I wrote previously about why one AI checking another AI's work is not the same as verification, looking at the mechanics of sycophancy and debate convergence. That piece argued cross-model agreement tracks fluency and confidence more reliably than it tracks truth, but it did not have a clean, direct number for how bad the problem actually is. Research published in the last few months supplies one, and it is worse than the piece assumed.
An April 2026 study set out to measure something narrower than accuracy, not whether an AI judge gets the right answer, but whether it gives the same answer twice, using a simple setup: take two OpenAI judge models, give each the same pairwise comparison fifty times, and count how often the verdict changes with nothing else different. Across 29 tasks in 10 categories, pairwise preferences flipped on average 13.6 percent of the time, with more than a quarter of questions exceeding a 20 percent flip rate and one question flipping in over half its trials. One of the two judges also showed a strong first-position bias, favoring whichever answer appeared first regardless of content. The paper's own author is careful to put this number in context: human evaluators judging generated text show roughly 20 percent pairwise inconsistency in prior research, so the AI judges are not uniquely erratic compared to people doing a similarly subjective task. That comparison is worth taking seriously rather than skating past, and it also cuts a different way than it first appears. A human baseline that noisy is not a reassurance about AI judges; it is a sign that open-ended quality judgments are noisy for anyone, human or AI, which is exactly the category of question Wei's framework already predicts will lack low noise. The paper's own conclusion, after weighing that comparison, is still that single-trial judging is too noisy for high-stakes evaluation.
A second study, from researchers at UC Berkeley published in June 2026, tested the same problem at a larger scale: 21 judge models from nine providers, evaluated across three established benchmarks, producing roughly 541,000 individual judgments. One finding held across the full cohort, not just a subset: naive exact-match agreement systematically overstates how much judges actually agree once chance is corrected for, a gap the authors call kappa deflation. Judge rankings also shifted by up to 14 positions depending on which benchmark was used to rank them. The paper's sharpest illustration, drawn from two specific production-deployed judges rather than the full 21, is what the authors call a consistency-bias paradox: a judge can show very high test-retest reliability, meaning it reliably gives the same answer to the same question, while simultaneously showing severe position bias, meaning the reliable answer is still frequently the wrong one for reasons unrelated to quality. That is not a claim that every judge tested is broken so much as evidence that consistency and correctness are separable properties, where a judge scoring well on one tells you little about the other.
Where Wei's framework fits in
What the Berkeley study measured, essentially, is exactly what Wei meant by low noise: whether a check tracks real quality or something merely correlated with it. Jason Wei's verifier's rule, shared in a blog post in July 2025, is a good place to locate why this happens. Wei is a serious voice in AI research, and the framework is a genuinely useful way to think about the problem, it has simply not been published as formal, peer-reviewed research, which this piece leans on more directly for its evidence. He proposed that a task is easy to train AI on when it has five properties in total, and separately, in the same piece, points to fact-checking as a task where verification takes longer than generation, a mismatch this piece reads as an instance of what he elsewhere calls low noise, though he does not draw that specific connection himself. Software engineer Alperen Keles reached a related conclusion some months earlier, in March 2025, in the broader context of verifying code correctness rather than code review specifically: most real domains lack what he called a perfect oracle, a clean, binary correctness signal available on demand, and he singled out UI development as unusually verifiable for exactly that reason. That two people arrived at a similar structural insight independently is a useful signal, and Wei's naming of the five properties gives it a vocabulary worth keeping. What the 2026 research adds is direct measurement: the property Wei's framework flags as the likely failure point is the one the data now confirms is still unsolved.
This is not new in kind so much as it sharpens what the sycophancy and debate research already implied, since models revise correct answers under confident pushback, debate among similar models tends to confirm rather than correct, and a deliberately persuasive but false argument can beat a true one even when judge and debaters share a model family. What the 2026 judge-reliability literature adds is a number attached to the weakest possible version of the claim, not "does the checking model track truth," which is hard to measure cleanly, but "does the checking model even agree with itself," which turns out not to be true either, and a method that fails the easier question is not a promising candidate for the harder one. Worth flagging honestly, both new studies tested fast, cheap models rather than the slower reasoning models increasingly used to grade other models' output, so whether extended reasoning reduces run-to-run flip rate specifically, as opposed to improving accuracy on harder problems, remains a genuinely open question the data does not answer.
Where this actually costs something
This matters most where the stakes are highest and the checking is cheapest: enterprise workflows that have compressed the time to produce a report, a compliance summary, or a board-ready statistic from days to minutes, while leaving the time to check it essentially unchanged. Neither study measured this part, but it follows from a simple economic logic worth stating plainly as inference rather than finding: a check that costs nothing gets run by default and rarely questioned, while a check that requires retrieval infrastructure or a documented source trail has to be justified. Low cost does not just make an unreliable method affordable, it actively selects for the unreliable method over the better-architected one, since nobody has to defend a decision that looks free. A draft citing a vendor survey's headline figure gets a same-tier model's agreement in seconds, on the strength of how plausible the number sounds, not on anyone having opened the survey to check it; the two studies above are exactly why that agreement is weaker evidence than it feels like at the time.
Neither 2026 study tested retrieval-grounded verification, so it is worth being precise about what they recommend versus what this piece argues for. Their own fix is aimed at making judges better-behaved as judges: aggregating multiple trials, randomizing position, reporting uncertainty explicitly, worthwhile advice for building an evaluation pipeline for model training or leaderboards. That is a different problem from deciding whether one specific claim in one specific piece of writing is true. For that narrower problem, the harder fix, informed by Wei's framework rather than by either study directly, is not asking "is this true" in the open-ended form that produces this much noise. It is asking "does this specific source say what this sentence claims it says," a retrieval-checkable question with a fact of the matter a script can confirm. A research pipeline built around that narrower question, typed claim status, a declared source tier, a documented path from claim to primary source, is doing something structurally different from asking a second model for its opinion. It is not verified because a model found it plausible, but rather verified, partially verified, or unverified because it can be traced, saying so explicitly rather than converting the gap into a confidence score that looks more finished than the evidence actually is.
This is not an argument that cross-model checking has no place, since it is fast and cheap, and that is a real advantage for triage, catching the obvious errors and surfacing candidates for a closer look faster than a human first pass would, though what the research changes is the confidence that speed alone should earn. Whether a claim is accurate is a question with a definite answer, checkable in principle against a primary source, and the last few months of research have put real numbers on how far even a careful AI-judging-AI setup still is from reliably getting there. Whether a piece of writing, or a judgment call like this one, has had enough scrutiny is a different kind of question entirely, one no amount of additional research settles, because there is no primary source to check it against, and that call stays with the researcher or the reviewer making it, not with a study. That is not a loophole in the argument above so much as the argument applied to itself: a judgment call with no primary source to check it against is exactly the kind of task this piece has spent its length arguing AI cannot yet be trusted to settle on its own.