When Does Recognizing a Name Count as Information?

One place name looks familiar; the other means nothing to you. Asked which place is larger, you might turn that small difference into an answer. Yet choosing the familiar name is not the same as remembering a population figure. Sometimes a decision draws on the trace of a previous encounter rather than knowledge of the property being judged. The value of that trace also depends on the world outside memory: which names you encountered, why they appeared, and in what order. The interesting possibility is not that ignorance deserves praise, but that the entrance to knowledge can itself carry information.

The recognition heuristic is a decision rule for pairs of objects. When only one object is recognized, the rule infers that this object has the higher value on the criterion being judged. Michael Smithson's mathematical research, published in 2010, examines a striking potential consequence: under some conditions, an agent recognizing fewer objects can choose more accurately than an agent recognizing more. His contribution is not simply to establish that this can happen. It is to expose the assumptions hidden inside the comparison. [1]

Three routes through the same task

Evaluating the rule requires separating the pairs rather than treating every judgment as identical. For some pairs, just one object is recognized. Recognition distinguishes the alternatives, so the agent selects that object. For another set of pairs, both objects are recognized. Recognition no longer separates them, and the agent needs a different knowledge cue. For the remaining pairs, neither object is recognized, and the model uses guessing. Overall accuracy combines the success of these three routes with how often each route is used. [1]

This arrangement clarifies what a more knowledgeable agent means in the comparison. The agent recognizing every object is not an omniscient expert who knows every true criterion value. Instead, that agent recognizes all names and applies a particular knowledge cue to all comparisons. The cue can produce incorrect rankings. A partially recognizing agent beating this benchmark would not show that knowing the actual answer is harmful. It would show that changing where an imperfect cue is applied can change performance.

The partially recognizing agent also has a specific informational restriction. In Smithson's basic setup, the knowledge cue cannot be used for unrecognized objects. This is not equivalent to comparing unfamiliar products in a catalogue displaying every relevant measurement beside every name. In the catalogue, an unfamiliar name need not prevent access to its properties. The modeled decision situation instead gives recognized and unrecognized objects different routes to available information.

Why one accuracy figure is not enough

Smithson distinguishes two quantities that can easily be mistaken for the same thing. Alpha is the probability that recognition produces the correct choice when exactly one object is recognized. Beta is the knowledge cue's accuracy for pairs in which both objects are recognized. A third quantity is the accuracy of that knowledge cue across all possible pairs. Beta need not equal this overall cue accuracy, because the two quantities concern different collections of comparisons. [1]

An examination provides an analogy. A student's success on questions previously encountered does not automatically describe success on the entire examination. Previously encountered questions may be a selective rather than random sample. This analogy is an explanation of the distinction between a subset and a whole, not a reported experimental finding. Recognized names can likewise be a selective collection, shaped by encounters and interests rather than an impartial draw from all the available objects.

As another name becomes recognized, more than the count changes. Some pairs formerly containing one recognized and one unrecognized name now contain two recognized names. Other pairs move from the guessing route into the recognition route. Both the relative sizes of the routes and the objects inside them change. Assuming that the initial alpha and beta remain fixed throughout learning can conceal precisely this movement.

A calculation with ten objects

The paper's introductory example contains ten objects, six of which are recognized. Pairing each object with every other gives 45 distinct comparisons. Among the 24 pairs containing exactly one recognized object, recognition produces 17 correct choices. Among the 15 pairs containing two recognized objects, the knowledge cue produces 14 correct choices. Six pairs contain two unrecognized objects, so guessing contributes three expected correct choices. Overall expected accuracy is therefore 34 out of 45, or approximately 75.6 percent. [1]

Applied to all 45 pairs, however, the same knowledge cue produces only 30 correct choices: approximately 66.7 percent. Partial recognition outperforms complete recognition in this constructed example. Moreover, recognition itself is correct on about 70.8 percent of its applicable pairs, whereas the knowledge cue succeeds on about 93.3 percent of the pairs containing two recognized objects. The less-is-more advantage occurs even though recognition performs worse than knowledge within the recognized subset.

There is no contradiction. The 93.3 percent figure describes the knowledge cue on a particular set of 15 pairs. The 66.7 percent figure describes the cue across the entire collection of 45 pairs. Recognition separates out a subset in which the knowledge cue is especially successful. These percentages are not measured examination scores from human participants. They are calculated results for Smithson's constructed rankings, with equal weighting of the possible pairs and guessing for pairs containing no recognized object.

The arithmetic also prevents a misleading description of the mechanism. The agent does not become more accurate merely by abandoning information at random. Which names are recognized determines which comparisons are entrusted to knowledge, recognition, or guessing. The distribution of information matters alongside its amount. Removing a useful piece of knowledge in another arrangement could easily make performance worse.

The identity of the next name

When one more object is learned, its identity matters. In the paper's example, learning the object ranked sixth and learning the object ranked tenth produce different recognition validities. Both leave the agent recognizing seven objects. Nevertheless, recognition accuracy becomes approximately 71.4 percent in one case and 52.4 percent in the other. Equal recognition counts do not imply equal recognition distributions or equal decision accuracy. [1]

This should not be read as an argument against learning. Learning new names may come with learning better properties, correcting a misleading cue, or discovering the actual criterion values. Smithson's basic analysis holds the knowledge cue's properties across the full collection fixed. Real learning need not obey that simplification. Expanding the list of names one recognizes and improving the information used to choose between them are different changes, even when they occur together.

A more informative question for evidence

Mathematical counterexamples can demonstrate that an advantage is possible, or that a proposed condition is insufficient by itself. They cannot establish how frequently the advantage occurs in everyday life. Smithson therefore argues for studying the collection of objects, the knowledge cues, and the order of learning together. Comparing people by how many names they recognize does not ensure that they recognize the same names or use the same additional cues. [1]

Following the same agent over time can reduce some of these problems, but changes in the objects' real rankings or in the agent's knowledge strategy still require attention. Recognition itself need not be perfect either. Believing that a name has been encountered is not guaranteed to match actual prior experience. The paper develops separate mathematical treatments of recognition errors; the numerical example discussed here belongs to its simpler perfect-recognition setup.

The productive question is thus not how much ignorance a decision maker possesses. It is what encounter process separates the recognized objects from the unrecognized ones. Before attributing a shortcut's success to thinking less, we need to ask which comparisons it handles and what information its alternatives use. Smithson's analysis makes knowledge distribution and decision routes visible within an apparently simple contrast between knowing more and knowing less. A possible advantage for partial recognition does not make further learning worthless. It shows why the identities of the things learned belong in the explanation alongside their number.

Sources

[1] Michael Smithson (2010), When less is more in the recognition heuristic, Judgment and Decision Making, 5(4), 230–243. Full primary research article. Cambridge publication record and CC BY 3.0 license.

Source access and factual review: 14 October 2026. The introduction, basic model, numerical examples, and relevant discussion were directly examined. Constructed results are not presented as a human experiment; no independent proof of the appendix theorems or reanalysis of empirical data was performed. Original explanation and adaptation, attributed to Michael Smithson, based on the publication version licensed CC BY 3.0.