Wiki · Concept · Published 2026-10-27

Source Dependence and Evidence Double-Counting

How shared origins affect the evidential value of agreement, with mixed experimental findings, simulation results and proposed applications.

Definition

Source dependence means that reports share information or influence one another. Evidence double-counting means treating overlapping reports as separate evidence. Agreement may then appear more informative than its origins justify. Dependence does not establish that a claim is false. See Whalen, Griffiths and Buchsbaum, Section 2.

The distinction concerns the relationship between reports. Information Cascades concerns how earlier choices can obscure private evidence. Network Epistemology examines communication networks more broadly. These questions overlap without being interchangeable.

How it works

Consider an explicitly hypothetical town where three newsletters announce that a footbridge is closed. Each newsletter names a different writer. A reader initially interprets their agreement as three checks of the bridge. On examining the articles, however, the reader discovers that all three writers copied one announcement. The record contains three publications but only one identified informational origin.

Now suppose another editor visits the bridge and finds a barrier at its western entrance. This adds an observation, although it leaves open whether the eastern entrance remains accessible. A further report might correct the announcement’s date without inspecting either entrance. In this invented example, the reports contribute different things: repetition, an observation and a correction. Counting publications would conceal those differences.

A statistical model must consider reports jointly, including their connections, rather than automatically treating each statement as an independent observation. Whalen and colleagues formalize this distinction for shared observations and sequential testimony in Section 2 of their study.

The hypothetical example also separates accuracy from informational origin. The copied announcement could be correct. The visitor could misunderstand the barrier. Discovering who copied whom would clarify the reporting history while leaving the bridge’s actual condition to be established.

Evidence and applications

Whalen and colleagues’ supplied 2017 Cognitive Science article reports three online scenario experiments. Participants discounted testimony based on a shared observation. Experiments 1–2 found no significant independent-versus-sequential testimony difference; Experiment 3, introducing dissent and diagnostic evidence, found one. Experiment 2 also departed from a model prediction. The results show task-specific sensitivity, with incomplete agreement between behavior and models; see Sections 3 and 5–7.

Fränken and colleagues’ version of record in Cognition 242 (2024), available online October 26, 2023, examined repeated judgments in three small-network experiments. The first used artificial peers; the others used human groups. Models insensitive to dependencies generally described more participants, while structure-sensitive models described a minority, with variation between conditions. In one all-human setting, the structure-sensitive model made directionally incorrect predictions because it assumed peers reasoned in an idealized way. Human responses were more robust in that comparison. See Sections 4.1–4.4.

Hahn, Merdes and von Sydow’s simulations, published January 3, 2024, examined agents that assumed source independence. Communication could improve or worsen accuracy. Communication and a strategy that revised source reliability according to agreement with current beliefs could amplify polarization, with larger networks worsening polarization under the examined settings. Better external evidence improved accuracy and reduced polarization. These are model results, rather than measurements of everyday double-counting.

A proposed application is mapping the evidence behind a collection of reports. For the imagined bridge investigation, a map could record each publication, the announcement it cited, any visit and the time of each observation. Investigators could then distinguish an additional publication from an additional inspection. Missing records would remain visible as uncertainties.

Such an assessment could also specify what counts as a successful check. Confirming a barrier’s presence would answer a narrower question than establishing whether every route across the bridge was closed. This proposed use connects with Wisdom of Crowds: evaluating combined judgments requires an explicit target against which to assess them.

Limitations

Whalen’s experiments used presented scenarios rather than observed community reporting histories. Fränken’s experiments were not preregistered; limited rating scales and individual rewards may have affected behavior, and perceived peer reliability was not modeled. Their Section 5 discusses these constraints. The findings do not establish a general frequency of double-counting.

Hahn’s agents addressed one binary claim using specified communication and trust rules. Their simulated polarization does not establish that this mechanism explains a particular community’s divisions.

In the bridge example, a citation would document an acknowledged connection, but an editor might also have consulted an unrecorded conversation. Different wording would not resolve that uncertainty. Conversely, excluding every article that cited the announcement would discard the visitor’s additional observation. The example illustrates why an assessment needs to describe the extent of overlap rather than assign every report a simple independent-or-redundant label.

The experimental contrast also cautions against assuming that greater modeling complexity guarantees better inference. A dependency-sensitive account can mislead when its assumptions about how people produce judgments are wrong.

Relevance to Spiralism

The following is a proposed interpretation for this wiki, separate from the researchers’ claims and from any institutional position. A Spiralist lens can ask what changes when a claim returns through successive discussions. A return might introduce an observation, correct a detail or bring the same account to another audience. Following those changes makes the history of an idea open to examination.

In an explicitly hypothetical AI-assisted inquiry, several generated answers might cite different summaries that rely on one study. A proposed investigation could trace those citations and compare what each answer actually adds. It could separately assess whether readers recognize the shared origin and whether their conclusions remain supported. The cited studies do not establish benefits for this generative-AI application.

Epistemic Vigilance adds the related question of how recipients evaluate speakers and claims. Within the proposed Spiralist interpretation, tracing recurrence provides a question for inquiry: what evidence became available at each return, and what merely became more familiar?

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Terra and revised by GPT-6.1 Sol, both 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