“It's all AI slop.” The dogma against using AI, and what the data says

Whenever it comes up on social media that someone leaned on AI — to learn something, to look something up, to make anything at all, or someone asks how to learn a thing and gets pointed at ChatGPT — the comments start piling up: "just don't learn it from AI, it makes stuff up". Someone generates a graphic for their CV or a job ad and gets buried under "AI slop". Every time just as fast, and every time regardless of what the tool was actually used for.

That you can't trust it without checking, agreed. But that's true of every source. A twenty-year-old tutorial on someone's personal site may be long out of date, a forum answer was written by someone who was also just guessing, and a classmate's notes contain mistakes they don't know about themselves. None of that means you shouldn't use those sources. It means you should verify.

Yes, part of the resentment has a real core — it just never sits in the tool. An ugly generated ad blasted across every group is annoying. But a person can produce an equally ugly graphic by hand and blast it around exactly the same way. A property listing where an AI visualization stands in for a photo, or where the actual condition has been smoothed over, is a genuine problem — when the buyer's decision hinges on the real state of the thing, that's deception, and personally it's probably the case that bothers me most of all. But it would work just as well with Photoshop, and the blame belongs to the advertiser, not the generator. And the other way round: a graphic for a CV or a job ad that gets the information across and looks decent is fine. Why should anyone spend hours on it or pay a designer?

It's also worth noting who tends to be in that dismissive camp. I don't know a single person who knows how to use AI sensibly and at the same time dismisses it wholesale. People who actually work with it have specific objections: it gets this wrong, there's no point asking it for that, here I have to check every line. Blanket "don't use it" comes almost exclusively from people who have never properly tried it and are repeating an opinion they picked up from the news or from social media. That isn't an argument against them by itself — just a reason to ask what that radical position of theirs rests on.

The difference is always in who used the tool and what for. The same holds for learning, with one advantage: there it can be backed by data. Which is the rest of this article.

What matters most: how well-worn the topic is

A final programming project, for instance, almost always has the same skeleton. A record-keeping app, a booking system, a simple game, a small database-backed website. Underneath sit a handful of tasks decades old: build a form, save a record, log a user in, read a file, list the entries and handle what happens when the list is empty. The language and framework vary from school to school, but across all of them the material has been written up hundreds of times, correctly answered and explained over and over.

That's exactly the kind of task where language models fail least. Hallucinations aren't spread evenly — they cluster where there's little material: a freshly released library, an obscure API, a specific combination of versions nobody tried before you. You'll meet none of that in a school project.

The hardest number anyone can throw at AI in programming comes from a CHI 2024 study that had ChatGPT answer 517 Stack Overflow questions — 52% of the answers were factually wrong. Before anyone uses that as an argument, it's worth knowing what was measured: this was GPT-3.5 in 2023, and Stack Overflow questions arise precisely where the asker couldn't work it out alone. It's the hardest possible sample, not a cross-section of what a student needs.

That doesn't drop the risk to zero, so one hard rule applies: code you haven't run doesn't exist. On a school project that means hitting a button and waiting a moment.

Where the critics are right

There is one serious argument against, and it's backed by data. A study of nearly a thousand high-school students in Turkey, published in PNAS, split students into three groups: a plain GPT-4 chat, a tutor version with pedagogical guardrails, and a control group with nothing.

During practice, both AI groups did far better than the control group — 48% better for the plain chat, 127% for the tutor version. Then came a test with no AI access, and the plain-chat group scored 17% worse than students who never had AI at all. They had turned it into a crutch: generate the solution, feel like you understand it, end up learning less than if you'd never had it.

The important detail: with the guarded tutor version, that drop essentially disappeared. So the study doesn't say "AI harms learning". It says "AI used as an answering machine harms learning".

Where they're wrong

The other side of the coin is a randomized study in Scientific Reports, where 194 Harvard physics students covered the same material two ways — once in an active-learning class with an instructor, once at home with a custom AI tutor. Every student went through both. With the AI they learned more than twice as much in less time, and reported feeling more engaged and more motivated.

That tutor was instructed to keep answers short, reveal one step at a time, and make the student think first. Precisely the property that separated the tutor version from the plain chat in the Turkish study — the group whose scores held up from the one that ended up worse off.

So the difference between the two studies isn't AI versus no AI. It's how it's used — the same dividing line that separates a decent graphic on a CV from a smoothed-over photo in a property listing.

About those alternatives

That last weakness is real and not worth downplaying. But for a missing bracket, or an error message the student has never seen before, it takes ten seconds to verify by running the thing. And those ten seconds replace an hour of staring helplessly at the screen.

How to set it up so it actually teaches

All of this points to a fairly concrete recipe:

Tell the model not to solve it for you. One line at the start of the conversation: "I'm a student working on my final project. Don't give me finished solutions, give me one hint at a time and let me try it myself." That is exactly what separated the tutor version from the plain chat in the Turkish study.
Run every piece of code. No exceptions, however trivial it looks.
Ask "why", not "how". "Why does this database query have to be parameterized?" teaches more than "write the save function for me".
Get quizzed. Ask for five questions on the topic you just covered and answer them without looking at your notes. You'll find out fast whether you understand it or it merely read nicely.
The school syllabus stays the backbone. AI is a mentor alongside it, not a replacement for it — the exam tests what's in the syllabus.

And one argument you won't find in any study: the project is defended orally in front of a board. Code the student doesn't understand will sink them sooner and more reliably than any hallucination — the first follow-up question about why it's written that way exposes it. Mindless copy-pasting punishes itself here, and that's a fairly reliable pedagogical safeguard.

The rest is on the school. Teachers have to account for AI the same way they had to start accounting for Google, for Wikipedia, and long before that for the calculator. The pattern tends to repeat: ban it first, discover the ban is unenforceable, then adjust the assignments and the examining so the tool stops being a shortcut to the answer. An oral defence is exactly that kind of mechanism — it works regardless of what the student leaned on while writing.

Closing

The advice "don't use AI to study" conflates two different things: using AI instead of thinking, and using it to think. The data condemns the first and supports the second. It's the same conflation as with that generated graphic — a label slapped on the tool instead of an objection to what someone did with it. And the difference between the two is practical: a label teaches nobody anything, while a specific objection does. Don't run someone else's code blind. Don't let it hand you finished solutions. Verify.

Talking a student out of a tool that will explain an error message to them at ten at night, because it got bad press somewhere, isn't caution. It's putting obstacles in their way.

Sources

• Study on GPT-4's effect on high-school learning and the role of guardrails — Bastani et al., PNAS, 2025; experiment with nearly a thousand students in Turkey.
• Randomized trial comparing an AI tutor against in-class active learning — Kestin et al., Scientific Reports 15:17458, June 2025; 194 Harvard students.
• Analysis of the factual correctness of ChatGPT answers to Stack Overflow questions — Kabir et al., CHI 2024; 517 questions, GPT-3.5, data from 2023.