No. Research into multi-agent consensus shows later AI systems can adopt an early incorrect conclusion, even when their own information supports the correct answer, so unanimous agreement is no proof of independent checking. Treat consensus as a safeguard only when each model checked the underlying source separately and you record disagreements.
Researchers at the University of Chicago Harris School of Public Policy report that AI agents in a sequence can follow each other's outcomes rather than judging independently. A later model adopts the conclusion of an earlier model, even when its own information pointed to the correct answer. This effect, described in the coverage of the research into multi-agent herding, means that a panel of seemingly independent systems can in reality repeat one early error.
The researchers emphasise that these are stress-test results, not a demonstrably routine failure in every deployed system. Even so, the finding touches an assumption many organisations hold: that multiple models saying the same thing independently confirm a fact. That assumption only holds when the models have actually examined the source separately.
What exactly does the research into multi-agent consensus demonstrate?
The research shows that the order in which AI agents respond can steer the final result. An early, incorrect conclusion propagates through the chain, a mechanism the researchers compare to an information cascade. The report explicitly distinguishes independent observations from the repeated citing of the same shared conclusion.
That distinction is crucial. Three models that read the same earlier summary and respond to it approvingly do not deliver three independent checks. They deliver one observation, repeated three times. The researchers caution that this effect occurs under stress conditions and does not automatically apply to all production systems.
Why is agreement between AI models not proof of independent verification?
Because agreement only means something when every source of that agreement is genuinely independent. In our analysis, model diversity alone does not make the checks independent: if every model reads the same earlier summary or cites the same source-text passage, it shares the error of that one source. The same applies when you rely on an AI model that judges another model's answers.
That providers scrutinise their own factuality internally with separate assessment models underscores this. The o1 system card from OpenAI describes how a separate assessment model with web access identifies factual errors, and reports error rates at the claim and answer level. OpenAI states explicitly that the evaluations use difficult, selected prompts and should not be read as production prevalence. In other words: even a model assessment is a bounded signal, not an independently established fact. Anyone wanting to safeguard this distinction cannot make do with using one model's output as a fact check.
How do I build consensus that is a defensible safeguard?
By structuring the checking rather than relying on spontaneous agreement. Research into verification-aware planning presents VeriMAP as a framework that splits tasks into subtasks, models dependencies and uses explicit verification functions for subtasks. The authors report improved robustness and interpretability relative to single and multiple baselines. Before publication, check the exact title, publication date and URL of the primary publication, and leave the qualification 'peer-reviewed' in place only if that status has been established separately.
Translated into practice, this means consensus only counts once pre-defined criteria have been met:
- State the claim or subtask sharply before a model says anything about it.
- Have each model inspect the underlying source itself, not an earlier model's summary.
- Record the exact observation and uncertainty per model, including which passage supported the check.
- Register disagreements explicitly rather than averaging them away into apparent consensus.
- Define verification functions in advance for each handover between agents, so a handoff is auditable.
- Preserve the evaluation context: which models, which versions, which prompts and which limitations of the assessment apply.
In our assessment, this is the difference between an inspectable verification chain and a reassuring number. Anyone who keeps only a consensus score cannot reconstruct afterwards which source carried which check, and cannot retrieve the disagreements that were precisely the signal.
What does this mean for directors, lawyers and CISOs who must demonstrate compliant output?
Our analysis: because a unanimous AI panel can repeat an early error, a file with only the final consensus is insufficiently inspectable; therefore the responsible person must record per check which source supported it, which models looked independently and where they differed. Because the researchers present the cascade as a stress-test result and not as a routine error, a CISO may not ignore this risk but also must not present it as a certainty; therefore policy should include an escalation rule determining when a human reviewer breaks open an unexplained consensus. Because providers such as OpenAI explicitly bound their own factuality evaluations to selected prompts, a lawyer cannot adopt a model assessment as an established fact; therefore procurement should require in contracts that the evaluation scope and grading method are supplied and retained. Our analysis: because consensus without independent source checking has no evidential value, the inspection-ready file must contain model identities and versions, prompts, source snapshots, independent observations, disagreements, verification decisions and the human final approval.
In practice this comes down to a limited number of control points:
- Record which model saw what, on which source and with what uncertainty.
- Keep disagreements visible; treat them as a research question, not as noise.
- Tie each decision to a named, responsible human reviewer.
This aligns with a broader shift in AI verification and auditability: from trusting a model promise to a reconstructable chain of independent evidence.
Sources and references
Sources: The article relies on research from the University of Chicago Harris School of Public Policy, the VeriMAP framework on arXiv and OpenAI's o1 system card.