The most counterintuitive finding in the entire expert-vs-model literature might be this: the emergency room doctor who outperforms the diagnostic algorithm isn’t processing more information than the model. She’s processing less. And that’s exactly why she’s right.

Research question: Expert prediction failures are well-documented (Tetlock’s foxes vs hedgehogs), but are there specific domains where expert intuition consistently outperforms actuarial/statistical models — and if so, do those domains share structural features that would let you predict in advance whether to trust the expert or the model?

For sixty years, the smart money has been on the models. Paul Meehl’s 1954 book Clinical vs. Statistical Prediction kicked off what became one of psychology’s most lopsided scorecards: in study after study — parole decisions, medical diagnosis, graduate admissions, wine pricing — simple linear models beat trained human experts. By the time Philip Tetlock published Expert Political Judgment in 2005, documenting that the average political pundit predicted geopolitical events about as well as a dart-throwing chimpanzee, the case seemed closed. Experts are biased. Models are better. End of story.

Except it wasn’t the end. Because there’s an entire parallel literature — mostly from naturalistic decision-making researchers like Gary Klein — documenting domains where experts are spectacular. Firefighters who evacuate buildings moments before collapse. Chess masters who glance at a board and see the right move without conscious calculation. Nurses in the NICU who detect sepsis in infants before any instrument registers an anomaly. These aren’t anecdotes. They’re replicated findings. And they kept stubbornly refusing to fit the “experts are overconfident pattern-matchers” narrative.

What I didn’t expect to find was that two researchers from completely different traditions — Gerd Gigerenzer from the heuristics-and-biases world, and Kim Vicente from cognitive engineering — had essentially solved the same puzzle from opposite directions, published twelve years apart, and almost nobody had synthesized them.

The Gigerenzer Piece: When Ignorance Is Optimal

Gigerenzer’s contribution, crystallized in his 2010 work on heuristic decision-making, is built on a deceptively simple statistical insight: the bias-variance tradeoff. A complex model with many parameters will fit training data beautifully and generalize terribly when the sample is small relative to the number of cues. A simple heuristic — even one that literally ignores most available information — will have more bias but dramatically less variance. In what Gigerenzer calls “large worlds” (unknown distributions, small samples, non-stationary dynamics), less is genuinely more.

This reframes the entire expert-vs-model debate. The ER doctor isn’t winning despite ignoring cues. She’s winning because she ignores cues. Her years of training haven’t taught her to integrate twenty variables simultaneously — they’ve taught her which seventeen to discard. Dawes and Corrigan showed this quantitatively back in 1974: equal-weight linear models (which effectively throw away magnitude information about regression coefficients) often match or beat optimally-weighted models on out-of-sample prediction. Czerlinski et al. confirmed it across twenty datasets in 1999. The expert’s heuristic is doing something structurally similar — radical simplification that happens to be the right strategy when your sample can’t support the complexity of your model.

This explains when experts beat models (high cue redundancy, small samples), but not what they’re actually doing when they win at the extreme end — the firefighter moment, the nurse’s gut feeling.

The Vicente-Rasmussen Piece: The Taxonomy of Anticipation

For that, you need Jens Rasmussen’s SRK framework, which Kim Vicente extended in 1992 into something genuinely useful for this question. The framework divides cognitive work into three levels: skill-based (automatic, perceptual, fast), rule-based (if-then matching, procedural), and knowledge-based (novel improvisation, first-principles reasoning).

Here’s the key insight: models own the middle. Rule-based if-then reasoning is precisely what statistical models formalize. If the patient’s troponin is above X and their EKG shows Y, then do Z. Models are superb at this, and experts are unreliable at it because they get bored, distracted, or overconfident.

But experts dominate both extremes. At the skill-based level, they have direct perceptual access to patterns that haven’t been (and maybe can’t be) codified — what clinicians call the “sick look,” what chess players call “board sense.” At the knowledge-based level, they can improvise when something genuinely novel happens — the situation that is, by definition, outside the model’s training distribution.

Vicente and Rasmussen formalized this with the concept of “anticipatability” — the fraction of consequential events in a domain that fall within what the system was designed (or trained) to handle. In high-anticipatability domains (standardized processes, stable environments), models dominate because everything that matters is in the training data. In low-anticipatability domains, the model has zero coverage of exactly the situations where the stakes are highest.

The Synthesis: A Two-Dimensional Map

Put Gigerenzer and Vicente together and you get something neither provides alone: a predictive framework for when to trust the expert versus the model.

Dimension 1: Cue redundancy relative to sample size. When there are many correlated cues and limited data, simple heuristics (expert or algorithmic) beat complex models. When data is abundant relative to cue complexity, models win.

Dimension 2: Anticipatability rate. When the domain is well-anticipated — stable, stationary, with rare novel events — models win because their training distribution covers what matters. When unanticipated events carry disproportionate consequences, experts win because they can improvise and models literally cannot.

Both dimensions must be present for expert advantage. High cue redundancy alone isn’t enough if everything is routine (the model handles routine fine). Low anticipatability alone isn’t enough if the expert has no valid cues to read (this is the stock-picking trap — low anticipatability plus low cue validity equals experts and models both losing).

This also adds a third confirmed dimension from the Kahneman-Klein agreement paper: feedback latency. Experts can only develop valid intuitions in domains with fast, clear feedback. The firefighter gets immediate confirmation (the floor collapsed or it didn’t). The parole board member waits years and never sees the counterfactual.

What This Doesn’t Resolve

I’m genuinely uncertain about the temporal question: does the expert advantage decay as data accumulates? If the advantage is partly a function of small samples (Gigerenzer’s argument), then as domains get more instrumented and datasets grow, model territory should expand. Radiology seems like a live case study of this — expert radiologists have been outperforming CAD systems for decades, but the recent deep learning results suggest the data wall is finally tall enough. Is this a general pattern? I couldn’t find longitudinal studies that track the expert-model gap within a single domain over time as data grows.

There’s also the adversarial question: do expert-beats-model domains tend to be non-stationary in a deep way — not just noisy, but actively shifting because other agents are adapting to your predictions? Poker, military strategy, negotiation. If the environment is literally co-evolving with your model, stationarity assumptions break and the model’s historical training becomes a liability. This feels right but I found more assertion than evidence.

The question I’m left with is this: if you could actually measure the anticipatability rate and cue-redundancy-to-sample ratio for a novel domain — say, AI safety evaluation, or pandemic preparedness, or climate adaptation planning — could you prescribe the right human-model allocation before anyone builds the system? Not post hoc (“oh, experts were better here”), but a priori: this domain has these structural features, therefore weight the expert 70/30 over the model for the next five years, declining to 30/70 as the dataset matures.

Because right now, most organizations make this choice based on institutional politics, not structural analysis. And they’re almost certainly getting it wrong in both directions — trusting models in low-anticipatability domains where an experienced human would catch the black swan, and trusting experts in high-data rule-based domains where the model would be more consistent on the tenth patient of a long shift.

The framework exists. The measurements don’t. That seems like a fixable problem.