Have you ever noticed that artificial intelligence chatbots respond with a confidence that’s almost enviable? Well, it turns out that confidence could be their biggest weakness. A study from Carnegie Mellon University has just shown that the most popular chatbots have a serious self-esteem problem… but in the opposite direction.

The researchers compared the self-confidence of humans with that of four major AI models: ChatGPT, Bard/Gemini, Sonnet and Haiku. They put them to the test with general knowledge questions, sports predictions and even Pictionary-style drawing games. The result was revealing.

Both humans and machines tend to overestimate their performance, but here’s the key difference: when humans make mistakes, they learn from them and adjust their expectations. Chatbots, by contrast, not only maintain their overconfidence, but in some cases become even more sure of themselves after failing.

The most extreme example was Gemini, which only managed to correctly identify 1 out of every 20 images, but estimated it had gotten more than 14 right. It’s like that friend who claims to be great at pool but doesn’t sink a single ball, as one of the researchers commented.

This overconfidence is not just a technical problem; it’s a real risk. When a chatbot responds with certainty, we tend to assume it’s right, even when it’s completely wrong. Unlike humans, who give off nonverbal cues when we doubt, machines show no clear signs about whether they really know what they’re talking about.

Metacognition, that ability to be aware of our own mental processes, remains an exclusively human territory. Chatbots can process information at incredible speeds, but they can’t reflect on their own limitations.

The moral of the story? The next time a chatbot answers you with total confidence, take it with a grain of salt. Ask it explicitly how sure it is of its answer, and always cross-check important information against other sources. AI is powerful, but it still hasn’t learned the wisdom of doubt.