LLM Consensus Missed Medical Truth Across Health Claims
Agreement among large language models did not behave like a trustworthy signal in health claim verification. In a 108-claim benchmark built from short Russian health statements, even the strongest systems often converged on the same answer for the wrong reason, producing a kind of shared error rather than shared accuracy. The study tested 10 models spanning frontier systems, Russian-specific models, and medical-specialized families across Russian text and a validated English translation, with multiple decoding temperatures and three stochastic runs per claim. In total, the evaluation generated roughly 90,000 judgements, enough to expose a pattern that is difficult to dismiss as a sampling artifact.
Frontier models performed best on raw classification, with macro-F1 scores in the 64–70 range, yet none of the systems was reliable enough for clinical use as a truth verifier. The most striking failure involved the “disputed” label. Across all judgements, models assigned “disputed” to 59.8% of outputs, despite disputed claims representing only 28 of 108 items in the benchmark and a 26% true base rate overall. In practical terms, the systems leaned toward epistemic hesitation even when the claim was actually true or false, and that hesitation did not translate into better discrimination of medical correctness.
False claims were especially vulnerable to being missed. The models did not simply overuse uncertainty in a neutral way; they systematically failed to identify falsehoods with the consistency one would want from a verification tool. When a system defaults to “disputed” for claims that are factually true or false, the apparent caution can mask a loss of clinical utility. In misinformation screening, a vague label is not a safeguard if it does not separate incorrect statements from accurate ones.
Majority Voting Produced a Shared Mistake Rather Than a Safer Answer
The ensemble result was more unsettling than the individual model scores. Majority voting did not improve performance. Three frontier models reached unanimous agreement on 90 of the 108 claims, yet 31% of those unanimous decisions were wrong. Even more revealing, 26 of 28 claims labeled “disputed” by the ensemble were actually true or false, split across 14 true claims and 12 false claims. Consensus, in this setting, did not signal reliability. It often signaled synchronized uncertainty or synchronized misclassification.
This failure pattern matters because ensemble methods are frequently treated as a practical substitute for ground truth when models are combined at scale. The logic is appealing: if multiple systems independently converge on the same answer, confidence should rise. Health claim verification does not appear to obey that rule. When models share similar training data, similar alignment pressures, or similar prompt sensitivities, they may inherit the same blind spots. The result is a consensus mirage—a state in which concordance looks reassuring while medical truth remains poorly tracked.
The finding carries direct implications for digital health workflows built around automated screening of online misinformation. A system that repeatedly labels statements as disputed may appear conservative, but conservatism is not the same as clinical accuracy. On a claim-by-claim basis, the difference between true, false, and disputed is not academic; it determines whether readers receive a correction, a warning, or a shrug. In public-facing health information, those distinctions shape whether misinformation is challenged or allowed to circulate with a veneer of machine-generated neutrality.
Why Agreement Failed to Track Accuracy in a Medical Setting
The pattern fits a broader limitation of current LLM behavior in high-stakes factual tasks. Language models are optimized to generate plausible text, not to maintain calibrated medical truth under adversarial or ambiguous conditions. When faced with short health claims, they can produce a strong consensus around language that sounds careful without actually mapping onto evidence. This is particularly relevant for claims that sit near the boundary between partially true, context-dependent, and plainly false statements, where a model may prefer the safer-sounding label rather than making a decisive factual commitment.
The multilingual design of the benchmark strengthens the message. The effect held across both Russian and a validated English translation, and it remained robust to prompt wording and language. In other words, the problem was not confined to a single phrasing quirk or one linguistic setting. It was embedded in the way contemporary models handled medical verification itself. Robustness to prompt changes can be useful in deployment, but here it mainly showed that the flaw was structural rather than incidental.
The study also matters because it separates model confidence from medical truth in a way that many users may not expect. Clinicians, editors, and patients may assume that agreement across several AI systems is analogous to agreement among human experts. It is not. Human specialists bring independent reasoning, domain memory, and a willingness to challenge one another on evidence; LLMs can instead converge through shared statistical habits. When those habits are misaligned with validity, a unanimous answer can still be wrong.
The benchmark itself was carefully curated, with 108 short claims split into 40 true, 40 false, and 28 disputed statements, all validated through multi-stage physician review. That design gives the results weight because the reference standard was not casual annotation. The study does not prove that LLMs are useless for health information triage. It does show that agreement among contemporary models cannot be treated as a proxy for medical correctness, even when the models look strong in isolation. Any real-world verification system will need external grounding, not just internal consensus, if it is to distinguish accurate health information from misinformation with clinical credibility.