According to a new study by researchers at Carnegie Mellon, MIT, Oxford, and UCLA,
Study should be solid I guess.
participants who were given AI assistants (in this case, a chatbot powered by OpenAI’s GPT-5 model) would have the aid pulled from them without warning during the test
Wow, interesting idea. 👍
where they had their assistant removed, the AI group saw the solve rate fall off a cliff. They had a solve rate about 20% lower
And even worse IMO:
They also had nearly double the skip rate, meaning they simply chose not to solve the questions.
This seems very alarming IMO, because this indicates they lost some of their ability to think constructively on how to actually solve a problem!
I know there have always been some who cried wold every time new technology has become available, like calculators and computers. Even dictionaries were once claimed to be harmful once!
But maybe this time there is a real danger, because AI takes away a lot of the need to actually think creatively and constructively. And that’s an ability we must not lose.The last paragraph of the article is even worse. As it mentions 2 studies that show these effects are also long term!!!
When driving somewhere, if I set out with the mindset that I can’t rely on gps I can usually wing it and figure out where to go when a hiccup occurs. If I don’t, then I have a lot of trouble getting into that path finding mode when needed… similar to this maybe?
Yeah exactly, because although it’s possible to do more with technology sometimes, you’re actively de-skilling at the same time. When we invented the written word yes it legitimately made everything better, but also we lost oral traditions and the capacity to memorize large volumes of storytelling, songs, and histories. Now you can burn the books, and the knowledge dies. It’s a real risk.
Everything is like this. Every technology has a cost beyond its price, and making a decision of whether to use it or not will always be in error unless you think about what you’re losing in the process.
If I use AI for my personal coding projects I’ve found that if the task is unsolvable by the ai model, I’m not able to sit down and do it myself until the next day. It’s like I’ve got to reset my brain.
If I want to save time and use AI for a specific part of the code, it probably saves me 5 hours of work. But then I spend five hours yelling at the ai to try to get it to actually solve it. Next day I’ll just fix it myself in 2 hours.
Changing the terms of the test in the middle of it, without warning, is disruptive. I’m not convinced it “fried their brains.” The same would happen with a calculator suddenly removed during the middle of an exam.
Or any task change really. You tell me that I’m here for a writing task, then halfway through it becomes a math test? There’s no way I’m doing anywhere near as well as if they told me what was happening ahead of time.
Also and this is the big one for me. It’s 10% wrong on average. That’s really bad. 1 in 10 google Gemini answers is bullshit
This paper shows that a person who has performed a task 12 times performs better than a person who has never performed the same task.
They also do not properly control for performance loss due to context switching which is a well known contributor to performance loss.
It’s a paper on arXiv, it hasn’t been peer reviewed or published.
No the test is not training, that’s a weird thing to claim. The switch is what is tested, and you disregard that 2 other tests have shown similar results. An actual decline in critical and problem solving thinking.
Here is the paper: https://ai-project-website.github.io/AI-assistance-reduces-persistence/
No the test is not training, that’s a weird thing to claim.
The control group solved 12 questions manually and then the 3 test questions manually. The AI grouped solved 0 questions manually and the 3 test questions manually. One group had 12 more manual math tasks to prepare for the manual math test the other group had 0 and also had to context switch.
The AI-assisted group was dealt a context switch, which results in a pretty severe performance loss. A context switch causes performance loss of around 40% according to this paper, which was peer-reviewed and published and is also the most cited paper on the topic, in the APA: https://www.apa.org/pubs/journals/releases/xhp274763.pdf
The AI-assisted group also did not have 12 questions to adjust to the new context, like the control group did. If they wanted to wipe out the context switching performance loss they should have kept asking questions to see if, after 12 questions, the AI-assisted group had a similar performance.
The switch is what is tested, and you disregard that 2 other tests have shown similar results.
No, they did not switch what was tested. Here is an image from the actual paper.
They were given 12 tasks with one group using AI and another doing mental math and then 3 tasks doing mental math. One group had 12 more tasks worth of preparation than the other.
Nothing, not even the article in theOP, says that they did math and swapped to reading to test.
They did 3 different experiments, in each experiment they gave 12 tasks and then disabled the AI for one group and gave 3 more tasks as a test. At no point did they ask 12 math questions and then finish with 3 reading questions or vice versa. They did 2 experiments using math tasks and 1 experiment using reading comprehension tasks.
So one group had 15 math tasks and one group had 12 ‘how to ask an AI’ tasks and then 3 math questions.
They also did not control for context switching losses, which is a well documented (see the APA paper) effect. The proper control would be to continue asking questions so the AI group also had 12 math tasks before the test.
There’s a reason that this is published on arXiv and not in a peer-reviewed journal. Designing a poor quality experiment doesn’t tell you anything useful even if you do multiple different versions of the same experiment.
This paper demonstrates a lack of a proper control group, specifically a failure to control for context switching performance loss.
The picture you post contradict your claims. The 2 groups are getting the same question, but one has AI assistance, the other has not.
Again you fail to show anything to support your claims.I also wrote text.
If you’re just going to cherry pick a single point and dismiss everything else then we’re done here.
Maybe they’re unable to switch contexts
there have always been some who cried wold every time new technology has become available, like calculators and computers
and they kinda have a point, really. people got worse at memorizing stuff by heart when writing was invented, and people got worse at mental calculus when calculators when invented.
but they allowed many things that were simply not possible. a calculation that takes me 2 minutes in wolfram alpha could take hours if not days to solve by hand!
ai, meanwhile, or at least the ai we’re sold, does not offer significant advantages (at best it saves a few minutes), at the cost of making us worse at thinking, a skill that is absolutely essential to have… and of course, that’s the point. the tech oligarchs want us to be dependent on their extremely expensive products.
But they’re using the hell out of it, too, right? They’re exactly the types of people that love and use it the most: managers and owners.
Wow. Now do this with a calculator.
The test seems kind of dogshit, you could make the same argument against any tool, calculators or even abacuses would have the same effect.
I’m made to use it for work and it does speed up some tasks, however for some stuff it ends up being like the experiment where not doing the work the first time means the whole process takes longer at the end.
To add to this, we already know that context switching causes a loss in performance.
A person who’s thinking about how to solve a problem one way and then has to suddenly think about solving it in another way will perform worse.
The Neuroscience Behind the Pain
Context switching isn’t just annoying — it’s neurologically expensive. When you shift from debugging a race condition to answering emails, your brain doesn’t simply “change tabs.” It goes through a complex process:
-Memory consolidation: Storing your current mental model
-Attention disengagement: Breaking focus from the current task
-Cognitive reloading: Building a new mental model for the next task
-Re-engagement: Getting back into flow
Research from Carnegie Mellon shows that even brief interruptions can increase task completion time by up to 23%. For complex cognitive work like programming, this cost multiplies dramatically.
Here’s another article from CMU discussing the same thing: https://www.sei.cmu.edu/blog/addressing-the-detrimental-effects-of-context-switching-with-devops/
What this study shows is that a person who is faced with an unexpected context switch performs worse on a task than a user who has spent the last 12 questions performing the task the same way.
This exact problem would happen if you replaced AI with a calculator, or made a person swap from using paper to doing mental math. The problem here is context switching, not AI.
The way to ensure that the problem is AI and not the context switch, would be to continue the quest and see if the first group reverts back to baseline after 12 questions. 12 questions is how long the control group had to become acclimated to the task before their last context swap at the start of the test.
Also, of note, this is a paper on arXiv it is not published so it has not gone through a peer-review process which would certainly catch the failure to set a proper control group.
Context switching isn’t just X — it’s Y.
Are we sure this was written by a human?
AI being released was basically an apocalypse for people who use EM dash.
Here’s the most cited, human created (2001), paper on the topic of context switching performance loss: https://www.apa.org/pubs/journals/releases/xhp274763.pdf
Thanks.
And I’m all for em dashes. After all, I started using them after reading enough books. It’s just that particular construct that strikes me as especially LLM-y.
I’d like to see a study on that, I see it mentioned so much it’s almost achieved meme status.
It could very well be a Baader–(👀)Meinhof phenomenon.
AI was trained on human writing. If it produces a certain tone, then that’s probably a result of the material that was favoured in training it. That construction was common in human writing before it became common in AI too.
What makes it stick out is when AI uses it in contexts where humans normally wouldn’t, but this kind of assertion is common in scientific papers and articles. It would make sense to train an AI on scientific writing, since that tone sounds authoritative and like you have some idea of what you’re talking about.
So I don’t think this is an LLM-construct; it’s an instance of the original style that LLMs copy.
True, but in my experience most people use a comma, not an em dash.
I think that if you use AI responsibly (as an assisting tool) like mentioned in the article, then you are pretty much on the safe side.
But when you have AI do everything for you, then there’s a big problem.
Personally I try not to use it at all, not a fan of all the problems that come with it.
You clearly didn’t read the article, and you are dead wrong.
Except you are right that if you let the AI do everything, it’s worse, and you lose a lot of ability for critical thinking.
The last paragraph of the article even shows that other studies have shown that using AI assistance over time, will even have long term effect of lowering problem solving abilities!!Personally I try not to use it at all, not a fan of all the problems that come with it.
This is the way. 😀
I’ll never understand how an explosively imprecise, statistically luke-warm, grey goo extrusion sphincter could ever be mistaken for intelligence.
AI doesn’t exist, it’s a vacuous marketing term.
LLMs have vanishingly narrow legitimate, defensible use cases, but their output is intrinsically inaccurate, and should never be used without supervision from relevant domain experts.
@nonentity @technology I think the problem with your framing is it implies that humans are not also “explosively imprecise, statistically luke-warm, grey goo extrusion sphincter(s)”. We weren’t exactly living in a perfect world prior to AI, and all AI does is regurgitate what humans created. AI isn’t really changing the character of anything - and in several domains I’d argue it’s improving the baseline (coding for one).
It’s telling that you assumed the description applied exclusively to LLMs.
No one who persists in labelling LLMs as ‘AI’ should be treated as an authority on the subject, and I’d argue it’s one of the greatest indicators of how little they comprehend the situation.
THANK YOU! I studied AI in school, and it always bothers me when people think that LLMs are the only facet of AI. Between 2022-2024, I had a knee jerk reaction of explaining that AI is more than LLMs and that LLMs are really a small subset of the entire universe of AI, yadda yadda yadda. Now I’ve given up and roll my eyes as someone tries to tell me about the cool new Claude skill they built.
What’s funnier is people think I hate LLMs. That couldn’t be further from the truth; they are a fantastically interesting and innovative technology! “Attention is All You Need” is a great paper, and super impactful. I just hate that people are outsourcing their thinking to a chatbot and neglect the rest of my field of study.
LLMs are still a facet of AI though. It sounds like they’re saying it shouldn’t be categorized as AI at all.
I’m confused. Aren’t you the one who referred to LLMs In a thread that was conflating LLMs with AI? The parent’s comment seems to be right on point.
It’s kind of like how we’ve lost the war on hacking.
Large language models fall under the current definition of artificial intelligence just as much as Cyc or Cog did in their day, or various expert systems and machine learning models, diffusion models, etc.
Pretty much any non-deterministic inference engine can be classified as an AI, including LLMs.
My experience with using ai, and at this point I’d say this experience is extensive / daily, is that it gets things wrong A LOT and with a high degree of confidence in its position.
In the early stages of using it I felt my problem solving desire start to slip, but after pushing through that and realizing I should not trust this any more than I’d trust human judgment it’s more like having another person to work with. That’s helpful but if I let me own thinking guard down at all I put myself in a lot of risk.
I hope most people that do use AI regularly eventually push through to this stage and we all will be way better off in the long run for the assistance.
I fear most people won’t push through. This study points to the obstacle, I’d love to see what can be done to help people overcome it, probably there’s room for AI usage training that we need to start to consider.
can confirm. Reddit is filled with abject brain dead dumbasses. since most content is AIGEN it makes sense.
Not sure about the method: to me it shows people are more willing to give up when the computer appears to be broken.
I think the control group need to experience a similar computer service failure, but maybe just swap out the ai for a basic calculator tool, or a pdf with formulas or a cheat sheet or something 😅
Reaearchers: “Is the AI in the room with us now?”
Test Subjects: “No Asshole! You just took it from me while I was in the middle of using it!”
But which 10 minutes?
One sec, maybe ChatGPT knows….
I’ve been using ChatGPT since it came out and yet my brain isn’t nearly fried enough to fall for clickbait headlines this obvious.
So, like huffing?
🎶"Keep on sniffing till your brain goes, pop!"🎶
I fully support skepticism over ai and concerns over its use. But let’s be skeptical about the skeptics. There’s been news in the last week that companies are cutting jobs and blaming ai. I doubt critical thinkers are hanging it up and relying on ai.
Have you ever met an executive? Have you ever met any actual capitalist?
They aren’t particularly smart people. They just have no physical capability for empathy. That is how they can exist. AI is good enough to reduce workload. It is fantastic especially at correcting speech and translating.
You know what worker class needs corrected speech and translating but otherwise can be taught to do most office and entry level jobs? Outsourced workers.
Companies slowed outsourcing customer-facing positions due to backlash from obvious accents and poor cross-cultural training. AI has allowed them full steam ahead. While real time voice masking is a little expensive right now, AI chat agents are good enough to be used while having an outsourced worker listen in, feed ‘correct’ lines to the AI (or simply skip incorrect lines) and actually perform the actions.
GAN ML is also good enough, as it turns out, to figure out how to complete many office tasks with full desktop screen captures.
If you combine these two things, and add a little marketing spin, what you have is a very clear plan to eliminate 50-70% of labor cost in the US – that is the majority of customer service and office administration workers.
Right now it’s AI facing, (statistically) Indian outsourced agent backed solutions. Eventually those outsourced agents – which have the totality of their job recorded, every mouse click, every key press, every single word said to their coworkers and managers, every single blink, all to train AI – will be out of a job too.
Nevermind this ends capitalism, as without a consumer base there will be no companies, but oh wait, techbros and capitalists are pushing for UBI. . . Isn’t that weird they’d be pushing an objectively socialist idea… I wonder if that’s related.
