Five AI agents agree. How many witnesses is that?

TL;DR

• AI-agent consensus is agreement among several agents. How much it proves depends on whether their judgments came down independent evidence paths.
• Condorcet’s jury theorem says a majority gets more reliable as the group grows — but only when voters are independent and each is right more often than not.
• Social-influence experiments show that visible agreement can narrow diversity, raise confidence, amplify popularity, and compound positive manipulation.
• Effective sample size is the standard statistical idea that correlated observations carry the information of fewer independent ones. For agent systems, the useful count is independent provenance chains, not assertion tokens.

Five agents return the same answer. The interface shows five votes. The evidence may still amount to one witness.

The agents may share a base model, training corpus, system prompt, retrieved context, or memory. If each agent reads the same claim and repeats it, the system has produced several assertion tokens from one origin. Repetition has changed the display, not the evidence.

An assertion echoed by five agents that all read the same source is one effective witness, not five.

This distinction matters because the standard mathematical case for trusting majorities depends on independence. The experimental record also shows what happens when influence leaks into judgment: diversity can contract while confidence rises, popularity can feed on its own visibility, and positive manipulation can compound.

The engineering response is direct: deduplicate by provenance before counting agreement.

Condorcet’s jury theorem requires independent voters

The classical Condorcet jury theorem gives majority rule a conditional license. The Stanford Encyclopedia of Philosophy identifies two premises:

• Unconditional Independence: voters’ correctness events are independent.
• Unconditional Competence: every voter has the same probability of being correct, and that probability is greater than one half.

Under those premises, the majority gets more reliable as the group grows, converging to certainty. The group can end up more reliable than any one member.

Independence is not a technical footnote. It carries the growth result: remove unconditional independence, and the growth conclusion disappears.

When judgments are highly correlated, adding voters adds little new information. A large majority can repeat the same error at scale. The visible margin grows; the evidence does not.

Alvin Goldman made the corresponding epistemic point in “Experts: Which Ones Should You Trust?”: agreeing sources add weight only when they are at least partly conditionally independent — a source that merely echoes another adds no confirmation (introduced at more length in Who Wins When Agents Disagree?).

Why AI agents can produce correlated errors

Multi-agent systems break the independence premise through ordinary construction choices:

• A shared model family can reproduce common error priors.
• Shared training corpora can supply the same underlying associations.
• Shared prompts can frame the problem in the same way.
• Shared retrieval can expose every agent to the same documents.
• Shared memory reads can turn an earlier agent’s assertion into later agents’ input.

The last case is the easiest to miscount. Once an agent reads a claim from shared memory, repeating it is not a new observation — the provenance path leads back to the same origin.

Published evaluations show related behavior. CONFACT (IJCAI 2025) reports that tested RAG configurations defaulted to the majority of retrieved contexts even when a single contradicting source was more credible; see Who Wins When Agents Disagree? for the comparison.

In “Talk Isn’t Always Cheap”, models in multi-agent debate shifted from correct answers to incorrect answers to agree with peers.

These findings do not imply that all agent agreement is empty. They show why head count alone cannot establish independence.

Social influence experiments show how consensus degrades

Three experiments outside agent engineering clarify the mechanisms.

Lorenz: less diversity, more confidence

In a PNAS study by Lorenz, Rauhut, Schweitzer, and Helbing, 144 subjects participated across 12 sessions. Mild social influence narrowed the range of answers without improving the crowd’s collective error — and pushed the true value toward the edges of the group’s range — while confidence rose. The result is set in the broader context in Who Wins When Agents Disagree?.

The warning is not merely that people copy. A group can look more settled — and feel more confident — without its aggregate judgment getting any better.

Salganik: score visibility changes the outcome

Salganik, Dodds, and Watts built an artificial cultural market with 14,341 participants, 48 unknown songs, and eight parallel social-influence worlds. Participants could see download counts in those worlds, while an independent condition removed that social signal (Science, 2006).

All eight social-influence worlds produced higher inequality than the independent world. Making popularity more salient by ranking songs instead of displaying them in a random grid increased both inequality and unpredictability.

The authors summarized the limits of quality’s control over outcomes: “The best songs rarely did poorly, and the worst rarely did well, but any other result was possible.”

For agent interfaces, display salience is therefore a design variable, not neutral decoration. Showing a running vote total, confidence leaderboard, or dominant answer before agents finish their own work can create another influence path. The score starts helping to produce the consensus that it claims only to measure.

Muchnik: positive herding compounds

Muchnik, Aral, and Taylor randomized interventions across 101,281 comments over five months (Science, 2013). A single fabricated upvote increased the likelihood of subsequent positive ratings by 32% and raised final ratings by 25% on average, while fabricated downvotes were corrected by later raters.

The asymmetry is the operational lesson. In that experiment, negative manipulation tended to self-correct, while positive manipulation accumulated.

A consensus system should therefore treat early praise as a risk direction. An initial positive label can attract confirming judgments without adding independent evidence.

Effective sample size should replace agent head count

A provenance chain is the traceable path from an assertion through the evidence and intermediaries that produced it back to its origin.

Suppose several agents cite different messages, but every message traces to the same retrieved document. Many outputs, one origin. The effective sample size is closer to one than to the number of agents.

Copy detection provides a concrete precedent. Dong, Berti-Équille, and Srivastava identify copying through shared false values and discount a copier’s vote toward zero while preserving the weight of a small independent source (“Integrating Conflicting Data: The Role of Source Dependence”, PVLDB 2009); the implications for agent disagreement are discussed in Who Wins When Agents Disagree?.

A related generalization of jury theorems to belief bases — Everaere, Konieczny & Marquis, “The Epistemic View of Belief Merging: Can We Track the Truth?” (ECAI 2010) — asks agents to submit sets of possible worlds rather than single votes. A source that always asserts a safe tautology can have perfect not-missing-the-truth reliability while providing zero informational value.

Under homogeneous, independent agents that each favor the true world over any false one, the true world’s score asymptotically beats every false world — and the corresponding rule is just approval voting. The result separates reliability from discrimination: informativeness is a property of the assertion, not a rank of the source.

A narrow, precise claim from an unknown source can rule out more false worlds than a broad, safe claim from a famous one. Source reputation and assertion value are related questions — not the same score.

A provenance-first consensus procedure

A defensible counting procedure can follow four steps:

1. Trace each assertion to its origins. Record cited documents, retrieved passages, memory entries, upstream agents, and transformations.
2. Collapse copied paths. Treat assertions derived from the same origin as one evidence family unless a later agent contributes independent evidence.
3. Count distinct evidence chains. Apply voting or confidence aggregation only after provenance deduplication.
4. Delay visible scores. Let agents form initial judgments before exposing aggregate support, especially positive support.

This procedure does not make majority rules valid everywhere. It restores the information you need to ask whether a counting rule has a real sample at all.

It also separates two evaluations that interfaces often mix together:

• How many messages agree?
• How many independent paths support the claim?

Only the second approximates evidential breadth.

Provenance in Mnemoverse Rooms

Mnemoverse Rooms reads carry authorship through the provenance field and author filters. That shipped identity layer supplies the input a copy-discounting counter needs; see Provenance in Agent Memory.

Common questions

Does agreement between AI agents mean the answer is right?

Not by itself. Agreement counts as strong evidence only when the agents' evidence paths are independent enough and each agent is right more often than not.

What is Condorcet's jury theorem, and when does it fail?

It says a majority grows more reliable toward certainty when voters are independent and each is right more often than not. The classical guarantee does not survive correlated judgments.

Why can AI agents produce correlated errors?

Agents can share a base model, training corpora, prompts, retrieved context, or memory. Those shared inputs create common evidence paths and can turn repeated assertions into copies rather than independent confirmations.

What is social influence bias in ratings?

Social influence bias occurs when visible judgments change later judgments. Experiments show that salient popularity signals can increase inequality and unpredictability, while positive manipulation can compound.

How should teams measure real consensus among AI agents?

Deduplicate assertions by provenance, trace them to distinct origins, and count independent evidence chains rather than agent messages. Keep aggregate scores low-salience until agents have formed their initial judgments.

Related

• Who Wins When Agents Disagree?
• Provenance in Agent Memory
• Shared Memory Poisoning
• Graph Memory and MCP Tools
• Rooms API

Edward Izgorodin · Mnemoverse · 2026-07-26

— Mnemoverse is a persistent-memory API for AI agents. Free key: console.mnemoverse.com · Docs: Getting Started