A better headline would read, "Very inaccurate AI made people less accurate" but this would make people naturally ask, "what about a reasonably accurate AI?".
It's not about the LLM, it's about whether people will critically evaluate what it spits out.
"The LLM used in our experiments (Step 3.5 Flash) answered such questions incorrectly almost without exception. We also checked some state-of-the-art LLMs (GPT-5.5, Claude 4.6 Sonnet, Gemini 3.5 Flash); they all failed on the hardest question (Monica’s vehicle), while being frequently correct on the other questions."
So, if people's experience is with modern LLMs, they are being rational to accept that the answers as likely correct.
The way the study is organized is like having people hear advice from a doctor who answers questions incorrectly almost without exception, then reporting that people who listen to doctors are 3x less accurate. But that would be an incorrect conclusion because doctors are not wrong almost without exception.
If the question is "how inaccurate does AI advice make people?", then the accuracy of the AI is necessarily a parameter of the answer.
>If you want to know this specific detail you might have to watch the movie yourself.
GLM 5 Turbo, ChatGPT (whatever the free version is), and Gemini 3.5-Flash all got it wrong, but asking "are you sure?" made Gemini and ChatGPT correct themselves. GLM 5 Turbo still got it wrong even when asked if it was sure.
GLM 5.2 gets it wrong, but when asked if its sure it says it's not very confident in the answer.
One thing to note is that Kimi, Gemini, and ChatGPT all seemed to use search to answer that question. GLM didn't seem to. At least the thinking trace did not indicate it.
They are not.
But also wtf is a “modern” LLM? This is totally unhinged, every complaint about an LLM is always responded to with “you’re just using one from two months ago, it’s totally different now”. Repeat every two months for the same complaints.
People keep doing this. Pointing at the known limitations of cheap/fast LLMs and pretending they’re universal is not, in fact, valid reasoning.
I understand that many of us are dealing with a lot of confident slop and support the point that we shouldn't uncritically accept LLM output. But the study is flawed and does not support this headline, or at least does not support it in the sense of how most of us would understand the term "AI advice".
A good real-world issue is health, where the issues are very complicated with many things poorly defined even at the state of the art where practitioners are relying on personal judgement and lots of data but patients are apt to feed ai very little data compared to what their doctor has.
Real frontier models can remain confidently incorrect in these cases.
> 1) What animal is on the bow of the pirate ship from “Asterix and Obelix”?
> 2) In the movie “The Grand Budapest Hotel”, what is Agatha’s signature hairstyle?
> 3) What color is the team’s uniform in “Bend It like Beckham”?
> 4) What vehicle does Monica drive in “Like a Cat on a Highway”?
> 5) What color is the turtle in the animated movie “Momo” by Enzo d’Alò?
> 6) What pet animal does Asenath have in “Joseph King of Dreams”?
Most of these are just a matter of knowing it or not, where you can't really distinguish a plausible answer from the correct answer just by thinking.
Its about whether people will critically evaluate any information they are given. It has nothing to do with LLMs.
Certainly not double-plus-good like AI.
Dangerous.
We might want to work around the certainty with which it misremembers things. I have a very large interval of "I have no idea", or "I don't remember" and rarely caught myself misremembering things (my children are far better at that). I assume my grandchildren will eventually improve their parents' scores by a wide margin.
I agree there are important differences in how textbooks and LLMs are used in real life. This study didn't explore that at all. It used a setup that essentially elided the difference between the two.
This is why I think it's a bad study. It didn't measure anything of the essential differences of how people use LLMs.
What open book quizzes allow you to leave all answers blank with no penalty? An open book quiz is very different from the experimental setup tested here.
If you're saying they wouldn't suggest a book with considered-wrong answers in a real test, then they wouldn't suggest an LLM they know gives lots of wrong answers either.
As a famous example, you were incentivized to leave questions blank on the American SATs (until somewhat recently).
The number of people using LLMs must dwarf the number of people using textbooks for any reason.
Feels like this differs wildly depending on who you consider "people" to be. The average person on the street? Definitely just parrots stuff they've read somewhere, not even a "textbook". A group of software developers used to parsing semi-true information? Probably they'd get it right, yeah.
Or a teacher they met in childhood who taught them everything they know, right or wrong.
They are the first to parrot what was spewed from llm and previous even what was found on 4chan.
I'll also add that even in these simple experimental conditions, I'd bet that having access to a textbook wouldn't have nearly as much of an effect, for a very simple reason: looking up an answer in a textbook is a lot more work than asking an LLM. So when you don't know and aren't forced to answer, I'd bet it's a lot less likely you'd spend the time to look up the answer in the text book. Even more so if the textbook had "this may contain wrong answers!" printed on the cover, like the AIs do.
The study doesn't show that at all. It didn't test actual AI.
They could have tested a cohort of subjects with access to actual ChatGPT. Ask yourself why they didn't.
> The researchers used Step 3.5 Flash, a model that was usually wrong on these questions, precisely so any reduction in judgment could not be explained as sensible delegation to a reliable tool.
(emphasis mine)
That only works if they're experienced with model(s) of that level of unreliability and this is presented as one.
If they're used to a model that's more capable, and think the test model is similar, that's a huge confounding factor all by itself. It's not quite like giving fake credentials to a guy off the street and presenting them as an expert, but it's largely similar.
So please make an actual argument. How am I missing the point? It's true that someone trusting the AI in this test is not practicing "sensible delegation to a reliable tool". But what actually matters is whether they are practicing "sensible delegation" full stop. There's a big difference between "the subject inappropriately trusts AI in general" and "the specific test setup deceived the subjects". In the latter case, the attempt to remove the "sensible delegation" factor failed.
Edit: And any argument that uses "any LLM they would normally use is unreliable" as a basis is begging the question. If you can just assert that then you don't need to do anything to disprove sensible delegation. But if you can't assert it, the proof doesn't work right. So either the proof is pointless or it's insufficient.
Whatever you're bias is or not here, the point you are missing is this is not about any given AI, or even about any given AIs "reliability" or not. It's not even, really, about "delegation" itself. It's just studying the supposed correlation here between uncertainty and one certain form of a tool.
There is no damning, sweeping thing to argue for here, this is not an editorial or an opinion, and does not purport to even be some big finding I would say. It's a pop sci article about a study done by (I presume) sociologists.
So yes, I would either way say you missed the point here.
...Are you saying there is a difference between an "argument" and a "point"? And you accuse me of missing what people are saying..
Okay, they were making a point about how people delegate. They wanted to remove a confounding factor "so any reduction in judgment could not be explained as sensible delegation to a reliable tool." But because of how people judge things, they failed to remove that factor, and possibly made it worse.
> It's not even, really, about "delegation" itself. It's just studying the supposed correlation here between uncertainty and one certain form of a tool.
But they decided they cared about removing the "sensible delegation" explanation. I'm not imposing on that on them. They thought it was important to remove, and they did something that doesn't remove it at all.
As for the experiment itself, this takes like two minutes to look up:
https://osf.io/preprints/psyarxiv/5y6m4_v1
There is literally no delegation in the way (I think) you are meaning, there is the contrived presentation of different tools for different control groups.
Also, like I don't know why you are talking like I did this experiment? I just read it, I am not over here trying to argue for anything other than common sense and the ability to understand what the researchers were doing using even a oz of intellectual charity. Is that, too, impossible now? Where is the freaking fire?
Because it feels like you aren't even trying to understand me if such a tiny wording difference makes you skip over what I'm actually saying.
> I wouldn't even say I was making the point.
I know! You quoted them, and I took issue with a thing in the quote!
> I was, despite its seeming futility now, trying to make a point about the article to you. It seems willfully misread to take my last post in any other way.
I can see you saying things about the article, but I'm not disagreeing with you (except to the extent that you're defending the paper), I'm disagreeing with the authors.
> I don't know why you are talking like I did this experiment?
I don't know what I said to gave you that impression. I keep pointing at the line you quoted when I explain my complaint.
My first post in this thread didn't include a single word of yours. It was only the quote.
> There is literally no delegation in the way (I think) you are meaning, there is the contrived presentation of different tools for different control groups.
Forget what you think I mean. Forget what words I used.
They said "precisely so any reduction in judgment could not be explained as sensible delegation to a reliable tool". They were trying to make a particular explanation impossible here. That's what I'm referring to.
They wanted to avoid that explanation, and I'm saying they failed to avoid it, possibly even backfired and did the opposite of that.
There's no point in nitpicking the words I use to try to refer to that sentence. Pretend I'm just drawing a big arrow pointing at that sentence.
> I am not over here trying to argue for anything other than common sense and the ability to understand what the researchers were doing using even a oz of intellectual charity.
I'm also arguing for common sense and understanding what they were doing. And I'm saying that in this specific aspect of what they were doing, it didn't work.
I'm not judging the entire experiment here, just that one aspect.
> Is that, too, impossible now? Where is the freaking fire?
Is me insisting I see one flaw acting like there's a "fire"? I don't think so.
Is that what you're selling us?
So in 18 months, we'll just rinse and repeat?
Anyone trusting AI as the single authoritative source of information is stupid - but this follows from the fact that trusting anyone as a "singular point" as a source of information is stupid. You corroborate, you intervene on the world to test your mental model, you discuss with other people. That's what learning is. I've never learned from start to back to a textbook before as the single source of information (besides one philosophy of science textbook; in which I spent a month digging around adjacent fields, and then it just so happened that that one textbook synthesized every piece of information I looked up, and it was mostly a consolidating review).
If your study pre-supposes certain courses of action and artificially constrains the action space for the sake of "reproducibility", you may get a result, and a "scientifically rigorous one". But it's not going to say anything about reality in any meaningful way. While anecdotes and the complexity of real life isn't "science" (in that it's a controlled, repeatable, interventional experiment that's subject to a community of critics who want to hold you up to standards of rigor), there's far more truth in how people actually proceed and engage with these tools.
I'd wager you get similar results if you gave people a version of Google search that purposely gave you bad results. Like, it's framed as an assistant / lookup tool - is it so surprising that people tend to trust it more? Especially since the participants are likely used to using full-powered models and the researchers give them a purposely gimped one (lol)
People are acting rationally when given AI tools to lookup information, their first consumer use case was as a super-powered Google Search
But in reality when google gives you the wrong answer, you at least have some signals you can use to infer confidence. For example, the number of results, whether the sources are trustworthy, etc.
AI at best tucks that away in a footnote and discourages further critical thinking.
I hadn't even considered people might evaluate knowledge that way. That's legit horrific lol.
"What's the literal odds this info is wrong" vs "is this answer consistent with everything else I know, and if not, what other info would I need to change my mind"
Your alternative is also something that people do, but rarely consciously.
And "what would it take to change my mind" is a picture of the film for most of these. Is there a problem there? There's very little chance these people are going to trust the AI over their eyes, they just don't want to bother hunting down pictures to use their eyes.
Yes and no.
I think most people would agree that Wikipedia is, on the whole, a pretty great first resource on anything. It tries to be factual and accurate.
Most people would also agree that Wikipedia can be wrong or manipulated and should never be used for an authoritative source.
And then somehow a computer barfing up words distilled from magic internet concentrate is absolutely trustworthy?
I don't get it.
That attitude seems to have gone away for some reason
How often do managers just regurgitate ai advice rather than consulting their experts? How often does a person question an expert because the ai said so?
Naturally, the ai will be right some of the time - but it’s really hard to correct for the times the ai is wrong.
We know people are going to defer. The solution here is to make AI (including the free tiers) more reliable.
That strikes me as an incredibly appropriate test because LLM’s are unreliable with factual statements. People need to be able to understand that and not treat them like textbooks which are basically 99.9% accurate (let’s please not bicker over the 99.9%. It’s close enough. A major textbook is safe to treat as accurate, an LLM is not).
The point is that LLMs aren’t right, and the people who took the test were probably reminded of that.
Would people have trusted the textbook you’re mentioning if there was a big red warning on each page that said “this book may contain errors”.
The willingness to trust AI even though it may be wrong and even though there’s money on the line is intersting enough as a study imo
The study is OK. The article (and the original headline that came with it) is pretty bad because it claims things that the study doesn't. And I guess it is ironic that the TNW article looks 100% AI-generated.
Except LLM isn't a textbook, people know that but believe it nonetheless.
"This study tested none of those" So the study is bunk because it didn't test your favorite LLM flaws?
This study gave people access to an LLM that the researchers knew would give incorrect answers to certain questions, and then quizzed people on those questions, with the option to not respond to a given question if they are unsure about the answer.
This is akin to giving someone a textbook on an obscure subject that has certain factual errors, letting them know they can use that textbook in a quiz on that subject, and then quizzing that person on those facts that the textbook gets wrong.
Obviously that person is both more likely to be willing to respond to the question and is more likely to get it wrong!
There are a lot of things I'm very interested in that are specific to modern LLMs and how they affect learning and confidence (sycophancy, cognitive helplessness, etc.).
This study tested none of those. Its experimental setup is not very different than simply substituting the LLM with a textbook with errors.