Stop teaching (through) failure

Why “failure as pedagogy” is a comforting yet dangerous myth

“Fail fast.”
“Learn from failure.”
“Failure is a stepping stone to success.”

These slogans have become so familiar in education that they now pass as wisdom. They’re offered as reassurance to students, justification for risky instructional designs, and moral cover for systems that withdraw support too early. If learners fail, the story goes, they’ll reflect, adapt, and come back stronger.

But what if this story is wrong?

Education has a persistent habit of mistaking outcomes for causes or methods. For example if the goal is to solve problems, then we promote – wrongly – problem-solving to the method to learn to solve problems. When something sometimes follows learning, we quietly turn it into a mechanism of learning. Engagement becomes instruction. Struggle becomes depth. And, increasingly, failure becomes pedagogy.

The idea is appealing. If learners fail, they’ll reflect. If they reflect, they’ll learn. From this logic flows a family of instructional approaches that tolerate or even engineer failure early in the learning process. Sometimes this is expressed informally as “learning from failure.” Sometimes it appears in more formal guise, such as productive failure. The common assumption is that failure itself initiates the cognitive work that leads to understanding.

A recent paper by Lauren Eskreis-Winkler and colleagues (The exaggerated benefits of failure) should make us pause before accepting that assumption. Across eleven studies, the authors show that people consistently overestimate how much improvement follows failure. Observers expect those who fail to attend closely to what went wrong, engage with feedback, and correct/change course. Those who actually experience failure do considerably less of this than observers think and believe.

This gap between expectation and reality is essential. Failure-based pedagogical arguments depend on the idea that failure naturally captures attention and motivates learning. The evidence suggests that this is often not the case. This isn’t a small methodological quibble. It’s a direct challenge to a deeply held educational belief: that failure is, by itself, a reliable engine for learning.

The core finding of the study is: We expect failure to work, but it rarely does.

The attention gap

The authors explain this result in terms of an attention gap. Observers believe that failure commands attention, but it doesn’t. Failure is, in addition to being demotivating (I’m dumb. I’m a failure) is threatening. It challenges competence and identity. The common response is not deep processing but disengagement, avoidance, and emotional self-protection. This isn’t a deficit of character or motivation; it’s a predictable human response.

This finding should sound familiar to anyone who has read the feedback literature, cognitive load theory, or research on self-regulated learning. Information is only useful if it’s processed. Feedback only works when learners have the capacity and willingness to attend to it. Failure, especially when it is global and public, tends to reduce both.

From a cognitive load perspective, failure typically occurs when learners are operating at or beyond the limits of their working memory. Novices lack the schemas in long-term memory that would allow them to recognise relevant task features, evaluate solution attempts, or meaningfully interpret feedback. When they fail, the problem isn’t merely that they don’t know the answer; it’s that they don’t yet know what the answer would or should consist of.

Failure under these conditions imposes substantial extraneous cognitive load. Learners need to simultaneously cope with task demands, emotional responses to failure, and poorly structured feedback that they’re not yet equipped to use. Rather than freeing cognitive resources for learning, failure consumes them. Approaches that rely on failure to prepare learners for subsequent instruction or learning assume that learners will extract useful information from their unsuccessful attempts. This is precisely what novices are least able to do! Without sufficient prior knowledge, failure is of little diagnostic value. Learners can’t distinguish between relevant and irrelevant aspects of the task, can’t evaluate their own solution attempts, and can’t easily map feedback onto stable mental representations.

Confusion and misappropriation

This is where the elephant in the (class)room - productive failure - deserves to be mentioned explicitly. The intention behind productive failure is not misguided. It aims to prepare learners for later instruction by activating prior knowledge and exposing gaps in understanding. But this preparation is assumed to occur through failure itself. The Eskreis-Winkler findings highlight why this assumption is risky. Failure doesn’t reliably increase attention to feedback; it often reduces it. When learners lack the schemas needed to interpret what went wrong, failure becomes effortful activity without productive cognitive payoff.

Cognitive load theory suggests why this is unreliable. For novices, failure doesn’t activate useful prior knowledge; it reveals its absence. Without schemas to organise information, learners can’t easily extract structure from their failed attempts. The failure generates cognitive activity, but much of that activity is unproductive; increasing load without increasing learning.

None of this implies that learning should be easy, smooth, or free of challenge. But it does require us to distinguish carefully between failure and difficulty and this is where desirable difficulties is often misappropriated. Desirable difficulties don’t involve asking learners to fail at tasks they’re unprepared to perform. They introduce challenge into successful learning processes. Retrieval practice, spacing, and interleaving are difficult because they require effortful processing of knowledge that already exists in long-term memory. The difficulty strengthens learning because the learner has something to work with.

Failure-based approaches reverse this sequence. They introduce difficulty before there’s understanding. The learner is asked to generate solutions without the conceptual tools needed to evaluate or refine them. The resulting difficulty isn’t desirable because it’s not functional. It doesn’t deepen processing; it overwhelms it.

The distinction is not semantic; it is architectural. Desirable difficulties make learning harder in order to make it stronger. Failure makes learning harder because learning has not yet been made possible. And this distinction matters because it goes to the heart of instructional responsibility. When we believe that failure teaches, we’re tempted to withdraw guidance prematurely. Eskreis-Winkler et al. show that overestimating learning from failure reduces willingness to provide support, while correcting this belief increases it. In education, the same logic applies. If failure is assumed to be formative by default, the burden of learning shifts from instructional design to learner resilience.

But failure is not formative by default. Errors can be informative when they’re small, local, and embedded in instruction that makes their meaning visible. Failure is global. It signals that instruction has already broken down. Teaching is not about exposing learners to what they can’t yet do and hoping that insight will follow. It is about designing conditions under which effort leads to learning. That means ensuring that struggle is intelligible, that feedback is usable, and that difficulty is introduced when learners are cognitively ready to benefit from it.

A little cognitive load theory

Decades of research show that novices learn more effectively from studying worked examples than from attempting to solve problems independently. The reason is not that worked examples make learning easy or passive. It’s that they manage cognitive load. Worked examples reduce extraneous load by making the problem structure explicit. They allow novices to allocate limited working-memory resources to understanding relationships between steps rather than searching blindly for solutions. In doing so, they support schema acquisition which is the very process novices need before problem solving becomes educationally productive.

Failure-based approaches reverse this logic. They ask novices to engage in problem solving before they possess the schemas required to benefit from it. The resulting failure is then assumed to prepare them for later instruction. But without schemas, failure doesn’t clarify structure; it obscures it. Learners don’t know which aspects of their attempt were relevant, which errors were diagnostic, or how feedback should be interpreted. From a worked-example perspective, failure-first approaches remove precisely the support novices need at the moment they need it most.

On the other hand, experts can learn from failure because they possess rich, well-organised schemas. When an expert fails, the failure violates expectations in a structured knowledge system. This mismatch generates informative signals that can be analysed and corrected. Novices, however, don’t have this system. For them, failure doesn’t violate expectations in a meaningful way; it merely confirms uncertainty. Without schemas, there’s little to update. Cognitive load theory predicts that asking novices to fail before instruction increases extraneous load while contributing little to schema construction.

It’s more than a pedagogical problem

Our cultural romance with failure is understandable. It’s comforting because it reassures us that setbacks are secretly productive and that systems needn’t change. But the data tell a less flattering story. Failure doesn’t reliably teach. People don’t naturally attend to it. And when we believe it does, we design environments that make learning harder, not easier. If we care about learning, and by this I mean real, durable learning, then we should stop asking failure to do pedagogical work it’s not equipped to do.

But a more troubling implication of the paper isn’t psychological or pedagogical but rather ethical. When we overestimate how much people learn from failure, we become less willing to support them afterward. Translated into education: If we believe failure teaches, we feel less responsible to teach. We feel justified in withholding guidance and can tell ourselves that struggle is character-building and confusion is productive. Failure-as-pedagogy quietly shifts blame from instructional design to learner resilience, from system constraints to student mindset, and from support structures to personal grit. It allows educators and institutions to say, “They’ll learn from it,” when the evidence suggests many won’t unless we intervene.

What this means for classrooms

Let’s be clear. The authors don’t argue for protecting students from all nor does it deny the value of struggle, challenge, or feedback. Errors are informative when they’re local, bounded, and embedded in instruction that makes their meaning visible. What the paper does is undermine the lazy idea that failure itself is pedagogical. It also doesn’t claim that people never learn from failure. It shows something more important: Designing instruction that depends on failure to do the heavy lifting isn’t principled minimalism; it’s misplaced optimism.

And yes, failure can happen in learning and it’s natural that it does. These are learners/novices and they make mistakes and sometimes fail. But it shouldn’t be the method by which learning is expected to occur. In other words: Stop celebrating failure and start designing learning for success.

Remember: Failure is an outcome. Learning is a design problem. And confusing the two is a mistake we can’t afford.

Eskreis-Winkler, L., Woolley, K., Erensoy, E., & Kim, M. (2024). The exaggerated benefits of failure. Journal of experimental psychology. General, 153(7), 1920–1937. https://doi.org/10.1037/xge0001610