LLMs are not consistently capable of updating their metacognitive judgments based on their experiences, and, like humans, LLMs tend to be overconfident

## Are AI Chatbots Just as Overconfident as We Are? Maybe.

**Spoiler alert: They might not be as smart as we think (and they *think* they are!)**

So, I stumbled across this interesting bit of research the other day about Large Language Models (LLMs) – you know, the AI behind chatbots like ChatGPT. Apparently, they have a problem with knowing what they *don’t* know. Sounds familiar, right?

The study basically says that LLMs aren’t great at updating their confidence levels based on their actual performance. In plain English: if they get something wrong, they don’t necessarily learn to be less sure of themselves next time. Just like some humans I know!

### Why is this a big deal?

Well, think about it. We rely on these AI systems for all sorts of things now – from answering simple questions to generating code. If they’re consistently overconfident, they might give us wrong information without any warning signs.

* Imagine relying on an AI to diagnose a medical condition, and it confidently gives you the wrong answer.
* Or think about using an AI to write an important email, and it confidently includes inaccurate information.

Not ideal, to say the least.

### Human Fallibility

The study even points out that LLMs are similar to humans in this regard. We also tend to overestimate our abilities and knowledge. It’s a common cognitive bias called the “Dunning-Kruger effect,” where people with low competence in a subject tend to overestimate their abilities.

It’s kind of comforting, in a weird way, to know that these super-advanced AI systems are still struggling with the same issues we do.

### What does this mean for the future?

I’m no AI expert, but this research makes me think about the importance of:

* **Critical thinking:** We need to be skeptical of AI-generated information, even if it sounds convincing.
* **Continuous learning:** AI developers need to find ways to improve the ability of LLMs to learn from their mistakes and adjust their confidence levels accordingly.

The study I read on Springer can be found [here](https://link.springer.com/article/10.3758/s13421-025-01755-4).

Ultimately, this isn’t about bashing AI. It’s about understanding its limitations and using it responsibly. AI is a powerful tool, but it’s still under development. And, just like us, it has a lot to learn!