AI's Best Learning Trick: Designing 15% Failure for Faster Mastery

01: When You're Wrong 15% of the Time, You Learn Fastest

A 2019 study from Princeton and Arizona State University published in Nature Communications established the "85% rule": in simple learning models, practice at roughly 85% accuracy yields the fastest learning. If exercises are too easy (near 100% correct), the brain receives no error signal; if too hard (below 60% correct), the signal becomes noisy. The model shows 99% accuracy yields only about one‑quarter of the maximum learning speed, while 60% accuracy yields about 40%. Although human learning isn't this neat, the insight warns that a student getting everything right may simply be working on material that is too easy.

02: One Problem Set Cannot Fit a Whole Class

The optimal 85% threshold varies by individual. Simulating a class with mixed abilities given the same problem set shows correct rates scattered widely; fewer than 20% of students land in the high‑efficiency zone, and the class‑average learning speed drops to roughly 40% of the theoretical maximum. Traditionally, only a full‑time private tutor could continuously adjust difficulty for each learner. AI now makes this personalized calibration available to everyone — this, the author argues, is the real reason we are in the "best era to learn mathematics."

03: Six "Crazy" Learning Methods Built on Designed Failure

All six methods share a premise: the learner must do the work; AI only structures the difficulty and feedback.

Lock in 85%. Prompt the AI: "Adjust difficulty based on my right/wrong answers to keep my accuracy around 85%." A streak of correct answers increases difficulty; two consecutive errors steps it back.

Lose the first round. Ask the AI for a problem that requires knowledge you haven't learned yet. You will fail, but that failure pinpoints exactly what you're missing. When the explanation follows, absorption is much faster. This "productive failure" approach is backed by a meta‑analysis of 53 studies (Kapur) showing that attempting before instruction improves conceptual understanding, especially when well‑designed.

Teach the AI. Have the AI role‑play a middle‑school student with typical misconceptions; you teach it. Stanford research found students exert more effort to teach a "virtual student" than to learn for themselves. Previously this required a real peer; now it's on demand.

Let the AI poison the solution. Ask the AI to embed a single error in a worked solution without telling you where. You find and fix it. Research with elementary students learning decimals showed that the group that studied erroneous examples, located the mistakes, and corrected them outperformed others on delayed tests. Hunting someone else's error forces you to understand why each step holds.

Bet before you check. For each problem, write your answer, then record your confidence level, then see the result. Psychology's "high‑confidence error correction effect" shows that errors made with high confidence are remembered more durably after correction. The surprise of "I was sure but I was wrong" acts as powerful memory glue.

Shuffle everything, mix the practice. Have the AI interleave problems from several chapters without labeling which topic they belong to. An experiment with seventh‑graders found that the interleaved group scored 72% on a two‑week delayed test versus 38% for the blocked‑practice group. The difficulty of interleaving — deciding which method to apply — is exactly what exams test.

For advanced learners, the article suggests going further: use formal proof assistants like Lean to have a machine verify every step of a proof. Terence Tao's collaborative project used this approach to verify over 20 million algebraic property relationships, giving learners an unfatiguing, uncompromising referee.

04: Crazy, but with Boundaries

These methods collapse if the AI becomes an answer machine. A Turkish experiment with nearly 1,000 high‑school students showed that those who simply asked a chatbot for answers scored 17% lower on a closed‑book test. Dosage matters: 15% error is productive; above 50% error signals the need to step back. When a learner's emotional state deteriorates, pause. Parents and teachers must recalibrate expectations: practice accuracy is not a grade — 85% is better than 100%.

Good practice used to be scarce: appropriately difficult problems, well‑crafted erroneous examples, a patient partner willing to iterate. AI has made those abundant. The new scarcity is the courage to willingly fail 15% of the time in front of an AI every day.