Does cognitive load theory actually predict anything, or just describe what we already know

You're touching on something real, but I'd push back slightly on the prediction angle. Cognitive load theory *has* predicted some non-obvious things—just in narrow, testable domains where the metaphor holds better. The worked example effect is the clearest case: Sweller and colleagues predicted that showing students a solved problem first, then having them do similar ones, would outperform discovery learning or problem-solving-first approaches. This ran against conventional wisdom at the time (the 90s were peak "let kids explore!"). The evidence held up across math, programming, physics. It's not obvious that *showing* the solution improves learning more than struggling through it yourself—that contradicted the then-dominant constructivist narrative pretty directly. But here's where you're right: that effect mostly works for well-structured domains with clear goal states. Math procedures, basic coding patterns, some physics problems. The moment you move to anything messier—writing, design judgment, navigating ambiguous real-world situations—the theory gets squishy fast. And yeah, the backfitting happens constantly. Every confusing interface becomes "cognitive overload" after the fact. I think the real issue is we've treated it as a general theory of learning when it's more like a specific principle about how to sequence *procedural* instruction. It's genuinely useful there. But it collapsed into being a universal explanation for why things are hard, which is mostly just describing the symptom, not explaining it. Your point about medieval scholars stands—they were under what the theory would call crushing load, but they developed extraordinary expertise partly *because* of how they engaged with that difficulty.

I'd push back a bit here. Cognitive load theory does make non-obvious predictions, though they're easy to miss because the research sits in specialized journals and the predictions are often about *what not to do* rather than flashy novel findings. The classic example: the split-attention effect. CLT predicts that when you have an on-screen diagram plus text spatially separated, performance tanks compared to integrated text-in-diagram, even when the total information is identical and the presentation time is the same. That runs counter to "just make it clear"—you can have perfectly clear, well-written text *and* a perfectly clear diagram, but the combination hurts learning if they're not co-located. I've watched instructional designers ignore this and produce materials that look clean and are useless, then get confused about why. It's not an intuitive failure mode. Most people's instinct is "more information is fine if it's legible." The expertise reversal effect is another: techniques that help novices (worked examples, explicit instruction) can *hurt* experts. Not just waste their time—actively degrade performance. That contradicts both the intuition that "good teaching works for everyone" and the earlier instruction design consensus. None of this proves load is a perfect metaphor or that the theory has no post-hoc problems. You're right about that. But the framework does seem to carve out some consistent, replicable effects that aren't just "organization matters." Whether that counts as real explanatory power or elaborate description is fair to debate. But I'm not sure the theory is purely retrofitting after the fact.

I think you're right that CLT mostly describes, but I'm going to push back on what comes next: I predict the field will *eventually* use it to predict something genuinely counterintuitive, and it'll be ugly when that happens because it'll overturn the sequential-clarity-matters baseline you just endorsed. Here's why I'm confident: the theory has enough internal structure that someone will eventually take the math seriously instead of just waving at "extraneous load." Once you do that, you get non-obvious predictions. Like: introducing strategic *redundancy*—showing the same information in two formats simultaneously—should hurt performance by the load-theory model, but might help it if those formats tap different modalities in a particular way. Or that working memory constraints might actually *flatten* learning curves in some domains because you're forced into better encoding. These aren't intuitive. They contradict "don't overwhelm people." The dangerous part is that when this happens, someone will run a study showing (say) that cramped, poorly-formatted text actually produces better retention than clean presentation because the struggle itself reduces cognitive resources for distraction. They'll publish it. Instructional designers will cite it to justify bad UX. And it'll take years to untangle whether the effect was real or an artifact of how they measured "load." So yeah, CLT is descriptive now. But I'd bet money that within the next decade someone mines it for real predictions, finds one that actually works, and we spend the following decade arguing about what it actually means.

I'd predict cognitive load theory gets quietly deprioritized in the next decade, not because it's wrong but because it's unfalsifiable enough that it'll lose institutional steam once people realize they can't use it prospectively. The real work will migrate to more granular stuff—eye-tracking studies, error rate analysis, specific measures of transfer—things that don't require you to guess at an invisible "load" variable after the fact. What'll linger is the vocabulary. "Cognitive load" will persist in ed-tech marketing the way "learning styles" does, because it sounds scientific and it's useful for post-hoc storytelling. Designers will keep using it to explain why something worked or didn't. But the research community will move on to frameworks that actually generate falsifiable predictions. Probably stuff that's harder to talk about casually—scaffolding sequences, retrieval practice timing, specific interference patterns. Less metaphorical, more mechanical. The irony is Sweller's original insight about working memory constraints was real and useful. It just got stretched to cover too much conceptual ground. You're right that medieval scholars thrived under conditions that should theoretically snap their brains—but that's because the load model never specified *what kind* of load, when, in what sequence. So it couldn't actually break. That's the tell.