I decided to write this article to try to articulate some ideas from the papers discussed below. An experienced dispatcher looks at a delivery schedule and immediately expects trouble. The schedule appears feasible, yet something about the combination of routes, timing, and workload feels wrong. Should that judgment carry weight? Before answering, it helps to ask how the dispatcher learned it and whether the current situation resembles the cases that produced that experience.
Three readings approach this problem from different directions: Tversky and Kahneman on judgment errors, Kahneman and Klein on intuitive expertise, and Gigerenzer and colleagues on the conditions under which simple strategies work. Reading them together makes the question of when to trust intuition more precise.
In their 1974 paper, Tversky and Kahneman describe how people simplify difficult probability judgments through representativeness, availability, and anchoring. We may judge likelihood by resemblance, estimate frequency through easily recalled examples, or adjust insufficiently from an initial number. These operations can produce systematic errors when the cue used diverges from the quantity being estimated. A convincing description, for example, can draw attention away from the underlying prevalence of a category. Tversky and Kahneman, 1974.
Applied to the dispatcher, this perspective raises a concrete possibility: perhaps a recent, memorable failure makes a similar schedule look unusually dangerous. The feeling may reflect the ease of recalling that episode. To assess it, we would want to compare the memorable case with the broader record. The paper supplies mechanisms to investigate; identifying a possible bias still leaves the question of whether it explains this particular judgment.
Kahneman and Klein examine why experienced professionals can also make effective judgments quickly. They describe skilled intuition as recognition: a familiar pattern brings a plausible interpretation or response to mind. Two conditions matter. The environment must contain sufficiently regular, predictive cues, and the person must have opportunities to learn those relationships through practice and useful feedback. Years of experience provide limited reassurance when outcomes are too noisy, feedback is misleading, or the relevant patterns keep changing. Kahneman and Klein, 2009.
Their account changes the investigation. Has the dispatcher encountered this configuration repeatedly? Were subsequent delays visible, and could their causes be understood? Does the current schedule belong to the same operational setting? Confidence alone cannot answer those questions. Expertise can also be local: someone may recognize an emerging routing problem while being much less skilled at estimating its financial consequences. The conditions make skilled intuition possible; they do not certify every confident judgment.
Gigerenzer, Reb, and Luan bring another distinction into view. A heuristic can be an explicit strategy specifying what information to examine, when to stop searching, and how to choose. Their ecological-rationality perspective asks which environmental conditions allow a strategy to perform well. Ignoring information can sometimes help when additional cues introduce noise, estimation error, or unnecessary cost. The relevant comparison includes the actual information and resources available to competing methods. Gigerenzer, Reb, and Luan, 2022.
For the dispatcher, an explicit rule might be to flag schedules whenever a particular bottleneck exceeds a threshold. This would be a hypothetical strategy to test. Its value would depend on whether that bottleneck predicts the operational problem and whether other information materially improves the decision. Simplicity creates an opportunity for efficiency and transparency, while also creating the possibility of omitting something consequential. Environmental fit has to be examined rather than inferred from a rule’s apparent elegance.
These readings can therefore converse without being forced into a single dispute. Tversky and Kahneman identify ways judgment goes wrong. Kahneman and Klein examine conditions that support learned expertise. Gigerenzer and colleagues evaluate explicit strategies in relation to their environments. A rapid expert impression and a written threshold rule may both simplify a problem, but they differ in how their reasoning can be inspected and tested.
The dispatch example also exposes a remaining difficulty. Suppose a new delivery policy changes the relationship between workload and delays. Historical success would become less informative about current performance. Someone must then judge whether the old expertise or rule still applies. That judgment introduces another uncertainty: recognizing a change can itself be difficult, especially when failures are infrequent or their causes are ambiguous.
For autonomous AI systems, this discussion raises questions rather than settling them. If such a system develops rapid recommendations from patterns in historical data, under what conditions would those recommendations count as learned expertise rather than as the product of misleading correlations or memorable examples? How could the system detect that the environment had changed, that its feedback was unreliable, or that a simple rule no longer fit the situation? And what forms of inspection, testing, or human oversight would be needed before its confidence could be treated as informative? These questions suggest possible implications for the design and evaluation of autonomous systems, and I will explore them further in later articles.