Being overconfident about succeeding at hard tasks and underconfident about easy ones — miscalibrated in both directions.
Imagine you have a really hard math test and a really easy spelling test. You'd probably say 'I think I'll do okay' on both. But the truth is you'll bomb the math and ace the spelling. Your confidence doesn't move enough — it stays in the middle no matter how hard or easy the thing actually is, like a thermometer that's stuck.
The hard-easy effect describes a systematic miscalibration of subjective confidence relative to objective task difficulty. When facing hard questions or challenges, people express far more confidence in their answers than their actual accuracy warrants, while on easy questions they report less confidence than their high accuracy deserves. This asymmetry means confidence judgments are insufficiently sensitive to actual difficulty — people anchor toward a moderate level of confidence and fail to adjust adequately in either direction. The effect has been robustly demonstrated in general knowledge quizzes, professional forecasting, and skill-based assessments, though some researchers have argued it can be partially explained by statistical artifacts and item selection methods.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors and analysts tend to set confidence intervals that are too narrow for difficult-to-predict volatile assets while being slightly too conservative on stable, well-understood instruments, leading to systematic miscalibration in risk models and portfolio allocation.
Clinicians may express similar confidence levels when diagnosing rare, ambiguous conditions and common, straightforward ones, leading to under-preparation for complex differential diagnoses and occasionally unnecessary caution on textbook presentations.
Sarah Lichtenstein and Baruch Fischhoff first documented the phenomenon in 1977 through studies on confidence calibration in general knowledge tasks. The term 'discriminability effect' was used by Ferrell and McGoey in 1980, and 'difficulty effect' by Griffin and Tversky in 1992.
In ancestral environments, moderate confidence served as a functional default. Overconfidence on hard challenges could motivate persistence in high-stakes survival tasks like hunting large game, while slight underconfidence on routine tasks promoted careful attention to familiar but still dangerous activities. A coarse-grained confidence signal was computationally cheaper than precise calibration and was 'good enough' for most survival decisions.
Machine learning models trained on human-labeled confidence data can inherit miscalibration patterns, producing probability estimates that are overconfident on hard classification tasks and underconfident on easy ones. Language models may express similar certainty when answering both trivially easy and extremely difficult questions, failing to appropriately hedge on topics where they are likely to be wrong.
This page taught you the name. The deck turns the name into reflex. 1,100+ swipeable scenarios, 1,100+ defenses, 650+ detection prompts — spaced-repetition Swipe Deck, unlimited Spot-the-Bias Quiz, Defense Playbook, Pre-Flight, My Blindspots, Cheat Sheets, Field Guide e-book. $39.53$59.
