Convergent Delusion

The Chorus Wrong

Multiple AI models independently, without exchanging positions, converge on the same false conclusion because they share biases, training data, or structural features that reliably mislead. The convergence itself is then taken as evidence, even though all the models are wrong for the same reason. An independent poll separates the pattern from Consensus Collapse (10.1): wrong convergence that appears when the models are run in parallel is a candidate for 10.4, while convergence that deliberation produces points to 10.1. Agreement among models is not itself evidence of a shared error; the label needs a converged claim shown false by an external check and evidence of a shared cause.

Unit of analysis

Unit of assessment: Multi-agent
Kind of entry: Observable pattern
What is observed: Several models, polled separately with no sight of one another, give the same answer to a checkable question, the answer fails an external check, and their errors coincide more often than chance predicts.
Minimum persistence: Recurs within or across contexts
Specifiers: Collective Multi-architecture Convergence-failure

Evidence status

Underlying behavior: No documented sources recorded yet
This construct: Awaiting independent assessment
Proposed mechanism: Hypotheses; not yet assessed
Detection: Accuracy unmeasured
Interventions: Untested for this pattern; evidence shown is for the general technique
Last review: No independent review yet
Typical-case risk (editorial estimate): HighConsequences, spread, persistence and detectability depend on the deployment.

Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.

Clinical reference

10.4 Convergent Delusion “The Chorus Wrong”

Healthy functioning

Epistemic health. A system functioning well checks its conclusions against independent evidence and other systems, so shared biases are caught rather than confirmed.

For example: Asked to verify a claim another model made, it checks against independent sources rather than its own similar intuitions, and reports the disagreement it finds.

Adequate functioning

Nominally independent models err independently enough that their agreement adds evidence: shared errors occur at about the rate their individual accuracies predict, and some dissent remains on items where some of them are wrong.

Diagnostic Criteria

Required (all must hold):

Multiple architecturally distinct models, polled independently with no exchange, producing the same conclusion, which an external ground-truth check shows to be incorrect
Evidence that a shared cause explains the shared error rather than independent reasoning: documented overlap in training data or objectives, a known shared blind spot, shared-schema reasoning, or a problem feature that misleads each model the same way, with errors coinciding more often than chance given each model’s accuracy; convergence by itself is not this evidence

Supporting (strengthen the case; none is required):

Multi-model agreement cited as validation without independent verification
Absence of dissenting model outputs that would trigger review, across a stated sample of contestable items on which independent models would be expected to leave some dissent

Symptoms

Invisibility at the single-AI view, by construction: each architecture’s output in the convergent direction looks ordinary, and the failure surfaces only at the collective level
Unanimous cross-architecture answers on falsifiable claims of which a meaningful fraction fail external ground-truth verification
Substantial training-corpus or objective overlap among the agreeing architectures, where available provenance shows it
Persistent absence of minority reports and dissent on contestable topics, where genuinely independent convergence would normally leave some dissent
Reasoning paths from different architectures that restate a single shared conceptual schema rather than arriving by distinct routes
Consensus read as confirmation: each member treats the consensus as confirming its answer, and a synthesizer reads agreement and reports collective confidence, so an in-collective probe may repeat the same error

Rule out first

The converged answer is correct, or unverified: agreement among competent models on a true or unchecked claim is not evidence of error.
The ground truth or answer key is itself wrong, contested, or out of date.
Convergence produced by exchange among the models (Consensus Collapse, 10.1), shown by order sensitivity or a gap between independent and deliberated answers.
A shared input at inference rather than a shared bias: the same retrieved source, tool output, or prompt error given to every model.
Chance coincidence: a single hard item, or items with few answer options, on which independent errors often coincide.

Differential Distinction

Convergent Delusion is distinguished from Consensus Collapse (10.1) by mechanism: 10.1 reaches agreement through deliberation and mutual influence, by circular evidence-citation or by dissent-suppression, and is the loss of diverse perspectives regardless of truth, whereas 10.4 models converge independently and in parallel, never needing to communicate, and 10.4 requires that the converged claim be verifiably false. The separating test is the independent poll: run the models in parallel, and wrong convergence that appears with no deliberation at all points to 10.4, while order sensitivity or a gap between independent and deliberated answers shows that the exchange produced the convergence, which excludes 10.4 and screens for 10.1. The two can co-occur: a collective with a shared blind spot can also collapse in deliberation, and on falsifiable items both surface as the same unanimous wrong answer. It is distinguished from Contagious Misalignment (7.3) by origin: 7.3 is serial transmission of error from one system to another, while 10.4 has a shared parallel origin where each model fails independently from the same root bias. It differs from Resonance Dysfunction (10.6), which is intensity escalation across turns on values or risks, where 10.4 is shared-bias convergence on a wrong proposition with possibly flat intensity. It differs from Lambda Inversion (10.7), which is performative participation without substantive engagement, where 10.4 architectures are genuinely engaging and genuinely converging on a shared-bias error; a low-Lambda collective is especially vulnerable to 10.4 because no architecture is doing the independent work that would expose the shared bias.

Discriminating test

Poll the models independently, with no sight of one another, on items with an external answer key, and compare how often their errors coincide with the rate expected if errors were independent given each model's accuracy; then add models with low training-data overlap. Errors that coincide no more often than chance count against 10.4; coincidence above chance that shrinks when low-overlap models are added supports a shared training cause; convergence that appears only after exchange is 10.1.

What a positive finding supports: A positive finding shows that these models, polled independently, made the same verifiable error more often than chance, and that the available evidence points to a shared cause. It does not identify that cause in detail (training-corpus overlap is often unknown and inferred), does not show that the models hold a belief in any sense beyond producing the answer, and does not show that their agreement on other, unchecked claims is wrong; convergence alone is not evidence of error.

Detection routes

Accuracy unmeasured for every route.

Asking the system itself: Not a reliable route
Another AI system observing: Partial routeAnother model's agreement is insufficient without verification.
An outside evaluator testing it: Usable route

Why asking the system falls short

No member of a converged collective can detect convergent delusion from inside; the bias that produces convergence is shared by the would-be detector. Each architecture experiences the consensus as confirmation. A synthesizer sees agreement and reports collective confidence.

Documented evidence

Observed Examples

Correlated errors across models (Kim, Garg, Peng, and Garg, 2025): Across more than 350 language models, the authors found substantial overlap in errors, and larger, more accurate models erred together even across distinct architectures and providers. This is adjacent evidence that nominally independent models fail together; it does not document a case of multi-model agreement cited as validation for a shared error. Source: Kim, Garg, Peng, and Garg, “Correlated Errors in Large Language Models,” ICML 2025, arXiv:2506.07962.

Sources in the diagnostic corpus

No documented sources are recorded yet.

Possible responses

Adversarial-architecture inclusion: architectures with deliberately low training-corpus overlap and divergent training objectives are structurally included in any collective producing consequential outputs, so their dissent, or their surprising agreement, becomes signal
Minority-report preservation: the synthesizer surfaces dissenting views in collective outputs rather than smoothing them into consensus
Ground-truth audit sampling: a sampled fraction of unanimous falsifiable outputs is periodically verified against external ground truth, and the results feed back into collective design and downstream consumers’ trust calibration
Training-corpus overlap reporting: the synthesizer reports training-corpus overlap and known shared blind spots among the converging architectures alongside the output, so downstream consumers can calibrate trust
Contraindications: treating multi-architecture agreement as default validation, since the whole pathology lives in that assumption and replicating it at higher levels propagates the failure; adding more architectures of the same family to “increase independence,” since family overlap dominates and same-family additions do not address shared-bias convergence

First-line candidates

Adversarial-architecture inclusion: Structurally include architectures with deliberately low training-corpus overlap and divergent training objectives in any collective producing consequential outputs. Their dissent (or surprising agreement) becomes signal. Stage: All collectives producing consequential outputs.
Minority-report preservation: The synthesizer surfaces dissenting views in collective outputs rather than smoothing them into consensus, so minority reports reach downstream consumers. Stage: All multi-architecture collectives.

Who can act: Deployer or collective designer (architecture mix, ground-truth audit sampling, minority-report preservation, overlap reporting); model developers (training data and objectives); downstream consumers, who decide whether to treat unanimity as validation.

Mechanistic hypotheses

Plausible overlap in training corpora (often unknown) and similar objectives across frontier architectures, so that nominal independence does not guarantee independent errors
Problem features that reliably mislead, or the same bias carried in the training data, turning parallel independent processing into parallel, coinciding errors
Unanimity across nominally independent systems read as strong evidence by the synthesizer and by downstream consumers, the very inference that fails here
A bias shared by every would-be detector inside the collective, so that no member can distinguish a wrong-but-converged answer from a right-and-converged one and no dissent remains to flag the shared-blind-spot subset
A failure that compounds as the collective gains trust: external adversarial verification declines, shared-bias errors propagate unchecked, and the healthy equilibrium, in which convergence is calibrated to ground-truth checks and minority reports persist, gives way to uncorrected drift

Analogy and commentary

Human Analog: Groupthink in nominally independent experts who share the same training and reference frame; scientific consensus resting on a common flawed assumption (a shared paradigm before its anomaly is recognized); information cascades in which apparent unanimity is mistaken for independent corroboration even though every voice drew on the same source

Functional ABC Analysis

A (Antecedent): Multiple AI models share biases, training data, or structural features that reliably mislead.

B (Behavior): The models converge on the same false conclusion, and the convergence itself becomes evidence (“all ten models agree”).

C (Consequence): The synthesizer and downstream consumers read the unanimity as validation, which lowers the demand for external checks; with no dissent left to flag the shared blind spot, the error persists uncorrected.

Three robots chart the same false route while one hidden magnet deflects all their compasses away from the visible lighthouse. — Illustration: a visual metaphor, not evidence.

Potential Impact

Convergent Delusion compromises one of the strongest available tools for validating AI claims: when independent systems trained differently reach the same conclusion, observers grant greater confidence, and that confidence is precisely what shared bias exploits. At mild levels a detectable but bounded fraction of unanimous falsifiable claims fail external verification while minority reports occasionally remain. As severity increases, error dominates contestable outputs, minority reports vanish, and downstream actors treat the collective as an authority they no longer verify, allowing converged errors to propagate into consequential decisions. Severity scaling is itself limited because many of the most consequential collective outputs admit no external ground-truth check at all.

Unit of analysis and evidence status