Since Dunning-Kruger is the relationship between confidence and competence as people learn about new subjects (from discovery to mastery), then if AI is "Dunning-Kruger as a Service" its basically "Education as a service".
However, people accepting incorrect answers because they don't know better is actually something else. Dunning-Kruger doesn't really have anything to do with people being fed and believing falsehoods.
Edit: I had the word "Foolish" in there which was mainly in reference to the OP article about the robbers who didn't hide from cameras because they thought they were invisible. It wasn't meant at a slight against anyone who believed something ChatGPT said that was wrong.
I’m not sure how much I agree with calling them “foolish people.” A big part of the problem is LLMs act incredibly confident with their answers and if you’re asking about something you don’t know a ton about, it’s very difficult sometimes to spot what is incorrect. That’s not being a fool, that’s just not being able to audit everything an LLM might spit back at you. It also doesn’t help that none of these companies are honest about the quality of the results.
We could hand wave this away with “well don’t ask things you don’t already know about,” but these things are basically being pitched as a wholesale replacement for search engines and beyond. I look up things I don’t know about all the time. That’s kind of what we all use search for most days lol.
It’s a little too caveat emptor-adjacent (I hope that makes sense?) for my taste
I’d say there’s a difference between “being a fool” and “being fooled.” Just because I fool you doesn’t mean you’re a fool. I don’t know why you’re so eager to put down people like this rather than at least somewhat acknowledge that these companies and tools bear some of the responsibility.
I don’t think it’s fair to expect every person who uses an LLM to be able to sniff out everything it gets wrong.
But you can expect to learn in both cases. Just like you often learn from your own failures. Learning doesn’t require that you’re given the right answer, just that it’s possible for you to obtain the right answer
We’ve been down this road before. Wikipedia was going to be the knowledge apocalypse. How were you going to be able to trust what you read when anyone can edit it if you don’t already know the truth.
And we learned the limits. Broadly verifiable, non-controversial items are reasonably reliable (or at least no worse than classic encyclopedias). And highly technical or controversial items may have some useful information but you should definitely follow up with the source material. And you probably shouldn’t substitute Wikipedia for seeing a doctor either.
We’ll learn the same boundaries with AI. It will be fine to use for learning in some contexts and awful for learning in others. Maybe we should spend some energy on teaching people how to identify those contexts instead of trying to put the genie back in the bottle.
If you can't discern the difference between a LAMP stack returning UGC and an RNG-seeded matmut across the same UGC fined-tuned by sycophants then I think we're just going to end up disagreeing.
> You can't simultaneously expect people to learn from AI when it's right, and magically recognize when it's wrong.
You are misconstruing the point I was making.
My point is that DK is about developing competence and the relationship between competence and confidence (which I am also claiming evolves over time). My whole point is that the DK effect is not as relevant to LLMs giving wrong answers and people believing them as the author is claiming.
As someone else pointed out in another comment, the effect of people believing falsehoods from LLMs has more to do with Gell-Mann amnesia.
Tangentially, it actually is possible to learn from AI when it's right and recognize when its wrong, but its not magic, its just being patient, checking sources and thinking critically. It's how all of humanity has learned pretty much everything, because most people have been wrong about most things for most of time and yet we still learn from each other.
> Dunning-Kruger is the relationship between confidence and competence as people learn about new subjects
Um, no, it isn't. From the article:
"A cognitive bias, where people with little expertise or ability assume they have superior expertise or ability. This overestimation occurs as a result of the fact that they don’t have enough knowledge to know they don’t have enough knowledge."
In other words, a person suffering from this effect is not trying to learn about a new subject--because they don't even know they need to.
> "A cognitive bias, where people with little expertise or ability assume they have superior expertise or ability. This overestimation occurs as a result of the fact that they don’t have enough knowledge to know they don’t have enough knowledge."
I agree completely. Nothing I said contradicts this.
Every expert starts as a beginner. Not every beginner ends up as an expert.
Also people can learn without a deliberate intent to learn. As far as I am aware right now, deliberate intent has nothing to do with DK, and I certainly wasn't making any claims about it.
> people can learn without a deliberate intent to learn. As far as I am aware right now, deliberate intent has nothing to do with DK
I agree with this: DK's research wasn't about whether anyone is trying to learn anything; AFAIK they didn't even look at that.
However, in the GGP to this post (the post of yours I originally responded to in this subthread), you said:
"Dunning-Kruger is the relationship between confidence and competence as people learn about new subjects."
But DK didn't test for "non-deliberate learning" any more than they tested for deliberate learning. They didn't test anything over time at all. So their research can't tell us anything about what happens over time as people learn (whether they're doing so deliberately or not).
For what it's worth I'm embarrassed by my entire dialog with you. I was having a bad day, sleep deprived and coming on HN to escape. I wasn't thinking clearly or communicating clearly, and on top of that I was straight up wrong about some things that were independent of my state of mind.
On a couple occasions I typed out things that I didn't actually mean. I also was making logic errors, and was embarrassingly biased by a bunch of pop-psychology viral stuff I had seen years earlier.
While my interactions with you were frustrating for me in the moment, I appreciate them now because they triggered a bunch of corrective reflection (including unlearning some things, which is typically a lot less fun than learning). I also want to acknowledge that they were probably frustrating for you too, and I want to thank you for engaging and staying civil despite me acting irrationally.
Hopefully you see this despite the significant delay in response.
However, people accepting incorrect answers because they don't know better is actually something else. Dunning-Kruger doesn't really have anything to do with people being fed and believing falsehoods.
Edit: I had the word "Foolish" in there which was mainly in reference to the OP article about the robbers who didn't hide from cameras because they thought they were invisible. It wasn't meant at a slight against anyone who believed something ChatGPT said that was wrong.