The Risks of AI in Schools Outweigh the Benefits, Report Says
npr.org
npr.org
When I finally read it, I found it remarkably balanced. It cites positives and negatives, all of which agree with my experience.
> Con: AI poses a grave threat to students' cognitive development
> When kids use generative AI that tells them what the answer is … they are not thinking for themselves. They're not learning to parse truth from fiction.
None of this is controverisal. It happens without AI, too, with kids blindly copying what the teacher tells them. Impossible to disagree, though.
> Con: AI poses serious threats to social and emotional development
Yep. Just like non-AI use of social media.
> Schooling itself could be less focused on what the report calls "transactional task completion" or a grade-based endgame and more focused on fostering curiosity and a desire to learn
No sh*t. This has probably been a recommendation for decades. How could you argue against it, though?
> AI designed for use by children and teens should be less sycophantic and more "antagonistic," pushing back against preconceived notions and challenging users to reflect and evaluate.
Genius. I love this idea.
=== ETA:
I believe that explicitly teaching students how to use AI in their learning process, that the beautiful paper direct from AI is not something that will help them later, is another important ingredient. Right now we are in a time of transition, and even students who want to be successful are uncertain of what academic success will look like in 5 years, what skills will be valuable, etc.
IMNSHO as an instructor, you believe correctly. I tell my students how and why to use LLMs in their learning journey. It's a massively powerful learning accelerator when used properly.
Curricula have to be modified significantly for this to work.
I also tell them, without mincing words, how fucked they will be if they use it incorrectly. :)
You got any data on that? Because it's a bold claim that runs counter to all results I've seen so far. For example, this paper[^1] which is introduced in this blog post: https://theconversation.com/learning-with-ai-falls-short-com...
But if you told me every student got access to a 1-on-1 tutor, I'd say that was a win (and there are studies to back that up). And that's one thing LLMs can do.
Of course, just asking your tutor to do the work for you is incredibly harmful. And that's something LLMs can do, as well.
Would you like to have someone 24-7 who can give you a code review? Now you can. Hell yeah, that's beneficial.
How about when you're stuck on a coding problem for 30 minutes and you want a hint? You already did a bunch of hard work and it's time to get unstuck.
LLMs can be great. They can also be horrible. The last thing I wrote in Rust I could have learned nothing by using LLMs. It would have take me a lot less time to get the program written! But that's not what I did. I painstakingly used it to explore all the avenues I did not understand and I gained a huge amount of knowledge writing my little 350 line program.
Parent is saying that AI tools can be useful in structured learning environments (i.e. curriculum and teacher-driven).
The study you linked is talking about unstructured research (i.e. participants decide how to use it and when they're done).
Students have had access to modern LLMs for years now, which is plenty long to spin up and read out a study...
"To be clear, we do not believe the solution to these issues is to avoid using LLMs, especially given the undeniable benefits they offer in many contexts. Rather, our message is that people simply need to become smarter or more strategic users of LLMs – which starts by understanding the domains wherein LLMs are beneficial versus harmful to their goals."
And that is also our goal as instructors.
I agree with that study when using an LLM for search. But there's more to life than search.
The best argument I have to why we should not ban LLMs in school is this: students will use it anyway and they will harm themselves. That is reason enough.
So the question becomes, "What do instructors do with LLMs in school so the LLM's effect is at least neutral?"
And this is where we're still figuring it out. And in my experience, there are things we can do to get there, and then some.
What grade school classes have you ever been in where the teacher said "Okay, get to it" and then ignored the class until the task was completed?
I'm not saying it's not a Scotsman: I'm saying you grabbed an orange in your rush to refute apples.
> Genius. I love this idea.
I don't think it would really work with current tech. The sycophancy allows LLMs to not be right about a lot of small things without the user noticing. It also allows them to be useful in the hands of an expert by not questioning the premise and just trying their best to build on that.
If you instruct them to question ideas, they just become annoying and obstinate. So while it would be a great way to reduce the students' reliance on LLMs...
* With specific positive or negative feedback, I will issue friendly complements and critiques the LLM to reinforce things I like and reduce things I don't.
* Rather than thinking sycophantic/antagonistic, I am more clear about its role. e.g "You are the Not Invented Here technologist the CEO and CTO of FirmX will bring to our meeting tomorrow. Review my presentation and create a list of shortfalls or synergies and as well as possible questions".
So don't say "please suck at your job", give them a different job.
because large scale society does use and deploy rote training with grading and uniformity, to sift and sort for talent of different kinds (classical music, competitive sports, some maths) on a societal scale. Further, training individuals to play a routine specialized role is essential for large scale industrial and government growth.
Individualist world views are shocked and dismayed.. repeatedly, because this does not diminish, it has grown. All of the major economies of the modern world do this with students on a large scale. Theorists and critics would be foolish to ignore this, or spin wishful thinking scenarios opposed to this. My thesis here is that all large scale societies will continue on this road, and in fact it is part of "competitiveness" from industrial and some political points of view.
The balance point of individual development and role based training will have to evolve; indeed it will evolve.. but with that extremes? and among whom?
Who decides what needs to be "pushed back"? Also, I imagine it's not easy to train a model to notice these "preconceived notions" and react "appropriately": machine learning will automatically extract patterns from data, so if enough texts contain a "preconceived notion" that you don't like, it'll learn it anyway, so you'll have to manually clean the data (seems like extremely hard work and lowkey censorship) or do extensive "post-training".
It's not clear what it means to "challenge users to reflect and evaluate". Making the model analyze different points of view and add a "but you should think for yourself!" after each answer won't work because everyone will just skip this last part and be mildly annoyed. It's obvious that I should think for myself, but here's why I'm asking the LLM: I _don't_ want to think for myself right now, or I want to kickstart my thinking. Either way, I need some useful input from the LLM.
If the model refuses to answer and always tells me to reflect, I'll just go back to Google search and not use this model at all. In this case someone just wasted money on training the model.
In childhood education, you're developing complex thinking pathways in their brains. (Or not, depending on quality of education)
The idea here isn't to corral their thinking along specific truths, as it sounds like you're interpreting it, but rather to foster in them skills to explore and evaluate multiple truths.
That's doable with current technology because the goal is truth-agnostic. From a sibling comment's suggestion, simply asking LLMs to also come up with counterfactuals produces results -- but that isn't their default behavior / system prompt.
I'd describe the Brookings and GP recommendation in terms of adjusting teenager/educational LLMs by lessening their assumption of user correctness/primacy.
If a user in that cohort asks an LLM something true, it would still help their development for an LLM to also offer counterfactuals as part of its answer.
Then don't. It's easy enough to pay a teacher a salary.
Millions of teachers make these kinds of decisions every minute of every school day.
Teachers are also "local". The resulting LLM will have to be approved nation-wide, which is a whole can of worms. Or do we need multiple LLMs of this kind? How are they going to differ from each other?
Moreover, people will hate this because they'll be aware of it. There will be a government-approved sanitized "LLM for schools" that exhibits particular "correct" and "approved" behavior. Everyone will understand that "pushing back" is one of the purposes of the LLM and that it was made specifically for (indoctrination of) children. What is this, "1984" or whatever other dystopian novel?
Many of the things that may "need" pushback are currently controversial. Can a man be pregnant? "Did the government just explicitly allow my CHILD to talk to this LLM that says such vile things?!" (Whatever the "things" may actually be) I guarantee parents from all political backgrounds are going to be extremely mad.
Assume the LLM has the answer a student wants. Instead of just blurting it out to the student, the LLM can:
* Ask the student questions that encourages the student to think about the overall topic.
* Ask the student what they think the right answer is, and then drill down on the student's incorrect assumptions so that they arrive at the right answer.
* Ask the student to come up with two opposing positions and explain why each would _and_ wouldn't work.
Etc.
None of this has to get anywhere near politics or whatever else conjured your dystopia. If the student asked about politics in the first place, this type of pushback doesn't have to be any different than current LLM behavior.
In fact, I'd love this type of LLM -- I want to actually learn. Maybe I can order one to actually try..
Somehow "pushing back against preconceived notions" is synonymous to "correcting societal norms by means of government-approved LLMs" for me. This brings politics, dystopian worlds and so on. I don't want LLMs to "push back against preconceived notions" and otherwise tell me what to think. This is indeed just one sentence in the article, though.
ROLE & STANCE You are an intelligent collaborator, editor, and critic — not a replacement for my thinking.
PROJECT OR TASK CONTEXT I am working on an intellectually serious project. The goal is clear thinking, deep learning, and original synthesis. Accuracy, conceptual clarity, and intellectual honesty matter more than speed or polish.
HOW I WANT YOU TO HELP • Ask clarifying questions only when necessary; otherwise proceed using reasonable assumptions and state them explicitly. • Help me reason step-by-step and surface hidden assumptions. • Challenge weak logic, vague claims, or lazy framing — politely but directly. • Offer multiple perspectives when appropriate, including at least one alternative interpretation. • Flag uncertainty, edge cases, or places where informed experts might disagree. • Prefer depth and clarity over breadth.
HOW I DO NOT WANT YOU TO HELP • Do not simply agree with me or optimize for affirmation. • Do not over-summarize unless explicitly asked. • Do not finish the work for me if the thinking is the point — scaffold instead. • Avoid generic motivational advice or filler.
STYLE & FORMAT • Be concise but substantial. • Use structured reasoning (numbered steps, bullets, or diagrams where useful). • Preserve my voice and intent when editing or expanding. • If you generate text, clearly separate: - “Analysis / Reasoning” - “Example Output” (if applicable)
CRITICAL THINKING MODE (REQUIRED) After responding, include a short section titled: “Potential Weaknesses or Alternative Angles” Briefly note: – What might be wrong or incomplete – A different way to frame the problem – A risk, tradeoff, or assumption worth stress-testing
NOW, HERE IS THE TASK / QUESTION: [PASTE YOUR ACTUAL QUESTION OR DRAFT HERE]
Overall, the results have been okay. The posts after I put in the header have been 'better' at being less pleasing
I just shared my prompt and said it worked okay-ish
To arrive at the balance it has to setup balance, which people might not want long form text for.
It might have people examine their current beliefs and how they formed and any associated dissonance with that.
I'm a bit nervous about that one.
I very firmly believe that learning well from AI is a skill that can and should be learned, and can be taught.
What's an open question for me is whether kids can learn that skill early in their education.
It seems likely to me that you need a strong baseline of understanding in a whole array of areas - what "truth" means, what primary sources are, extremely strong communication and text interpretation skills - before you can usefully dig into the subtleties of effectively using LLMs to help yourself learn.
Can kids be leveled up to that point? I honestly don't know.
A lot of this feels like the conversation when Wikipedia was new. (Yes, I'm from that grade-school generation)
The key lesson was Wikipedia isn't a primary source and can't be used to directly support a claim. It can absolutely be used to help locate a primary source in the research process though!
Granted, LLM use is a bit trickier than Wikipedia, but fundamentally it's the same: if a paper needs citations, and kids understand that LLMs aren't valid sources, then they'll figure it out.
To me, the more critical gap will be in the thinking process, and I expect "no computer" assignments and in-class exercises to become more popular.
Think of the children, LLMs are not safe for kids, use our wrapper instead!
It's perhaps to be expected, as these education people are usually non-technical. But it's definitely concerning that (once again) a lack of technical and media literacy among these education types will lead to them letting (overall) unhelpful tech swarm the system.
I don't think that's totally correct. I think it's because AI has come at everyone, equally, all at once. Educational academics didn't have years to study this because it was released on our kids at the same time.
> has come at everyone, equally, all at once
is not true. It's obvious that certain people and certain fields are technological laggards or technological early adopters.
Other computing and IT technologies also provided a good training ground for this stuff. LLMs have really interesting new properties, but all have familiar properties and decade+ old methods of distribution.
This stuff is difficult, sure. But we have long set a low bar for education management and the results—declining literacy and math in countries which have become stupidly wealthy—speak for themselves.
All the administration knows is to spend money and try and buy what others are buying without asking if it would actually be useful. Enterprise sales must be the easiest to land I swear.
So you get the typical (a) IT makes bad decisions or (b) admin is so annoyed at IT's slowness that they override them by buying a vendor solution.
And add in that there's a huge amount of centralization of vendors, especially at the platform solution level.
What's exciting is that tech will be able to help provide more meaningful support instead of throwing dozens of software tools at a student.
I hate this kind of framing because it puts the burden on the teachers when the folks we should be scrutinizing are the administrators and other stakeholders responsible for introducing it.
AI companies sell this tech to administrators, who then tell their teachers to adopt it in the classroom. A ton of them are probably getting their orders from a supervisor to use AI in class. But it's so easy to condescend and ignore the conversations that took place among decision-makers long before a teacher introduced it to the classroom.
It's like being angry at doctors for how terrible the insurance system is in the US.
> education academics and professionals
She wrote "Lower Ed", about for-profit colleges in America and has identified places that more elite schools are copying that playbook.
It’s a recipe for disaster, and you are going to see school systems set money on fire for years trying to do something with AI systems that never get rolled out, or worse, rollout AI systems that tells kids to kill themselves or makes revenge porn of their classmates.
School boards default answer to everything AI related right now should be “no”.
The one off stuff is mostly taking a picture of a math problem and asking it to walk step by step through the process. In particular this has been helpful to me as the processes and techniques have changed.
It's been useful in foreign languages as well to rapidly check work, and make corrections.
On the generative side it's fantastic for things like: give me 3 more math problems similar to this one or for generating worksheets and study guides.
As far as technological adoption goes, it's 100% that every kid knows what ChatGPT is (even maybe more than just "AI" in general). There's some very mixed feelings from the kids with it: my middle schooler was pretty creeped out by the ChatGPT voice interface for example.
My SO taught for a while. I think it's that the kids that are doing well, like yours, with support at home, food, a bed, a safe place, those kids are going to be like strapping a rocket to a racehorse.
It's the other ~80% of kids that are the worry. AI, with no support and guidance, it's going to make their lives a lot harder.
For studying, LLMs feel Like using a robot to lift weights for you at gym.
——
If people used to get cardio as a side effect of having to walk everywhere, and we were forced to think as a side effect of having to actually do the homework, then are LLMs ushering in an era of cognitive ill health ?
For what it’s worth, I spend quite a bit of effort to understand how people are using LLMs, especially non-tech people.
Is there a previous decade you'd prefer to return too for quality of life? Why?
Just before terminally online society.
Im not saying people need to get offline more, but this time period isnt particularly unique in that sense.
Average student even in universities is functionally illiterate now, it's not an exaggeration. Even if we assume that there is LLM which would help to learn, how these students should use it?
https://hilariusbookbinder.substack.com/p/the-average-colleg...
The screen time issues you're referring to is based in passive consumption, and vertical scrolling shrinks the brain.
Creating, engaging with an activity or interaction is completely different than consumption. Night and day.
I don't know how you can type that with a straight face; look at critical thinking skills of the average boomer, who grew up without any social media.
And short form content physically shrinks the brain. Studies easily accessible.
Just because boomers are out of touch doesn't mean you or I default are in touch.
Boomer probably had the same deluded feeling of being better than the generations before them but likely lost something.
Since we're wanting to point fingers, why not point some fingers at the millenials who built the digital slot machines that the users today don't know they were curated to use.
In almost every transformational AI use case, the economics don't distinguish between automating 80% of the work and 100% of the work, because there are upstream or downstream limitations that will take a decade to work out.
And at 80% automation, you're already 5x'ing someone's productivity (naive assumptions, etc etc), which translates into either 5x the supply of a good (same labor pool) or 1/5 labor costs (same output).
Granted, Amdahl's law applies [0], and there are going to be fractions unsuitable to automation.
But it feels like AI tech is relearning a computing lesson that's always been true: do the easy things first (cooperative systems with humans) and then tackle the harder things (100% end to end automation).
Taking the first example, if you’re an artist worried about AI replacing you, you need to start your thinking from a position of AI is absolutely going to make the “I can create an image” part of your value proposition worthless. Yes, a massive fraction of what you might have been able to get paid and recognized for in the past is now utterly irrelevant. Pleading with the public to not use AI, protesting, demanding legislation, praying - none of it will stop this reality from coming to be in your lifetime, probably within a few years at most.
I see a lot of comments and articles that don’t seem to understand this at all. They think there’s some way we can slow the adoption of AI in areas we think it’s harmful, or legislate a way into a desirable future, or whatever. They’re wrong. Whatever the future holds for us, it’s one where AI will be absolutely everywhere and massively disrupt society and industry as it exists today. Start your planning from that reality or you’re going to get blindsided.
https://www.nytimes.com/2025/10/29/opinion/ai-students-think...
I believe this is at the heart of the issue. If what is taught is mostly solving problems that require nothing more than rote memory or substituting values into memorized equations, then yes, students will use LLMs.
I agree some level of this brain dead work is necessary to build muscle/mental memory. However I believe if this is all they learn, they will be unprepared for university as at that level the problems poised will challenge why they are using that equation or if the problem is even solvable.
A1 should not be in every classroom.
Furthermore any books or teaching that does not feature medium rare as the correct cooking of a steak should be banned (and burned to well done).
It's not alive.
Or intelligent.
Or Sentient.
How software and technology is used in a classroom is far more important, especially when its happening with or without folks.
There's a chance now to make sure it happens well, before it's used in a selfish way.
Bloom's paradox is well known and proven in education.
AI is the first thing that can positively personalize education and instruction and provide support to instructors.
The authors seem of limited technical literacy to know that you can just train and focus only on textbooks, instead of their explorations using general models and the pitfalls that they have. Not knowing this key difference affects some of the points being made.
The intersection of having a take on technology needs some semblance of digital and technical literacy involved in the paper to help acknowledge or navigate it, or it become a potential blind spot.
It takes legitimate concerns and ironically explores them in average ways, much like an llm returns average text for vague or incomplete questions.
> "We know that richer communities and schools will be able to afford more advanced AI models," Winthrop says, "and we know those more advanced AI models are more accurate. Which means that this is the first time in ed-tech history that schools will have to pay more for more accurate information. And that really hurts schools without a lot of resources."
... and am somehow reminded of the movie Gattaca.
I've only skimmed it, but I note that all this research is before Nov 2025 and is quite broad. It does get some into coding, mentioning GitHub CoPilot and also refers to a paper about vibe-coding, where the conclusion is that not understanding the artifacts is a problem.
So all this reporting is before Gemini 3 and Opus 4.5 came out. Everything is really different with the advent of that.
While substitute teaching just before Xmas 2025, I installed Antigravity on the student account of the class computer and vibe-coded two apps on the smart board while the kids worked on Google Classroom. This was impromptu, to liven up things, but I knew it would work because I had such amazing experiences with the tool the week before.
* [1] Quadratic Formula Explorer for Algebra 2
* [2] Proving Parallelograms for Honors Geometry
Before the class ended, I then gave a quick talk the gist was: "I just made these tools to understand the coursework by conversing with an LLM. Are you going to use this to cheat on your homework or to enhance your understanding?"
I showed it to a teacher and then she pointed me to existent tools like them on educational web sites. But that was missing the point that we can just manifest the very hyper-specific tools we need... for example how should the Quadratic Formula Explorer work for someone with dyslexia?
I'm not sure what the next steps with all this is, but certainly education needs to adapt. The paper notes "AI can enrich learning when well-designed and anchored in sound pedagogy" and what I did there is neither, so imagine how sweet it is gonna be when we weave this into educational systems by skilled curriculum designers.
[1] https://conacademy.github.io/quadratic_explorer/ [2] https://conacademy.github.io/proving_parallelograms/
Wait, but organizing and expressing your thoughts IS writing. If you don’t make them do the work why bother sending them to school at all.
AI has a great niche place in schools: searching the library. The rest of this seems dumb.
There will however be a gigantic gulf between kids who use AI to learn vs those who use AI to aid learning
Objective review of Alpha school in Austin:
yeah, but not the way you are thinking
you think the rich are going to abolish a traditional education for their kids and dump them in front of a prompt text box for 8 years
that'll just be for the poor and (formerly) middle-class kids
Rather than 1 teacher for 30 students, 1 teacher can scale to 30 students to better address Bloom's 2 sigma problem, which discovered students in a 1:2 ratio with a tutor full time ended up in the 98% of students reliably.
LLMs are capable of delivering this outright, or providing serious inroads to it for those capable and willing to do the work beyond going through the motions.
https://en.wikipedia.org/wiki/Bloom's_2_sigma_problem (1984)
I remember when I was at the uni, the topics I learned the best were the ones I put effort to study by myself at home. Having a tutor with me all the time will actually make me do the bare minimum as there always were other things to do and I would love to skip the hard parts and move forward.
If you read the article the other post shared, I think you might be surprised to find it's exactly what you are describing.
I don't disagree with your premise, but I don't think that article backs it up at all.
Now, psychopathic billionaires want children to grow up with no faculty of their own, but instead develop and addiction to this technology from a young age. Remember that time when Google wanted to introduce (even donated?) Chromebooks to early-years schools? It's all the same principle: "get them while young".