Why Getting Stuck Still Matters: The Case for Productive Struggle in STEM Learning

By Andrew B. Raupp, Founder & Executive Director, STEM.org Educational Research™

Short answer: students do not learn STEM by avoiding difficulty. They learn when difficulty is calibrated well enough to force observation, recall, comparison, revision and explanation—without turning confusion into abandonment. The goal is not to make learning harder for its own sake. It is to preserve the parts of learning that require the learner to actually think.

There is a simple mistake adults make when watching a child learn: we confuse smoothness with mastery.

A worksheet completed quickly looks successful. A program that compiles on the first try looks successful. A lab report with polished prose looks successful. A correct answer generated in seconds looks successful.

But learning is not the same thing as producing a clean output.

In STEM, some of the most important intellectual work happens before the answer settles into place: deciding what matters, trying a representation that fails, noticing an inconsistency, retrieving an idea from memory, comparing competing explanations, debugging a mistaken assumption, or revising a model because reality did not cooperate.

Those moments are inefficient by design. They are also where much of the learning lives.

The difference between performance and learning

Learning science has repeatedly found that conditions producing strong immediate performance are not always the conditions producing the strongest long-term understanding.

Robert Bjork’s work on desirable difficulties helped formalize this distinction. Strategies such as spacing practice, retrieving information rather than repeatedly rereading it, and interleaving different kinds of problems can make practice feel harder while improving later retention and transfer. The important point is not that difficulty is inherently beneficial. It is that certain forms of effort require the learner to reconstruct knowledge instead of merely recognizing it.

That distinction becomes especially concrete in research on productive failure. In two randomized studies of mathematics learning, Manu Kapur found that students who attempted complex problems before receiving formal instruction developed stronger conceptual understanding and transfer than students who received instruction first, even though both groups could develop procedural knowledge. A later meta-analysis by Tanmay Sinha and Kapur examined 53 studies and 166 comparisons and reported a moderate overall advantage for problem solving followed by instruction, with stronger effects when the design closely followed productive-failure principles.

This is not an argument for withholding instruction. It is an argument about when instruction becomes most useful.

If a learner first has to notice the shape of a problem, activate prior knowledge, generate possibilities and encounter the limits of those possibilities, subsequent explanation arrives into a mind that already has questions.

Information now has somewhere to attach.

Why this matters more when answers are cheap

The modern learner has access to extraordinary assistance. Search engines, worked examples, automated hints, calculators, code completion and generative AI can collapse the distance between a question and a plausible answer.

That can be enormously useful. It can also erase the very interval in which learning would otherwise occur.

Consider a student writing a short program. There is a meaningful difference between:

• reading an error message, forming a hypothesis, tracing the program state and testing a correction; and
• pasting the entire error and source code into a system that returns a corrected version.

Both paths may end with functioning code. Only one necessarily leaves behind a stronger debugger.

The same distinction appears in mathematics. A student who struggles to decide whether a problem calls for proportional reasoning, a linear model or a geometric relationship is doing categorization work. If a tool identifies the method instantly, the student may still complete the problem while skipping the decision that made the problem educational.

In science, the equivalent moment may be deciding which observation counts as evidence. In engineering, it may be determining which constraint caused a prototype to fail. In statistics, it may be noticing that a summary measure hides the structure of the data.

The educational question, then, is not simply whether a tool gives the correct answer.

It is: Which cognitive act did the tool remove?

Not all struggle is productive

There is a danger in romanticizing difficulty. Confusion alone does not teach. Repeated failure without feedback can produce frustration, guessing or withdrawal. A poorly explained assignment does not become rigorous simply because students cannot understand it.

A 2024 scoping review of research on productive struggle in mathematics found a growing literature, but much of it was qualitative or non-empirical and often focused on teachers rather than students. That is an important warning. “Let them struggle” is not an instructional method.

Useful struggle needs structure.

The learner needs a problem worth solving, enough prior knowledge to enter it, room to make meaningful attempts, feedback that eventually resolves misconceptions, and an opportunity to explain what changed between the first attempt and the final understanding.

The difference between productive and pointless difficulty is whether the effort produces information the learner can use.

A Useful Friction Test for STEM learning

Before removing a difficulty from a lesson, assignment or learning tool, ask five questions.

What this looks like in a real STEM lesson

In mathematics, give students a problem that can be represented in multiple ways before demonstrating the canonical procedure. Then compare the attempted representations and formalize what each reveals.

In coding, require a student to describe the suspected cause of a bug before requesting an automated fix. After the fix, ask the student to explain the failure in plain language and reproduce the correction without assistance.

In engineering, resist improving the first prototype for the student. Ask what the failure tells them about load, material, geometry, tolerance or constraint. Make the unsuccessful version part of the evidence trail.

In science, ask for a prediction before revealing the result. A prediction creates a model that can be tested. Without it, an observation can slide past without forcing revision.

In studying, replace some rereading with retrieval. Close the book. Draw the system from memory. Explain the concept to someone else. Solve a mixed set of problems instead of ten identical ones in a row.

None of these methods are dramatic. That is part of their value.

They return the learner to the work.

The purpose of assistance is capability

A useful tool should change what a learner can eventually do independently.

That standard is more demanding than asking whether a tool makes the task easier. Easier can be good. Calculators free attention from repetitive arithmetic. Simulations make invisible systems observable. AI can generate examples, challenge explanations and help a student interrogate an idea from several directions.

But assistance becomes educationally suspect when it consistently removes the decisions, retrieval, debugging, modeling or explanation the learner was supposed to develop.

This is the deeper reason human agency matters in responsible AI use. The issue is not preserving hardship. It is preserving authorship of thought.

A capable learner should emerge from assistance with more judgment than before.

If the system becomes more capable while the learner becomes less able to proceed without it, something important has been reversed.

We should not optimize learning until nothing is left to learn

Education has always contained a tension between efficiency and formation.

We want to remove needless obstacles. We should remove needless obstacles. But a classroom can become so optimized for completion that it accidentally optimizes away the mental actions that create competence.

The scratch work, false start, imperfect diagram, failed prototype, forgotten formula, wrong hypothesis and stubborn bug are not always evidence that learning has gone badly.

Sometimes they are the visible trace that learning is actually taking place.

The challenge for teachers, parents and learning designers is not to maximize struggle. It is to protect the right kinds of friction long enough for learners to become the people who no longer need the help.

---

Research referenced

• Kapur, M. (2014). Productive Failure in Learning Math. Cognitive Science, 38(5), 1008–1022.
• Sinha, T. & Kapur, M. (2021). When Problem Solving Followed by Instruction Works: Evidence for Productive Failure. Review of Educational Research, 91(5), 761–798.
• Kapur, M., Saba, J. & Roll, I. (2023). Prior math achievement and inventive production predict learning from productive failure. npj Science of Learning, 8, 15.
• Bjork, R. A. & Kroll, J. F. (2015). Desirable Difficulties in Vocabulary Learning. American Journal of Psychology, 128(2), 241–252.
• Young, J. R., Bevan, D. & Sanders, M. (2024). How Productive Is the Productive Struggle? Lessons Learned from a Scoping Review. International Journal of Education in Mathematics, Science and Technology, 12(2), 470–495.