The 85% Rule: Why the Best Difficulty Isn't Easy

Hand a class the same worksheet and you guarantee one thing: it will be too easy for some and too hard for others. Decades of research — and a striking result from machine learning — point to a better target. There is a measurable sweet spot of difficulty where learning runs fastest, and it sits closer to the edge of struggle than most teaching dares to go.

What is the right level of difficulty for a learner?

Ask most people what difficulty suits a learner and they will say "somewhere in the middle" — not too easy, not too hard. The intuition is right. The useful version is just more precise. The right level is not a fixed point on a worksheet; it is wherever a particular learner is challenged enough to grow yet successful often enough to keep going. Set work too far below that point and a learner coasts, learning little. Set it too far above and they stall, collecting errors they can't use. The sweet spot sits at the edge of what a learner can almost do on their own. What matters is that this edge is personal: the same question can be trivial for one learner and out of reach for another. So "medium difficulty" only means something once it is measured against the individual in front of it, not against the class as a whole.

What is the 85% Rule?

In 2019, a team led by Robert Wilson published a striking result in Nature Communications: across a broad class of learning systems, learning is fastest when training is answered correctly about 85% of the time — an error rate near 15%.¹ They derived this "Eighty Five Percent Rule" mathematically for common learning algorithms, and showed it held for artificial neural networks and biologically inspired models alike. The intuition is simple. Each attempt teaches most when it is neither guaranteed nor hopeless. At around 85% success, every practice item still carries useful surprise, but not so much that the learner is overwhelmed. Below that, too many failures stop being informative; above it, easy wins add little. The figure is specific to certain training regimes rather than a universal law of every classroom, but it gives a concrete, testable shape to an old intuition — that there is a measurable best level of challenge, and it is harder than "easy".

Why does too easy or too hard slow learning down?

Both extremes waste effort, for opposite reasons. When work is too easy, a learner succeeds without having to change anything. There is little gap between what they did and what they should have done, so there is little to learn from. Boredom follows, and attention drifts. When work is too hard, almost everything is wrong, and the errors don't point anywhere useful: a learner who misses nearly every question can't tell which step broke, only that the whole thing collapsed. Frustration follows, and effort stops. Learning lives in the gap between the two, where mistakes are frequent enough to be informative but rare enough to be diagnosable. That is why the fastest progress comes from a steady diet of problems a learner can mostly handle, with just enough failure to show where to push next. The goal is productive struggle, not struggle for its own sake.

How does this connect to the zone of proximal development?

Long before the mathematics, Lev Vygotsky described the same idea in human terms. His "zone of proximal development" is the band of tasks a learner cannot yet do alone but can manage with support — just beyond current ability, not far beyond it.² Teaching aimed inside that zone advances a learner; teaching aimed below it bores, and above it loses them. The 85% Rule can be read as putting a number on Vygotsky's zone: the difficulty that keeps a learner mostly succeeding is, roughly, the difficulty that keeps them inside it. Robert and Elizabeth Bjork reached a related conclusion from memory research, coining the term "desirable difficulties" for challenges that feel harder in the moment but build stronger, more durable learning.³ Three traditions — developmental psychology, memory science, and machine learning — converge on one practical claim. Effective learning is uncomfortable by design, but only by the right amount.

Why must difficulty be personalised, not fixed?

If the best difficulty is the level at which a given learner mostly succeeds, then a single fixed task cannot be right for a whole class. The same worksheet that keeps one learner at 85% success leaves another at 40% and a third at 99% — one stretched well, one drowning, one asleep. A fixed curriculum, however carefully pitched, can only hit the sweet spot for the learners who happen to sit near its assumed average. Everyone else drifts out of their zone. Keeping each learner near their own sweet spot therefore means adjusting difficulty per person, and over time, because the right level moves as they improve: today's stretch becomes tomorrow's warm-up. This is precisely the problem adaptive learning sets out to solve. Not to make work uniformly easier or harder, but to keep each individual in the narrow band where their effort buys the most learning.

What does this mean for teaching and technology?

For a teacher, the practical takeaway is to aim for productive struggle rather than comfort or punishment — to pitch work so most learners get most of it right, then nudge the difficulty up as they improve. That is hard to do by hand for thirty different learners at once. This is where technology earns its place: a system can watch how each learner is doing and adjust the challenge continuously, keeping them near their own sweet spot without the teacher recalibrating by hand. It is the principle CogniTrace is built around — matching the level of challenge to the individual, so effort is never wasted on work that is too easy or too hard. The aim isn't to make learning effortless. It is to make every bit of effort count, by keeping each learner exactly where the research says they learn fastest.