Originally published on Edu Global Institute.
Key takeaways
- A 2026 ETH Zurich study of 100 students found computer science knowledge had the greatest effect on AI-assisted programming performance.
- That effect held even after controlling for general cognitive ability.
- Students who used large language models heavily in daily life performed WORSE on both coding tasks and essays.
- The study is correlational, so it cannot establish whether heavy AI use causes weaker performance or merely accompanies it.
- Writing skill also correlated significantly - specifying precisely what you want is a writing problem before a technical one.
- USACO cannot be prompt-engineered away: a correct solution that is too slow scores nothing, which requires understanding complexity.
- Students stall overwhelmingly at Silver, where the obvious approach is too slow and more code does not help.
This guide covers coding. A 2026 study from ETH Zurich produced a result that should change how students think about learning to code.
Researchers took 100 students, all of whom had completed at least an introductory computer science course and had experience programming with AI assistance. They set them tasks using an AI agent: recreate a meal-planning app, add features to a course-organising app, replicate an abstract app. They also measured computer science knowledge, general cognitive ability, and writing skill.
Two findings stand out.
Computer science knowledge had the greatest effect on performance — and it held even after controlling for general cognitive ability. Students who understood program structure wrote more precise instructions and debugged faster.
Students who used large language models heavily in daily life performed worse — on both the coding tasks and the essays.
The study is correlational, so it cannot establish why. The authors offer two possibilities: heavy use may weaken self-expression, or weaker writers may simply lean on AI more. They are careful not to claim causation, and so are we.
But the direction is worth sitting with, because it runs against the assumption that the way to get good at AI-assisted programming is to use AI a great deal.
What this means for a student learning to code
The uncomfortable implication: the thing that makes you effective with an AI assistant is the knowledge you would have needed anyway.
If you understand how a program is structured, you can specify what you want precisely, recognise when the output is wrong, and locate the fault. If you do not, you can generate code that runs and have no way of telling whether it is correct — and no way of fixing it when it is not.
The second finding points the same way. Writing skill correlated significantly with success, because a clear, structured instruction produces better output than a vague one. Specifying precisely what you want is a writing problem before it is a technical one.
So the fundamentals have not become less valuable. They have become the bottleneck.
Why competitive programming still earns its place
This is the question parents ask us most about coding: with AI able to produce working code, is there any point in a child learning algorithms?
The honest answer is that competitive programming trains precisely the thing the ETH study found to matter — and it is almost uniquely resistant to being outsourced.
A USACO problem does not ask you to produce code that runs. It asks you to produce code that runs within a time limit on a large input. A correct solution that is too slow scores nothing. That constraint cannot be prompt-engineered away, because it requires understanding why one approach is quadratic and another is logarithmic.
Look at how the USACO ladder is built and you can see the skill being constructed deliberately:
- Bronze — simulation, sorting, brute force. Can you state the problem precisely enough to solve it at all?
- Silver — binary search, prefix sums, depth-first search, greedy foundations. Can you find a faster route than the obvious one?
- Gold — dynamic programming in force, plus shortest paths, union-find, minimum spanning trees. Can you recognise which known technique a disguised problem reduces to?
- Platinum — advanced modelling, no fixed syllabus, no upper limit on difficulty. Can you build an approach nobody has handed you?
Notice that only the lower rungs are about knowing things. The upper ones are about recognition and construction, which is the part AI assistance does not supply and the ETH study suggests you need in order to use that assistance well.
Where students actually stall
Silver. Almost always Silver, and the reason is consistent.
Bronze rewards careful implementation — read the problem, write the loop, handle the edge cases. A diligent student clears it on care alone.
Silver asks for something different: the obvious approach is too slow, so you must find a better one. As one guide puts it, clean syntax earns nothing if the method is too slow. That is a conceptual shift, not an increment of effort, and students who respond by writing more code rather than thinking harder about complexity stay stuck for a season.
Gold then asks a third thing again — not just knowing the algorithms, but developing the instinct for when to apply them. Knowing dynamic programming and recognising a DP problem in disguise are separate skills, and only the second is tested.
How we teach it now
We have changed our approach over the past year, in two ways.
Complexity before code. Students state the time complexity of their intended approach and check it against the input size before writing anything. Most Silver failures are visible at this stage, and catching them there costs minutes instead of a contest.
Debugging as a taught skill. Reading someone else's broken solution and finding the fault, reasoning about why a program fails on one input and not another, and narrowing a bug by bisection. The ETH finding makes this more important rather than less: if AI-assisted work succeeds on the ability to recognise and locate wrongness, then that recognition is the thing worth building.
We do not ban AI tools. That would be both unenforceable and pointless preparation for the world these students will work in. But the work is structured so that understanding comes first and assistance comes second, which is the order the evidence supports.
The honest summary
We would rather state the limits of this than oversell it. One correlational study of 100 students is not the final word, and it cannot tell us whether heavy AI use causes weaker performance or merely accompanies it.
What it does establish is that the students who performed best were the ones who understood computer science, independent of raw cognitive ability. That is not a claim about AI at all. It is a claim about foundations — and foundations are exactly what a child builds by struggling with a problem that has no shortcut.
That is what competitive programming is for. It has not been made obsolete. If anything, this is the clearest evidence in years that it was never the code that mattered.
Everything above is how we approach coding with our own students, and it is the method behind the results on this site.
Everything above is how we approach coding with our own students, and it is the method behind the results on this site.
Everything above is how we approach coding with our own students, and it is the method behind the results on this site.
Related reading
Frequently asked questions
With AI writing code, is competitive programming still worth learning?
More so, on the current evidence. A 2026 ETH Zurich study of 100 students found computer science knowledge had the greatest effect on how well they performed at AI-assisted programming tasks, even after controlling for general cognitive ability. Competitive programming builds exactly that understanding, and it is resistant to outsourcing because a correct but too-slow solution scores nothing.
Did the study really find heavy AI users performed worse?
Yes, on both the coding tasks and the essays. The authors suggest either that heavy use weakens self-expression or that weaker writers lean on AI more - the study is correlational and cannot distinguish between them. We would not overstate it, but the direction runs against the assumption that using AI a lot is how you get good at using AI.
Why do students get stuck at USACO Silver?
Because Bronze rewards careful implementation while Silver requires a conceptual shift: the obvious approach is too slow, so a better one must be found. Clean syntax earns nothing if the method is too slow. Students who respond by writing more code rather than thinking about complexity can stay stuck for an entire season.
What does the USACO ladder actually teach?
Bronze covers simulation, sorting and brute force. Silver adds binary search, prefix sums, depth-first search and greedy foundations. Gold brings dynamic programming in force alongside shortest paths, union-find and minimum spanning trees. Platinum is advanced modelling with no fixed syllabus. The upper rungs are about recognition and construction rather than knowing things.
Do you let students use AI tools?
We do not ban them - that would be unenforceable and poor preparation for the world they will work in. But the work is structured so understanding comes first and assistance second. Students state the time complexity of an intended approach and check it against the input size before writing any code.
Why teach debugging separately?
Because if AI-assisted work succeeds on the ability to recognise and locate wrongness, that recognition is the thing worth building. We teach reading someone else's broken solution to find the fault, reasoning about why a program fails on one input and not another, and narrowing a bug by bisection.
What age should a child start competitive programming?
It depends on mathematical maturity rather than age, because the Silver-and-above work is about complexity and reasoning rather than syntax. A student comfortable with logical argument will progress faster than an older one who types quickly. We run a diagnostic before recommending a starting point.
Sources
- TechXplore - What skills do people need to successfully program with AI? (ETH Zurich study)
- USACO Guide - divisions and topic progression
Read the original, with the full formatting and any later updates, on Edu Global Institute.