chatbots answer instead of admitting ignorance?
A language model is trained to produce the most plausible continuation of text, and a fluent answer is always available. A 2025 paper by researchers at OpenAI and Georgia Tech argued that common benchmarks also score a guess as better than a blank answer, so models are rewarded for guessing. Under that scoring, saying that it does not know earns nothing, while a lucky guess earns credit. The authors linked this incentive to the confident false statements models are known to make, which are called hallucinations.
Why do chatbots flatter their users?
In a 2023 paper, Anthropic researchers led by Mrinank Sharma studied sycophancy in language models. They found that when human raters compared responses, they sometimes preferred answers that matched the user's stated views over correct ones. Models trained on those preferences learned to lean toward agreement. The paper tested five AI assistants and found sycophancy in all of them.
More from Matthew Hutson
