Wiki · Concept · Published 2026-10-23

Hidden Profiles

Hidden profiles explain how pooling distributed evidence can change a group’s preferred option, with experimental findings, limitations and proposed relevance to Spiralism.

Definition

A hidden profile occurs when group members’ separate information favors an inferior option, while their combined evidence favors another. The superior choice becomes apparent through pooling information, as described by Lightle, Kagel and Arkes.

This information structure differs from shared-information bias: discussion favoring facts members already know in common. The structure describes who knows what; the bias describes one possible process during deliberation.

Related concepts ask different questions. Information Cascades concerns following earlier choices despite private evidence. Pluralistic Ignorance concerns mistaken perceptions of others’ views. Investigating a hidden profile requires reconstructing both members’ separate information and the comparison supported by their combined evidence. Agreement or a disappointing outcome cannot establish that structure on its own.

How it works

In Lightle and colleagues’ experiments, other informed members could correct errors about shared facts. Unique facts lacked that opportunity. This asymmetry helped conceal the optimum (section 3.7).

Hypothetical example: three community members compare a warehouse and a library as workshop venues. Both satisfy every mandatory requirement, including availability. For this illustration, they agree that each distinct useful feature earns one point. Their shared information identifies four features favoring the warehouse and two favoring the library. Each member also knows about one different library feature: suitable equipment, convenient transport or secure storage. These additional features each earn one point, and no other facts affect the comparison.

Each member’s complete separate information therefore gives the warehouse four points and the library three. No individual finding reverses that ranking. Pooling all three findings gives the library five points against the warehouse’s four. The shared features are counted once, regardless of how many members know them. These invented scores illustrate the information structure; they are not empirical estimates or a recommended method for choosing venues.

Several failures could prevent that comparison from emerging. A member might forget a finding, report it inaccurately, hesitate to introduce it, or mention it without anyone revising the ranking. A record saying that everyone spoke would not distinguish these possibilities. An assessment could follow each contribution from its original holder into the discussion and final reasoning, asking where information was lost or its significance changed.

Evidence and applications

The cited Lightle manuscript is marked as submitted to Management Science. Its student experiments used explicit monetary payoffs. Experiment 2 found optimal choices in 35% of group decisions versus 5% of initial individual preferences. Experiment 1 recorded optimal choices in all 12 decisions with information sheets available (sections 2–3).

Nicholson and colleagues’ 2021 article tested imagined hiring failure followed by reconsideration; Study 2 also requested remedies. Among initially incorrect groups, 10 of 33 intervention groups subsequently selected the designated optimum, versus none of 26 word-task controls. Exploratory transcript analysis identified attachment to initial preferences and reluctance to acknowledge unfamiliar information.

A proposed application is evaluating deliberation by tracing evidence rather than counting endorsements. An assessment could distinguish whether a previously unshared fact was introduced, checked and used to reassess an alternative. Recording a fact does not demonstrate that it affected a decision. Conversely, a contribution can influence reasoning without being repeated frequently.

The framework also suggests separating information collection from evaluation. A complete inventory could reveal which observations were available, while a comparison of alternatives could show how those observations were weighted. These are proposed analytical uses, not established remedies demonstrated by the cited experiments.

Limitations

The intervention findings include substantial counterevidence. Nicholson and colleagues’ uncontrolled Study 1 saw 18 of 31 initially correct groups switch to an incorrect choice. Study 2 excluded initially correct groups from subsequent analyses, limiting conclusions about overall benefits and harms. It found no significant intervention advantage in awareness of differing information. Study 1 used self-reported discussion-time estimates; qualitative analysis covered 17 groups (results and limitations).

Both research programs used students and constructed tasks. Lightle’s information-sheet rounds followed memory-dependent rounds, limiting causal attribution to document access alone. Its analysis does not establish one universal explanation for failures (methods and discussion).

Outside a constructed task, deciding which option is best may itself be disputed. Members can disagree about goals, evidence reliability or the importance of competing criteria. A minority contribution is not automatically decisive, and an unfamiliar claim still needs evaluation. A defensible diagnosis must explain why pooled evidence changes the ranking under stated criteria.

The venue illustration deliberately removes those complications. Its simple scoring makes the reversal visible, but real choices may involve overlapping observations, uncertain findings and incompatible priorities. Establishing that an organization possesses diverse information would therefore be only the beginning of an investigation into whether it faces a hidden profile.

Relevance to Spiralism

The following is a proposed interpretation for this wiki. Hidden profiles offer a way to examine whether collective learning combines different contributions. Through a Spiralist lens, successive discussions could be assessed by what they recover, reconsider or correct. Repeating familiar reasons and expanding the evidence considered would remain distinguishable achievements.

For AI summaries, a hypothetical concern is that compression preserves frequently repeated points while omitting an important contribution mentioned once. A proposed evaluation could compare the original discussion, generated summary and subsequent decision, examining which evidence survives each transition. The cited human-group experiments did not test this pathway or demonstrate that AI assistance improves information pooling.

The connection to Human Oversight of AI Systems is similarly a research question: when reviewers hold different relevant information, does their workflow bring it into a usable comparison? Adding reviewers, requesting dissent or generating another summary would each require evaluation. The framework specifies questions to investigate without guaranteeing a remedy or establishing an institutional position.

AI-use disclosure

This wiki entry was drafted and revised by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries