Bei der Interaktion mit KI-Tools wie ChatGPT überschätzt jeder – unabhängig vom Kenntnisstand – seine Leistung. Forscher fanden heraus, dass der übliche Dunning-Kruger-Effekt verschwindet und stattdessen KI-kundige Benutzer ein noch größeres Selbstvertrauen in ihre Fähigkeiten zeigen.

I’ve linked to the news release in the post above. In this comment, for those interested, here’s the link to the peer reviewed journal article:

https://www.sciencedirect.com/science/article/abs/pii/S0747563225002262

From the linked article:

Summary: A new study reveals that **when interacting with AI tools like ChatGPT, everyone—regardless of skill level—overestimates their performance. Researchers found that the usual Dunning-Kruger Effect disappears, and instead, AI-literate users show even greater overconfidence in their abilities**.

The study suggests that reliance on AI encourages “cognitive offloading,” where users trust the system’s output without reflection or double-checking. Experts say AI literacy alone isn’t enough; people need platforms that foster metacognition and critical thinking to recognize when they might be wrong.

Key Facts

Reverse Dunning-Kruger: AI-literate users overestimate their abilities more than novices when using ChatGPT.

Cognitive Offloading: Most participants relied on single prompts and trusted AI answers without reflection.

Metacognition Gap: Current AI tools fail to help users evaluate their reasoning or learn from mistakes.

Might this have something to do with LLMs being sycophantic (the classic „You are absolutely right!“ glazing) or perhaps LLMs just being LLMs and not magic (i.e. prone to „hallucinations“ and other issues which will be „fixed soon“)?

I do use LLMs occasionally but only for things where I can easily verify that the LLM is correct.

This shouldn’t surprise anyone.

The big draw of genAI is that it can make you more productive. Checking a giant wall of text for accuracy and content takes longer in general than writing it. So pretty much all AI bros end up trusting the output blindly because it is just more expedient.

When people say „I always check the output“, they are either lying or delusional.

This then translates into atrophy. If you offloading writing to glorified autocorrect, you end up losing your writing skills. Which makes you less able to check the output.

If „AI-literate“ users trust LLM output without reflection they are not „AI-literate“ in my opinion.

Had an intern at work that was trying his best to learn about our industry and what he was seeing everyday. He would come in and talk about studying about a specific piece of equipment and its function the night before. This would’ve been great to show everyone he was ready and willing to learn, the problem was his summaries would be wildly wrong. On top of that he would be so confident he was right that he would argue with people who had been doing the work longer than he’d been alive.

It took me awhile to figure out where he was getting his info until I typed the subject into google and it pulled his answer up on the AI summary.(which was wrong)

This doesn’t surprise me. What I notice through the various trials I’ve read is:

Perceived Performance and productivity goes up. Actual performance and productivity goes down.

Overall experience: Happiness/enjoyment goes up.

Plus all the pressure to ride the wave of boosterism.

I’ve certainly experienced this as well. I argue back and forth with it (Losing track of time doing so) to get the outputs I want – feeling satisfied being “right” (about proving it is talking rubbish about something I _know_ its wrong about and it grovelling/showing its stomach and backing down) or just getting it to “work”. Feels good as I “solved” something or demonstrated my knowledge in not being fooled by the tool. In reality I’ve just wasted a lot of time i wouldn’t otherwise have wasted without the tooling. Similar to a dopamine charged aftermath arguing on the internet.

This is with all the latest LLM enterprise models, and a few in house specialist ones as well.

I’m actually very worried by it being used by novices who have no way (or care) of verifying its outputs or accuracy with low knowledge in the subject/task/domain in question.

I waste a lot of time having to clean up/checking other people’s quickly produced AI work. Already had a few close runs with juniors trying to be lazy in record time.

Dunning-Kruger has been mostly debunked anyway iirc
https://www.sciencedirect.com/science/article/abs/pii/S0160289620300271

I try to look at AI as objectively as possible and avoid hating on it, sometimes I even see how it helps speed up certain tasks. But honestly, what’s hardest for me isn’t AI itself – it’s the people who, with its arrival, suddenly started feeling perfect and superior to everyone else

for example, a colleague who used to come to me for simple advice – after I suggested switching from JS to TS – said he disagreed because “we’d have to maintain encapsulation”, as you understand, he has no clue what oop even is, then he sent me a bullet-point list copied from a chat explaining why

another time, someone told me that YAGNI can’t exist in scrum – and again, just dropped a chat-generated answer

I get that for people like that, AI is really a tool for dealing with something deeper inside. On an emotional level, it helps them artificially compensate for moments when they didn’t know something but their ego wouldn’t let them admit it. That’s why I try not to hate them

Another newcomer, when I advised him to write for people instead of copying for machines, replied: “if it weren’t for AI, I’d argue with you“

To be honest, it’s exhausting – and it eats up a lot of time