A Crack in the [Universal] Design

This is uncomfortable for me. I’m about to critique something I believe in deeply.

A year ago I wrote that when we build ramps for wheelchairs, everyone ends up using them. Parents with strollers, workers with carts, people nursing a sprained ankle. The ramp built for the edge becomes preferred by everyone. I stand by that and ‘designing for the edges’ is the closest thing I have to a professional creed.

One popular expression of that creed is called Universal Design for Learning. UDL is founded on the metaphor of the curb cut - a dip in a sidewalk with an angled cut in the edge of a curb. Remove one small barrier and benefit everyone.

It’s a beautiful idea and, as I’ve said, one I stand behind. Which is exactly why I needed to look closely at the research supporting UDL.

Here’s the thing about a curb cut: it’s subtractive. The person who doesn’t need it pays nothing for its existence. They don’t even notice it and even if they do, it’s likely a benefit to all.

Now let’s look at how UDL can be seen in the classroom. Multiple formats of the concept (as a video, a text, and a hands-on station), multiple ways students can choose to engage (e.g. a choice board), and multiple options for students to demonstrate what they’ve learned (e.g. an essay, a poster, a new city in Roblox).

The problem is, none of that is subtractive - it’s all extra stuff. And, as we’ve discussed, everything you add to a learning environment draws on the same resources students need to dedicate to learning: working memory. As a refresher, the working memory of an adult has roughly four slots in which information can be stored and manipulated, and it’s even less for the eight-year-old in your classroom.

A curb cut doesn’t consume anyone’s attention, but a choice board introduces extraneous cognitive load. Every option a student has to evaluate, every presentation format they have to consider, is a withdrawal from an already-limited account of cognitive resources.

The metaphor breaks down because physical access is subtractive and free, but cognitive “access,” implemented as options, is additive and expensive. Even the organization that built UDL, CAST, acknowledged early on that they couldn’t directly apply the architectural principles because learning isn’t a building. The metaphor doesn’t hold.

This doesn’t make UDL wrong, but it does make it incomplete. To start with, how empirically valid is UDL?

In 2008, the Higher Education Opportunity Act defined UDL in federal law as “a scientifically valid framework for guiding educational practice.” ESSA imported that definition word for word in 2015.

The first meta-analysis of UDL research was published in 2017.

If the order of those things feels backwards, you’re not alone. Federal law declared the framework scientifically valid without a whole lot of systematic evidence. And when Matthew Capp published that first meta-analysis, his conclusion was carefully split: UDL appears to improve the learning process, but “the impact on educational outcomes has not been demonstrated.” Advocates and skeptics have been quoting their favorite part of that sentence ever since.

The picture has improved since, but only a little. A meta-analysis from King-Sears and colleagues in 2023 found a ‘moderate positive effect’ on achievement across twenty studies. So “UDL doesn't work” isn't a claim this evidence supports. But the studies underlying that meta-analysis are small and uneven, effects were weaker in STEM, and engagement outcomes consistently outperform achievement outcomes, when studied. Further, a longitudinal study of about 1,500 German secondary students, found no relationship between UDL implementation and reading growth and a negative one with math. The authors flag the fact that classrooms with more struggling students use more UDL and those students tend to gain less regardless of instructional practice, so we also can’t say “UDL harms math learning.”

But it also clearly isn’t the big win for UDL we’d hope for. At best, we can say, “the best available data can’t tell us a whole lot about the effectiveness of UDL.”

Why, after all these years, are things so muddy? The UDL research community has named it themselves, but the framework isn’t so easily testable. There are three principles, nine guidelines, thirty-one checkpoints, all combinable in essentially infinite ways.

Two studies “testing UDL” can be testing completely different practices, and often neither measures whether the practices actually happened. Back in 2010, researcher Dave Edyburn asked this question: would you recognize universal design for learning if you saw it? The field still hasn’t come up with a good answer.

So, the problem isn’t UDL is ineffective, we don’t know enough to say that. But there is a very big problem: a framework that can mean anything is very hard to prove wrong. In science, that’s a serious issue.

One more thing, and as an educational neuroscientist I’d be ducking my job if I skipped it.

UDL’s three principles are officially grounded in three brain networks: affective networks for the why of learning, recognition networks for the what, and strategic networks for the how. It sounds rigorous and clever. It is, at best, a loose organizing heuristic - a metaphor. When Gregory Boysen went looking in 2024 for the neuroscience behind the guidelines, he found that none of the cited research was about brain function at all. Not misinterpreted neuroscience, but not actually neuroscience at all.

The failure mode here is different from the neuromyths I usually get on my soap box about. Left brain versus right brain is falsified; the evidence exists and says it’s wrong. The three-networks story is unvalidated. It’s neuroscience worn as a costume, borrowed authority rather than borrowed mechanism. And we know from research on the “seductive allure” of neuroscience that adding brain language to an explanation makes people rate it as more convincing even when the explanation is bad. So the “neuro costume” inflates how much evidence people think there is.

I use brain language constantly, so here’s my own standard: if I invoke a mechanism, I should be able to point to the research on that mechanism. Working memory has decades of it. “Recognition networks explain why we need multiple representations” has none. That’s the difference

Cognitive science has spent fifty years on some of the exact questions UDL raises, and, as usual, it’s not black and white.

In support: Richard Mayer’s multimedia research shows words plus pictures beat words alone, reliably. Multiple representations can genuinely help, but the same research program produced the redundancy principle: presenting the same content in overlapping formats simultaneously ends up reliably hurting learning outcomes. Together, those findings mean “more means of representation” is not automatically better. It depends on whether the added representation carries new information or just duplicates the cognitive load.

Choice research says a similar thing. Katz and Assor’s review found choice helps only when options are few, relevant, and matched to student values and competence. A six-option choice board is not automatically autonomy support, it can instead be working memory overload before students even get to work.

And then there’s the finding I most wish every school leader knew, because it quietly detonates the word “universal.” It’s called the expertise reversal effect, documented across dozens of studies by Slava Kalyuga and colleagues. The expertise reversal effect is the idea that supports that help novice learners (worked examples, integrated explanations, added scaffolds, etc.) become useless and then actively harmful as learners gain knowledge. The expert’s brain has to process the scaffold and reconcile it with what she already knows, and that costs more, cognitively, than the scaffold saves. The same instructional move is a curb cut for one student and just a higher curb for another.

In UDL terms, that means there is no option set that is universal because the value of any support depends on each learner’s current knowledge and capacity. The variability that matters most in your classroom isn’t learning styles (neuromyth), it’s prior knowledge. Luckily, prior knowledge is something you can assess, design for, and watch change.

Here’s where my journey lands.

The curb cut logic is right for a specific learning category: access. Captions, readable fonts, accessible materials, multiple ways to physically reach the content. These subtract barriers at essentially no cost. That part of UDL is the part every classroom should have.

Past access, into instruction, cognitive science has to take over. Which representations to offer, how many choices, which scaffolds and when to fade them: those are cognitive design questions, governed by working memory limits, redundancy effects, and where each learner sits on the novice-to-expert path. UDL supplies the goal: design for real variability instead of the mythical average. Cognitive science supplies the mechanisms, and with them comes something UDL has needed since inception: testability.

“We implemented UDL” is unfalsifiable. “We reduced redundant representations in the novel-content unit and faded scaffolds as students demonstrated mastery” is a claim you can check. The UDL research community knows this, which is why they’ve been building reporting criteria and observation tools to pin down what implementation even means.

Minds aren’t sidewalks or driveways. If we want universal design for learning to live up to its name, we have to stop assuming what works for concrete works for cognition, and we need to start designing for the cognitive variability of our students - both the novice who needs the scaffold, and the expert who needs us to take it away.