Learning Science
Why You Cannot Judge Your Own Learning
People are not neutral judges of what they've learned. The miscalibration is not random noise — it runs in a consistent, predictable direction, is caused by identifiable mechanisms, and is measurably correctable. None of the correction happens through introspection.
Four mechanisms, four studies
Foresight bias
Koriat & Bjork (2005): judgments of learning are inflated exactly when the answer is visible during study but has to be produced from memory at test. Seeing the answer while studying creates a false sense that retrieving it later will be just as easy.
Stability bias
Koriat, Bjork, Sheffer & Bar (2004) and Kornell & Bjork (2009): predictions of future performance are almost unaffected by whether the test is tomorrow, in a week, or in a year. People do not model their own forgetting curve at all when predicting how well they'll remember something.
The fluency illusion
Roediger & Karpicke (2006, Psychological Science): repeated studyingof the same material raised learners' confidence in what they knew while simultaneously lowering their actual one-week retention, relative to retrieval practice. The material felt more familiar to reread, and that feeling was mistaken for having learned it.
Revealed behavior
Karpicke, Butler & Roediger (2009, Memory) asked students open-endedly how they study. Only 11% named self-testing as a strategy; 84% named rereading. The strategy people actually favor is the one the fluency illusion makes feel most effective — and it is the weaker of the two.
The correction that works — and it isn't willpower
Rhodes & Tauber (2011, Psychological Bulletin, 45 studies, 112 effect sizes, N = 4,554): simply delaying when a judgment of learning is made — asking "will I remember this?" some time after study, rather than immediately — raises the accuracy of that judgment by g = 0.93, nearly a full standard deviation.
What this implies
A judgment made immediately at study time is unreliable regardless of how carefully a learner tries to introspect. The fix is structural — change when the judgment is elicited — not motivational. This is why self-rating scheduling controls (for example, an "Easy / Good / Hard" button pressed immediately after seeing an answer) inherit foresight bias, stability bias, and the fluency illusion wholesale: they ask for exactly the judgment this literature shows is least reliable.
Confidence and hypercorrection
Confidence errors are not symmetric with correctness. Butterfield & Metcalfe (2001, Journal of Experimental Psychology: Learning, Memory, and Cognition, mean within-subject G = .36) show high-confidence learners absorb corrections to their errors particularly efficiently — the "hypercorrection effect." This replicates in classroom settings (Corral & Carpenter 2023, Journal of Applied Research in Memory and Cognition). Metcalfe & Finn (2011) found high-confidence errors preferentially resurface at a one-week delay.
The double edge
The same efficiency that makes hypercorrection useful for fixing confident wrong answers means confident learners will also absorb incorrect corrective feedback efficiently. A confidently wrong correction is not a neutral failure to help — it actively teaches the wrong thing, and teaches it well.