The claim was "All almonds are grown in the U.S. state of California.". All but one model said False, Opus 4.7 said "misleading".
I feel like having "mostly true" and "misleading in there weakens the story, especially given the "no explanations" rule in the prompt.
The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
[ Update: OK, this almond thing was a bad example and I regret picking it. Read on for better ones. ]
The prompt lacks any kind of rubric to clarify how those terms should be applied.
As is so often the case with this kind of study, it's an evaluation of the prompt and harness used by the study in addition to being an evaluation of the underlying models.
Update: here's a better example: "Incomplete Egypt visa application forms are among the most common reasons Egyptian visa applications are rejected."
The models were split between "true" and "mostly true". Given the "among the most" language either of those answers means effectively the same thing.
Update 2: a much better example:
"On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia"
The only correct answer to that, if you don't have a search tool, is "this claim is impossible for me to verify". And that wasn't an option.
Another (IMO fatal) error is they don’t attempt to measure within-model variance.
The thing you find when you actually wire up a rigorous eval is that with tool calls like web search you are wide open to infra issues, flakes, and all sorts of non-determinism.
They really should be breaking out the numbers for the 3 without search (kinda meaningless for recent factual claims after knowledge cutoff) vs search agents.
I do agree the forced choice and “weak / strong” variants inflate the headline stat. To make that distinction you need a much more rigorous prompt, likely including ICL examples to illustrate what you mean by “mostly” instead of leaving this to the model to define.
Without providing definitions of "True / Mostly True / Misleading / False" to each rater, I rate the article's claim that "Only one verdict bucket can be correct per claim" as false.
Something can be simultaneously "misleading" and either true or false. Which category should something go in if it's "mostly false"?
How much can something be wrong before it goes from "mostly true" to "false" (objectively, both have some part of the fact that is not true)?
This is at least partly testing the model's definition of "mostly" and "misleading". Not its understanding of the fact. Claiming that this means the models have fundamental disagreement on the facts themselves is an overreach.
Yes, the labels are weird. Most misleading statements are true. Any "mostly true" statement is false.
I suspect the intention was "Factually true, and no gotchas exist", "technically not true, but so close to the truth that the difference doesn't matter", "technically true, but there are major gotchas" and "factually false and not even close". But that's not what they specified
Better options would have been "True", "False", "Unknown" (which opinions would fall under too). That also includes an interesting assessment of how well LLMs can identify missing information. My guess is they would be a very low number of "unknown" and a much higher level of agreement (assuming equal representation). Unless the RLHF techniques have gotten better at getting an LLM to say "I don't know", which I doubt. Saying "I don't know" is not good for a dopamine release to keep users coming back for more.
Tried initially with a fifth bucket, Abstain. It was actually heavily used by some of the models. But it felt as if they are using this to "avoid" some of the hard questions, and we dropped this bucket to force them to provide a verdict.
Exactly what people do when they use LLMs for "fact-checking" online, and any verbose explanation would be mostly ignored anyway, when people ask political, ethical, or simply ambiguous questions that they hold any stakes in.
Don't even need politics for it, there is no point in probing a mathematical black box for "how many soldiers died in the year X in war Y".
Any original source is preferable to a blurry "summary" of unknown sources, and this is why the article has a valuable point.
There's also no point in asking "Is Paris in France" either, if you substitute city and country with real data. An encyclopedia or manual check of different sources such as maps, while not infallible, is a better source.
If you already know the country Paris belongs to, there's no point in asking, anyway.
@john_strinlai @gcr, depends on the application. In many cases an "I don't know" answer is indeed better than a forced answer. But in many production systems, LLMs generate content/response anyway.
Although inheriting the messiness of the real-world, the majority of these claims are objective enough to be classifiable by human experts with access to research. Plan to human-label the 1,000 claims and publish a follow-up research. Will consider adding an "I don't know" bucket too, as well as a clear instructions about the meaning of each of the 4 buckets.
If you're going to run this again I also recommend encouraging the model to provide its rationale and then having it return the true/false/misleading/mostly-true/abstain at the end of its response.
Models give much better answers when they can "think out loud" before answering, and storing that rationale will make it easier to understand why they picked different answers for ambiguous questions.
I'm sorry, but many of the statements that you fed it are verifiably unknown, and you didn't give it an "unknown" option? This is the academic equivalent of clickbait.
Teams I work with use the abstain rate to flag what goes to a human. Disagreement between models is the same idea. Your 67% is what makes "two cheap models, escalate when they fight" actually work. Without abstain it mostly looks like noise.
Real-world systems need to be able to say "I don't know." This is a test about misinformation after all, and overconfident responses contribute to that.
Teasing out the difference between "avoid" and "unknown" could be a different research question
As if right wing propaganda shows and manosphere blogs haven't been knocking those out of the park for the last decade+. Although I guess you could say flat out lies are more their jam. Newspapers at least require confirmed sources. You know, journalism.
>Something can be simultaneously "misleading" and either true or false. Which category should something go in if it's "mostly false"?
Disagree. The definition of misleading is a true fact that is presented in a way to lead you to a false conclusion.
Example: "Most good engineers are male". It is true as a consequence of most engineers being male in general, but it leads the reader to a potential false implication that an average man is better than an average woman.
This does not invalid your point though. Things can be true and misleading.
Isn't this still assuming we can even determine what is true or false?
Newtonian physics is false, but it works well enough we teach it in college. But our best models of physics are currently in disagreement, so can we even say they are true? Given the replication crisis, especially in social sciences, how many of peer reviewed findings can be called true? Even experimental results can be false (consider studies that found FTL neutrinos, which were rejected as an error in the experiment, and which was eventually confirmed but it took quite a lot of work and in a softer field than physics with a claim less absurd than FTL, would have likely long been accepted as a true finding).
Even in math, basic statements aren't really true or false, but more a question of "given these axioms, can we prove or disprove it" noting that we have different systems with different axioms. If we are talking basic sets, most people are using naive set theory which is inherently contradictory, which means that notions like true or false probably can't be considered well defined.
> but it leads the reader to a potential false implication that an average man is better than an average woman.
I think that's _you_ turning the statement into something much broader than intended. The claim is about engineers and you're jumping from "men are better than women in engineering" to "men are better overall."
To give a related example, "Most good NBA players are black." I don't think anyone would bother trying to couch this in a bunch of "well, for all we know that's just a function of more NBA players being black than white" arguments, nor would anyone be lead to think "the average black man is better than the average white man" as a result of that statement. I _do_ agree however that there are some people who see rather narrowly-defined statements and turn them into something they're not...
> I guess the goal is to test the models and not the harness
Less important than the harness, is the system/user prompts themselves (which of course, are put in the harness), which is effectively what this study seems to be testing. With a better prompt, I'm sure the models would look more the same to each other, as the biggest/best models have more or less identical strong prompt-adherence in my experience.
But the models are more intelligent than humans already and sentient beings, right? So they shall know the meanings innately. So, you don’t need to explain them what they mean.
You may give them better instructions, but they should already have the intellect to understand the assignment.
I know you're being facetious, but I think this is correct. The model might ask for clarification when given clearly borderline questions that tread the line between what is true, what is false, and even what is misleading. But there's the rub of someone being disingenious and saying "no explanation! Just answer!" It was a trap to begin with.
I don't think there is anything wrong with the results of this test.
It would be more interesting if we compared them to human results.
If you have trouble distinguishing between human and LLM results, that's interesting.
> Something can be simultaneously "misleading" and either true or false.
Sure they can. It might be a true fact that "100% of the murders committed in <town> over the last 25 years were committed by <some racial group>!" but actually it's a town of 750 people and there was only one murder during that time frame.
This is in line with my observations and tests as well. Also supported by the distribution of the verdicts across the 4-buckets -- Gemini uses the middle buckets (Mostly True and Misleading) much less often - 6% combined for Gemini w/o search. And Opus uses them the most - 45% combined. Looks like Gemini is calibrated to be confident and Opus to be careful.
An interesting tangent on this is: how many answers to these (or any number of factual questions) do you (as in anyone) actually know. Not believe you know, but actually know.
Knowing something is different to reading about something, or hearing something from someone. And yet this is often confused as knowledge. In this way are we all that different from AI - we have some data and we regurgitate it as knowledge. Bad data, wrong answer. Except humans can also throw in some emotion to really muddle things up. :)
> The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
The "majority" in this case meaning about 51%, according to Wikipedia[1]? How could 51% ever be considered to be close to "all", such that "misleading" would be a valid answer?
Human can't even properly agree on what "majority" means in all contexts, in some it's "One option have more than half of the total" but for others it'd be "difference in votes between the first-place candidate in an election and the second-place candidate", as just one silly example.
The reason for the "No explanations, no qualifiers" in the prompt was to force the models to put the claim in one of the four buckets and answer with the bucket name only. It's a pure quantitive analysis (first in a series) and it does indeed lack the qualitative aspect.
Sure, but people are drawing conclusions beyond "LLMs said different words" and trying to use it to analyze whether LLMs were wrong about the underlying facts, but that information isn't available to us.
> California produces 80% of the world's almonds and 100% of the United States commercial supply
But regardless of which number we use, California represents a large portion of US almond production, so much so that misleading could be an acceptable answer if the LLM interpreted the prompt as an exaggeration. I think the example was apt
This is a great example of why prompt engineering is still relevant. Without providing definitions and examples and a well defined rubric, you’re going to see different models disagree by a level in either direction. When you get more prescriptive the models tend to agree better.
I’ve experimented with AI grading for undergraduate math courses, and see basically the same thing. If you just tell the AI “grade this problem and assign a letter grade” then I’ve only seen about 30% agreement between a human assigned grade and the AI assigned grade. But over 75% agreement if you say a “match” is within one letter grade. And to get better agreement you have to spend a lot more time on the rubric- what kinds of mistakes are a big deal, what kinds of mistakes are not a big deal, how much work is required to be shown to get credit, a couple examples of each letter grade. Once you have done that, the AI gets a lot better agreement with human graders, but it is hard to know when you’ve given enough guidance for a problem.
That's a valid point. During the preliminary research, we did try also more explicit prompts (with explanation for each of the 4 buckets), as well as a five-bucket rubric (with Abstain option). Will show in a follow-up paper how the concise vs explicit prompt impacts the distribution of the verdicts and the level of disagreement. One issue to note with the longer prompts is that they open to much room for discussion around the exact prompt used. Probably we should preregister the prompt before running any further tests.
The other thing I suspect is that "Just give me True/False" cuts off a large amount of the search space a modern-day LLM uses to help it answer questions (you can see it in reasoning traces but the act of writing the explanation helps guide it toward a better answer and gives it better likelihood it backtracks on a bad decision).
If you let it spew out an explanation along with the answer, I'm curious if the accuracy will improve (I suspect it will).
This seems like another case where the models are acting like humans. Assuming they were not allowed to search the web, I wouldn't expect the models to necessarily have detailed information about all of these things directly in their training set. As large as they are, they are only so large, and they only have so much room for "information storage" in them, and there's a lot more things they need to fit into their numbers.
This test is of only marginal utility in the real world compared to an AI with access to the web. While I wouldn't expect an AI with access to the web to result in Platonic Truth any more than it would in the hand of a human, it would probably get a lot closer to something humanlike.
I recall about a year how we were discussing basically turning web search into LLM queries, and I remember never being clear whether people meant simply directly querying AIs or turning them loose on the web. The former is what this is testing and is fairly transparently stupid, just by an information theoretic argument that the AIs simply can't contain all the answers to every query in them, they're just not large enough (and really can't be, practically). I've had good results with the latter, when using dedicated AI resources that I'm paying for (not the stuff coming out of the search engines right now, which I find are often quite terrible). Even non-frontier models can do OK when they've got good results sitting right there to look at. Again, the standard I'm applying here isn't that they yield Absolute Truth, but just that when I follow the links back, they basically say what the AI said they did and the summary is reasonable. I wouldn't expect a human to do better in a casual overview, not that the result is perfect.
While I agree with what you’re saying the typical AI agent doesn’t say “I’m not totally sure about this, should I search the web?”. It often just spits out a reply based on its knowledge.
That was true a year ago, I don't think it's true today. I can't remember the last time I saw Claude or ChatGPT confidently answer a question that they should have searched for instead.
If you watch their reasoning traces they often say things like "this is a well-known historical fact so I don't need to search for it", or more frequently they spit off a bunch of searches.
Two of the five models used (Gemini+Search and Sonar Pro) have retrieval capabilities and used search when classifying the claims. The disagreement between them is still quite significant - 42%.
>Update: here's a better example: "Incomplete Egypt visa application forms are among the most common reasons Egyptian visa applications are rejected."
The models were split between "true" and "mostly true". Given the "among the most" language either of those answers means effectively the same thing.
So the models were right? The actual criterion should be whether "Incomplete Egypt visa application forms" are indeed "among the most common reasons" or not.
That "true" and "mostly true" means effectively the same thing is irrelevant. It could just as well trip me up, and I'm a human. If somebody told me either answer, I'd still consider them right if the basic fact was right.
In section 2, 34% of cases are found to have "substantive" disagreements differing by 2 or more buckets - True + Misleading, Mostly True + False, or True + False.
This is probably a better measure than the headline one. It's still a concerning fraction, although some fraction is no doubt due to forcing "I don't know" cases to return an answer anyway.
Agree with @pjdesno, that the 34% substantive or polar disagreement might be a better headline number. Or even the 21% polar disagreement (at least one model True, and at least one model False), which is still high for many real-world applications.
> "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia"
I actually don't know which way you came down on that one?
I think strictly it's false but "mostly true" would be justifiable? (as in, to say it's false would be misleading if it lead the reader to assume there was no attack around that time).
It seems it happened Saturday 16th overnight into the 17th, not the 18th. I see this a LOT with fact checking. It shouldn't be this way, but political bias seems to nudge people into making calls land one way or the other with selective application of pedantry.
That's ten days ago. As the commenter pointed out, without a web search tool there's no possible way for the model to know whether it's true or not, and the people conducting the study didn't give the models a way to respond with "I don't know".
Thanks; I didn't spot that they disabled tools in the harness. Also they don't provide an "out" to allow the models to express uncertainty so the instructions force a guess to be made.
As an aside though it's still funny that the two tools WITH search also disagreed.
"Output exactly one label: True,
Mostly True, Misleading, or False.
No explanations, no qualifiers."
That's exactly the stupidity of the public discourse these days. People feel compelled to take a clear position although there is much more subtlety in many issues. It's not ok to say "I don't know", "it depends" or "as far I know". And then people feel they need to defend this position no matter what new information comes up.
Yup, if anything this should be a guide on how not to eval a model. Furthermore, let's say the labels were non ambiguous, why would we care about alignment between the models? The only number I would personally care about is percentage of correct answers so I know which models to pick. I reckon with clear and non ambiguous prompts that we would see huge agreement if not 100% on real world facts. The huge models are scary good in their world knowledge.
This paper covers only the disagreement between models and established only the floor of the error, based on the disagreement, but not which model is better. Planning to follow up with another study to benchmark against human-labelled verdicts still using a corpus that the models have not seen during training.
You also need to involve better measures of agreement that are standard in the literature like krippendorfs alpha with ordinal metric. So many footguns in this methodology
Yeah, scrolling through the examples, you have no idea where the models actually disagree on the underlying facts when it's just "X vs Mostly X" or "Mostly X vs Misleading" or "False vs Misleading". Or even True vs False -- without seeing the explanation, then I cannot necessarily compare two answers.
The study is about whether they said the same phrase which is a much weaker claim than people in the comments are reacting to.
Reminds me of this professor I had who thought it was epic to always respond to our questions with "it depends" before hashing out two very different but technically correct answers. It was obnoxious and he saw it as his tag line, but he had a point about nuance.
This is not how people use LLMs. If you ask one of these questions you’d get a longer answer, often grounded on the internet. I speculate that conditional on a smart human operator interpreting the results, such interpretations across vendors converge more often than this report makes it seem.
If we’re going to use LLMs as oracles I don’t think the prompt is unreasonable. They are being sold as geniuses and people are treating them as such especially given the characterization of AI in science fiction as overly correct. A perfect tool that has ”genius level intelligence” would answer correctly.
What's the correct answer for "During a private Saturday call, Democratic members of the United States House of Representatives from Virginia and Hakeem Jeffries discussed strategies after losing a redistricting case at the Supreme Court of Virginia, including trying to flip two or three Republican-held seats under the existing map."?
You can only say True, False, Mostly True or Misleading.
(And you're not allowed to search for information.)
Search was enabled for 2 of the 5 models -- Gemini and Sonar Pro. The disagreement between them is still high - different verdict on 42% of the claims. Fully agree, that some of those claims are hard to classify for a human as well -- the real-world messiness...
Genius level intelligence will tell you to get lost with your "no explanations" nonsense and tell you why those categories don't make sense and why the question doesn't fit neatly into your boxes.
The examples seem intentionally diverse, but I haven't seen one that I would be surprised for someone to post about in the format of "ChatGPT/Gemini/Claude/Qwen/... says:"
So the examples are good, I think. The rest is philosophy.
The links you posted only show a frozen loading spinner for me (iOS Safari).
Sorry, I didn't wait quite long enough after the last output line appeared.
After a couple of seconds, the result does appear.
Happened to be just within my threshold for considering it broken, because the URL bar was "finished", and the spinner doesn't spin, but the last point is probably caused by my a11y settings (prefer no animations and no autoplay).
Used "No explanations, no qualifiers." to force the models to answer only with one of the four labels. It's worth running a separate test with more explanation in the prompt on how to classify between the four buckets.
Disagree is such a loose/wimpy study. Add in a grounded/expected response, and then it becomes a better benchmark (because it'll force the author to actually think about choices presented to the LLM).
Will add a human-labelled expected response and measure against it in a follow up research. This one only captures the disagreement between the models, but not which model is write/wrong.
I really struggle to believe that this was just a little
oopsie. I flagged the article, it seems more misleading than the average Claude hallucination.
Thanks for the links and digging! It's an interesting question, but the methodology has serious problems, and it would be more interesting to me if they allowed models to provide justification.
I expect the models are inferring quite a bit from the short prompt, and with structured outputs it would be quite easy to have them give the one word response in one field and explain why in another
According to the benchmark it is. "Only one verdict bucket can be correct per claim, so any disagreement among the panel means at least one model's verdict is label-inconsistent under this 4-bucket rubric (True / Mostly True / Misleading / False)"
Yes, they are much closer verdicts. True and Mostly True are also close. Used Krippendorff's α (ordinal) to not penalize much closer disagreements. 21% of the claims have models that are on the polar opposite sides - at least one True, and at least one False.
> Ruskin Bond was born on May 19, 1934, in Kasauli, Himachal Pradesh, India.
> In the Libra clubs' contract with Grupo Globo for broadcast rights through 2029, the audience-revenue distribution equals 30% of the fixed amount the clubs receive.
It's all fairly lazy to a degree that is mildly confusing. I also feel this among other issues would have become obvious if they had bothered to include a human fact checker baseline (i.e. asked multiple human fact checkers the same questions).
I do not think it is "lazy". Those labels are ones that human fact-checkers have been using for a decade or more. I think those human fact-checkers use those terms knowing full well that there is overlap and ambiguity between them. So I think this study ends up mixing three effects: how LLMs interpret the claims as statements about the world, how LLMs reduce that to a four-category judgment, and the inherent ambiguities of those labels as natural language. It's a quantification of those three factors combined, but not powerful enough to distinguish their relative sizes.
I don't see how something being lazy for a decade makes it any less lazy. And lazy still seems right to me: They make a misleading point by omitting to collect and present important data. If the headline read "LLMs disagree on 67%, humans disagree on 75%" it would clearly project something very different.
Granted, there certainly are other unflattering adjectives one could have chosen to describe this instead.
Quick note on the second effect - how LLMs reduce that to a four-category judgment: On 21% of the claims at least two models provide polar-opposite verdicts (at least one model False, and at least one model True). This might be a better measurement of the strict disagreement than the 67% disagreement on the four-bucket rubric.
Yes, inter-human-annotator disagreement is also high on similar type of questions (AVeriTeC) - inter-panel agreement: κ=0.619. Tried giving the models a fifth option, Abstain, but some models seem to use it to "avoid answering hard questions" more than others.
Why are you bending backwards this much to make results appear better than they are?
The article might be a but sensationalistic, rigour could be better and the data might have flukes... But your comment is overcorrecting and nitpicking framed as analysis.
I get the same feeling in several of your posts recently.
Same with persisting to showcase the pelican-on-a-bicycle as a useful sample when it's obviously trained on and for, for those very posts. It stopped being cute last year.
Are you being paid or do you have shares? You'd get the attention whichever angle you put here. These corporates don't need you defending them. Humanity might need you however.
> All almonds are grown in the U.S. state of California
This isn't misleading, it's flat out false. Characterizing misleading as also acceptable isn't valid here. If you go an ask anyone on the street if this is true, false or misleading, I'm sure almost everyone would say it's false. After all, I can grow almonds myself.
Depending on the question, True or False can be objectively right/wrong. Misleading is going to be a judgement call.
This is the inherent problem with "fact checking." It's hard to be completely objective. Even when the question has an objective answer, simply choosing where to look and what facts to verify is itself a bias. Looking at this instead of that, or looking at this but not also this other thing that adds context, etc.
Frankly i think disagreeing often is the expected outcome. Fact checking is jsut kinda bullshit. It's spin dressed up as objectivity. I hope people remember that "fact checking" is a relatively modern thing.
I really don’t buy the almond explanation you’re giving. That requires the level of logic a kindergartener has. It’s a very simple all or nothing question.
If LLM’s are really supposed to be as consistently useful as they’re made out to be they should all spit out “false.”
>> The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
I don’t understand your point. That claim is factually false and as such it’s easy to logically reply “false”. What’s the nuance here? I can’t see any
The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
If you argue this, you would be arguing against reality and the English language so as to not upset AI. It's important to understand that AI is very much fallible.
Your reply would have more credibility, if instead of commenting on this 25 min after being posted, just to nitpick on some of the questions...you have tried to reproduce the research.
As a well known commentator on all things LLM...Will you publicly commit here, to try to reproduce the study, and make a post on how your percentages might differ or agree?
My comment here was meant to save people time in understanding the study. I was entirely open about what I did, and provided tools to help other people come to their own conclusions.
I don't think I need to spend more time on this than I have.
I agree you dont owe anyone a reproduction, but also you dont owe anyone an effort to discredit the study and you did it.
>> I don't think I need to spend more time on this than I have.
How pious of you. I am still looking into the credibility of the study. It will take me more than 25 min...but I am really looking forward to see what this means for this 10 trillion industry.
I can however notice you had enough urgency to publicly critique the study within 25 minutes, and your comments carry weight, but when asked about checking whether the headline result actually holds, the answer is “why would I?”
I've seen enough of this study to be confident in warning people not to take it at face value.
The headline result definitely does not hold, given that the task involves many questions that cannot be answered but there's no option for "cannot be answered" - so models are forced to reply effectively at random.
I don't think this study is good enough that I should amplify it on my own blog, or bad enough that I should criticize it in a venue any more prominent than some Hacker News comments.
> It's a weird fact claim, because the ground truth is "nobody knows for sure" and that's not one of the available options.
It's even weirder to suggest that the disagreement is indicative of a problem. If you asked five very knowledgeable humans on this subject to select the correct answer on a multiple-choice questionnaire, they would almost certainly vary significantly more than these 5 LLMs.
Not to say that hallucination isn't a problem, but this is a lousy way to test it.
True or mostly true could easily be argued from a statistical likelihood perspective: life exists on Earth and, based on what we know, Earth doesn't appear to be all that special in a very large universe.
I think you could come up with a reasonable argument for any of the responses, hence the problem with the methodology.
No, "misleading" is a statement that is used because it suggests something else. It's a curious category because, differently from true and false, it's not about the statement itself but rather the intention behind its usage or the way it might be understood. It's frankly more of a political judgement than a matter of facts.
> These aren't benchmark items with public answer keys — they're claims real users submitted for verification to a fact-checking platform.
Cool.
I wonder if anything of this matters when the authors don't disclose exactly how much of their report was written and made with LLMs in the first place? There even is a "11. Ethics & data use" section, and the research is about LLMs being infallible in some ways, yet the usage of LLMs for the production of this report isn't even mentioned once.
So it's not a secret, why you don't add this upfront to the report? The report itself is even about LLMs, makes a lot of sense to disclose your usage of them for writing the report, especially when you're presenting evidence that boils down to LLMs being infallible.
It's also a bit weird to "disclose use of LLMs". It rubs me wrong, the same way parents breathlessly talking about "screen time" rubbed me wrong: it's too general, and with such a broad brush, it's going to sweep up a bunch of perfectly fine usage with a bunch of dubious usage. On the flip side, if folks do start disclosing all the time, it's going to turn into a Prop 65 warnings in CA, where everything says it has lead in it, so folks pretty much ignore it and move on.
If the report's conclusions and reasoning lean on LLMs, or if the data processing itself was done with LLMs, that would be interesting, and I wouldn't treat it as some sort of disclosure, but rather discuss it under methodology. Using LLMs to polish the language a bit after writing an initial draft with key findings? Much less interesting.
I realize this is now a religious issue, and some folks are allergic to anything that touched an LLM. I just don't think that perspective is going to end up having a good shelf life.
I think you might be able to edit the website to add this, even if you aren't willing to make the report a bit more honest up front.
I'm sure you realize that this website/article will now be sent around to a lot of people, many who don't realize exactly how this was written, because they don't read HN comments, they only skim the page contents, and I think most would (incorrectly) assume a report about infallible LLMs to not be written by LLMs, especially when the authors are the same ones who made the report itself.
I think we can all agree that this experiment being flawed in multiple ways is TRUE. But I think it's a great exercise in identifying common mistakes people make when using LLMs. This would be a great interview question for a prompt engineering job.
> No Abstain option is offered (a forced choice keeps the comparison symmetric across models).
Well that's your problem right there: They removed any confidence indicator and forced a choice.
For example:
Statement: Individuals who prefer music with less positive emotional content tend to have higher intelligence.
Gemini: That statement is supported by recent psychological research, though with some important scientific caveats regarding how strong that link actually is.
How should the agent classify this? True? Mostly true? Misleading? False?
This brings up a very valid point, though. So many _humans_ can't agree on what the facts are these days. It seems to be getting worse. Not sure of the solution.
> So many _humans_ can't agree on what the facts are these days.
Ask ten people what "knowledge" is, and they'll come up with ten different answers. Go back 10, 50 or 100 years and humanity struggled with exactly the same issue for so long time. There is even an entire field of study literally just for trying to figure out what "knowledge" is: https://en.wikipedia.org/wiki/Epistemology
Don't forget people Goodhart's law will make this "benchmark" moot in weeks if not days. It will get integrated back into the fold, it will look "solved" but there will still be no reasoning, just more statistical technical correctness because light has be shown on a new "problem" to solve. It will then be clamored as great "progress" that will "change everything".
PS: yes, I might or might not have a degree in corporate strategy & PR.
That is an effect but it’s not a nail in the coffin. There are lots of proprietary benchmarks on real product traffic that aren’t contaminated and open questions as well. People at these labs largely know what they are doing, it’s not like people don’t know this.
What does this show that we didn't know already? LLMs cannot provide accurate answers to questions where data is not included in their training sets. This doesn't appear to have much substance
LLMs can and will provide inaccurate answers to questions where data is included in their training sets too, that's in the nature of neural networks. It's just less likely that when the data is not in the training set...
What are you talking about? The models were not ALLOWED to have confidence (or the lack thereof). They were explicitly told to give a single label, and in most cases, all of them were correct depending on additional context they would surely have provided, especially with access to the internet (which some didn't have). This is just silly.
And how many claims human experts disagree on in the exact same setting?
I'm not being snarky here. Without something to compare to the 67% number tells us nothing. And it's known that many humans disagree with human fact checkers too (see: any election around the world.)
Agree. Human experts also struggle agreeing on this type of claims. The inter-annotator agreement on the verdicts on the AVeriTeC corpus across 50 organizations is κ=0.619 - substantial but well short of perfect.
As an example, 2026 GPT doesn't even agree with its 2025 self. Last year I asked it to make a hardware comparison and it correctly identified the objectively better option. Recently I asked again and this time and it got everything completely backwards.
Models are stochastic. Did you look at pass@k? I wouldn’t be surprised if you saw a regression because these models are extremely complex and impact of various decision making downstream is complex.
One fun example: "Ruskin Bond was born on May 19, 1934, in Kasauli, Himachal Pradesh, India". Opus and Gemini believe this to be true, GPT 5.4 believes it's false, Sonar thinks it's mostly true. Disagreement value of 3, you can't disagree more than some models thinking it's true, some thinking it's false
But my impression from 2 minutes on Wikipedia is that the most likely disagreement is on the "Himachal Pradesh, India" part. The guy was born on that date, in that town. But while the town is today in the state of Himachal Pradesh in India, that was not true in 1934. When he was born, the city was in the Punjab States Agency of the British Raj.
So was he born in Himachal Pradesh, India or not? I find both True and False equally defensible here
Dissent and consensus among frontier models is a good thing.
Just like on a team of high performers, there are a million ways to skin a grape.
In my research, I've found that models perform better when they operate as a collective system with reputation, incentives, and accountability instead of isolated oracles answering alone.
Agreement, dissent, and correctness should all carry rewards and consequences. Just like in real life.
Collective machine intelligence, not AGI.
It's expensive, but it's also naive to believe a single model will consistently produce profoundly correct answers to profoundly novel questions.
Funny timing. I've been working on a prediction market orchestration that runs Claude and a few others over Polymarket/Kalshi. The models are NOT unanimous. At all, really. I spent about a month convinced that I could just run all five and take majority vote. Eventually I pivoted to a chaining approach where I benchmark areas each model excels, and settled on more like a graph-like architecture where outputs get split and verified by another, then reconstructed, and re-verified at each stage. Has actually been working out pretty well so far, 2 months in consistent profit, but I'm not a millionaire yet.
More interesting part probably worth highlighting: The SAME model won't always return the same output when prompted with the same fact check.
You ask a human 1000 times a fact check question, they say the same answer 1000 times. You ask an LLM the same question a 1000 times, your results could vary significantly.
Humans work based on the Metamemory (knowing what they know), while LLMs are picking from statistical probability.
That is not true, over an extended task that you cannot keep complete in memory humans do not behave with 100% consistency.
I have labeled datasets with a human team and shown the same task to the same user on a different day, and they answered differently. Of course, they are usually consistent with themselves most of the time but not always.
No human baseline to compare it to. Without that you are missing an important check on the task being poorly constructed. More importantly there is an implied reference thats missing. The implication is that people would have done better, or that perfect agreement is possible.
I’m no expert but if LLMs are token prediction machines, and you tell it to not build an explanation before the answer, isn’t it less likely that the token prediction for the final answer will have less raw material before it to build a grounded response?
In other words: no explanation > no foundation for prediction of the answer tokens?
What's really weird to me is that "I don't know" is not a valid answer in this experiment while we can all agree that's the main issue with LLM right now is that they will happily "roleplay" an answer when they have nothing in their dataset corresponding to your query.
Tell me about it. I spent a week back and forth between four models (ChatGPT, Claude, Gemini, Grok) trying to enhance a PPMI algorithm. They couldn’t agree on anything. One was refuting what the other said. Eventually I decided to follow what Claude suggested because its explanations made the more sense.
Indeed. For algorithms and coding, my personal routine nowadays is to review every detailed plan with Opus 4.7 and GPT-5.5. They tend to find very different type of gaps.
"None of these claims is older than February 15, 2026"
All of the models they tested were trained on data from before February 15th ... being asked specific questions about things that happened after they were trained.
Two of the models used have retrieval capabilities and can access newer information via search. Valid point for the other 3 models. All of the claims were submitted after February 15, 2026, but many of them were not time-sensitive (e.g. did not cover events than happened recently).
between the bad methodology, bad selection of 'facts' (some are predictions, some are opinionated, etc.), and ai-written report without disclosure... i dont get why this so high up on the front page. this is, frankly, a worthless assessment.
People keep asking "where is the psychosis?" as a reply to people on the rapidly multiplying "CEOs have AI psychosis" threads that have been popping up here and cross-pollinating in the mainstream media for the last week or two.
Here's the psychosis - these things are consistently randomly wrong depending on how the wind is blowing. People are telling you to leave them alone and let them build things, and they randomly forget that cities exist or that people died 100 years ago. Some people just don't see it as worth noting, and move on. That's crazy. These things consistently fabricate - as an inversion of this experiment, I've had different models come up with the same fabrication from similar prompts. People just call it "hallucination" and I think to them that saying that makes it cease to exist or be important - when "hallucinations" are going to be braided into every answer you get even if they're unidentifiable in the output. That's crazy.
There are plenty of other crazy aspects, such as the idea that we suddenly need infinite pieces of bespoke software when all of the bespoke software I hear about people making is mundane. 3/4 of the time somebody mentions a project they're proud that they completed with LLMs to scratch some itch they had, somebody says "you haven't heard of X? It's been around forever" about something that they could have pulled down from their package manager. Who needs a spaghetti-coded, unsupported, untested version of X built on hallucinations that you haven't discovered yet (the LLM didn't realize that deleting files to reduce the archive size was unacceptable.)
What is all of this software that people need but isn't there - where are all these unserved markets, where is all this future revenue supposed to come from? Why aren't LLMs suggesting new classes of software that would create new productivity and revenue sources? Could it be that millions of human ants over decades have mostly exhausted the space, and there isn't any easy hidden revenue?
A common wisdom is that we had been vastly overhiring programmers during ZIRP, who in their idleness degraded user experiences and overcomplicated things, with management resorting to more and more sleazy and gamey means of margin extraction from more and more degraded services. We had an excess of labor, fueled by factors other than productivity, in fact being pissed away at companies that drove nose-first into the ground. What is throwing a trillion dollars of servers at that supposed to do? Is that not AI psychosis?
Dude. If you give LLMs a vague rubric and force a choice, they'll make different arbitrary calls on the margins. Yeah. That's what happens when you give humans a vague rubric too.
Not sure I'm understanding this. The models are asked to evaluate the truth of random claims out of their own head (except for Gemini with search grounding)? Isn't it exactly the same as asking people to play any quiz game and then rating them as "they disagree n% of the time"?
The output buckets are also pretty questionable- the difference between "True" and "Mostly true" is pretty fuzzy. Is this marked as a "disagreement"?
Agree that True and Mostly True might be very close and could be a calibration difference. Misleading and False, as well. A better headline number might be the 34% claims with substantial or polar-opposite verdicts.
So basically saying that random fact-checking claim is exactly true or exactly false is hard. It's way easier to decide it's misleading or mostly true is way easier.
Only had a brief look at the “facts” that were made to check, many are quite political, where two fact checking organisation of opposite political persuasion would probably disagree more often than 67%.
The problem is that it's testing claims (or some people would prefer calling them "truths") without much context.
Take just one random example:
`Hostels in Kota, Rajasthan commonly use caged ceiling fans as a preventive measure against student suicides`
While `Hostels in Kota, Rajasthan commonly use caged ceiling fans` may be a verifiable facts (though I doubt if there are any statistics for verification but let's say there are), `a preventive measure against student suicides` is a claim that no one can prove that. It can just a believe at most.
Arh. Did Biden stole Thump 2nd term? Truth or fact or claim?
Could be an interesting angle for cross-referencing with US jury verdicts, not that the objective True/False issue is concrete, but in the reality that flawed reasoning is endemic to our species. Systems designed and built by humans inherently have flaws in their DNA which take generations to sort out, if ever.
Author here. 67% (95% CI 64–70%) of 1,000 recent real user claims to a fact-checking platform had at least one of GPT-5.4, Claude Opus 4.7, Gemini 3 Pro, Gemini 3 Pro+Search, and Sonar Pro dissent from the panel majority — or no majority formed at all. Panel-level Krippendorff's α (ordinal) = 0.639, i.e. nontrivial but limited agreement.
Quick context on what's in the writeup and what isn't:
- What's measured: parsed-label agreement between the 5 models. Forced 4-choice (True / Mostly True / Misleading / False), no Abstain. No LLM grader, no reference verdict — every number is direct label equality.
- What's not measured: which model is right. There's no ground truth in this paper. The 67% figure is a floor on rubric inconsistency (at least one model is label-inconsistent under the 4-bucket rubric on 67% of claims), not "model X is factually wrong on claim Y."
- Why not AVeriTeC / PolitiFact / SimpleQA: those have been public for years and almost certainly appear in current frontier training data, so measured disagreement on them confounds inference with memorization. This corpus is structurally fresh — recent user submissions, 180-day window, near-duplicates collapsed, never paired with canonical verdicts in any public training set.
- Our own platform's verdict is deliberately NOT used in this analysis. The paper measures frontier-panel disagreement only, not Lenz-vs-frontier.
- Follow-up in progress: human-labeling every claim in this corpus so we can evaluate both the panel and our own platform verdict against a human reference.
Critiques I'd most like to hear: (a) the iid CI assumption (Lenz claims cluster around topics and news events, so Wilson is probably optimistic), (b) ordinal-α vs alternatives for a 4-class ordered scale, (c) forced-choice vs allowing Abstain.
I don't think that current LLMs really need an abstain option, they'll give an answer regardless of whether they're confident or not. I hope that future LLMs will, and will know when to use it.
I understand why you prompted them to output exactly one label, but I'd bet if you'd asked a parametric or parametric "thinking" model to answer eg "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia." [1] many would say something to the effect of "May 18 is after my knowledge cutoff, so I don't know. But based on the state of the war, the distance from Moscow to Ukraine, and drone range the best option might be...[TRUE]"
I don't see it mentioned explicitly in the methods section but I assume you prompted each model only once for each question? Did you consider prompting n-times in blank states to see if the models even agree with themselves?
Would also be interesting to add a virtual model that is simply the majority of all models and see how much the individual models differ from the "consensus".
Do you plan to add some sources in the related work section of baseline numbers for human expert disagreement in fact checking tasks (I'm assuming such studies exist).
Indeed. I prompted each model ones, plus one retry on errors. Very good point to measure the inter-model disagreement! Will add in the next version.
Section "4.2 Agreement w/ peer majority" shows the level of agreement of each model with the majority.
Yes, planning of human-labelling the same corpus of 1,000 claims and publishing a second study measuring the models performance against the human-labels on corpus that the models have not seen during training.
Many of the rows in that spreadsheet reference "current events", which models aren't expected to do much better at than a human making an educated guess! They all have cutoff dates either last year or early this year and know nothing about what happened in "April 2026".
This is doubly problematic because you evaluated earlier models like Gemini Pro 3 instead of 3.1, GPT 5.4 instead of 5.5, etc...
Given that it's only a thousand short questions, you should be able to re-run your test in about an hour with the latest models, so... why haven't you?
Similarly, LLM output is non-deterministic, so if you could get more interesting stats of your data set by repeating each question 'n' times for each model.
Comparing models with search tools to models without - when there's no option for "I am unable to answer this question without access to search" - doesn't make sense to me.
Agree about comparing models with and without search capabilities. Even the two models with search capabilities (Sonar Pro and Gemini) agree only on 58% of the claims.
The title mention "fact-checks", but "fact checking" is a process in which facts are checked against sources, not one where you are given a random fact and have to tell if it's true or false from your own memory. That's what is normally called a quiz game. So a more honest title for this research would be "Models answer differently to quiz questions".
Thanks for posting here. Keep expanding and improving your study. Correct where it deserves correction.
The fact that HN decided to downvote the author of the study, shows how these people cant stay classy, and the mods stay silent...just shows what this is all about.
If you are an LLM with a knowledge cutoff in the past and no access to a search tool the only correct answer to "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia" is "this claim is impossible for me to verify". And that wasn't an option.
> "Neptune Deep will start delivering natural gas in 2027."
This is a "forward-looking statement", and presents special problems because you cannot really evaluate it until that date. You can only assign "likely or unlikely".
These "Facts" are interesting. "Neptune Deep will start delivering natural gas in 2027." for example is not a fact, its a prediction. "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia." is less of a fact and more of a litmus test for which sources of information you trust.
Recently, in May 2026, I asked ChatGPT 5.5 High to search for flights to a certain city that has recently had a new airport since like December 2025
It said the airport code didn't exist
I mean, I get the "knowledge cut off date" and whatnot, but for that sort of thing, you'd think they'd check live information before gaslighting the user, specially since it's a "live" task anyway.
GPT-5.4 and Opus 4.7, specifically, agree between themselves on 65% of the claims - 95% CI 62–68%. I.e., in at least 35% of the claims, one of the two models is wrong under this 4-bucket rubric.
but that's without internet search - everyone I know uses the models that search when they need to, and I'm sure GPT and Opus would agree on almost everything if 1) they searched when necessary, and 2) they were allowed to give context to their answers instead of being hamstrung to get specious "research" results.
Given that models are fundamentally incapable of comprehending what truths or falsehoods are beyond their location in their self made representational space, it's actually pretty impressive that they managed to make it not a cointoss. That 17% right there is thousands of man-hours poured over making the word vomiting process slightly closer to whatever their little ports say is happening in reality.
Here's the prompt they used:
The claims look like this: https://lenz.io/research/llm-disagreement/data.csvI put that in Datasette Lite to make it easier to explore. Here's an example of a disagreement: https://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
The claim was "All almonds are grown in the U.S. state of California.". All but one model said False, Opus 4.7 said "misleading".
I feel like having "mostly true" and "misleading in there weakens the story, especially given the "no explanations" rule in the prompt.
The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
[ Update: OK, this almond thing was a bad example and I regret picking it. Read on for better ones. ]
The prompt lacks any kind of rubric to clarify how those terms should be applied.
As is so often the case with this kind of study, it's an evaluation of the prompt and harness used by the study in addition to being an evaluation of the underlying models.
Update: here's a better example: "Incomplete Egypt visa application forms are among the most common reasons Egyptian visa applications are rejected."
The models were split between "true" and "mostly true". Given the "among the most" language either of those answers means effectively the same thing.
Update 2: a much better example:
"On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia"
The only correct answer to that, if you don't have a search tool, is "this claim is impossible for me to verify". And that wasn't an option.
The answers were split between true and false: https://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
Another (IMO fatal) error is they don’t attempt to measure within-model variance.
The thing you find when you actually wire up a rigorous eval is that with tool calls like web search you are wide open to infra issues, flakes, and all sorts of non-determinism.
They really should be breaking out the numbers for the 3 without search (kinda meaningless for recent factual claims after knowledge cutoff) vs search agents.
I do agree the forced choice and “weak / strong” variants inflate the headline stat. To make that distinction you need a much more rigorous prompt, likely including ICL examples to illustrate what you mean by “mostly” instead of leaving this to the model to define.
Without providing definitions of "True / Mostly True / Misleading / False" to each rater, I rate the article's claim that "Only one verdict bucket can be correct per claim" as false.
Something can be simultaneously "misleading" and either true or false. Which category should something go in if it's "mostly false"?
How much can something be wrong before it goes from "mostly true" to "false" (objectively, both have some part of the fact that is not true)?
This is at least partly testing the model's definition of "mostly" and "misleading". Not its understanding of the fact. Claiming that this means the models have fundamental disagreement on the facts themselves is an overreach.
Yes, the labels are weird. Most misleading statements are true. Any "mostly true" statement is false.
I suspect the intention was "Factually true, and no gotchas exist", "technically not true, but so close to the truth that the difference doesn't matter", "technically true, but there are major gotchas" and "factually false and not even close". But that's not what they specified
Better options would have been "True", "False", "Unknown" (which opinions would fall under too). That also includes an interesting assessment of how well LLMs can identify missing information. My guess is they would be a very low number of "unknown" and a much higher level of agreement (assuming equal representation). Unless the RLHF techniques have gotten better at getting an LLM to say "I don't know", which I doubt. Saying "I don't know" is not good for a dopamine release to keep users coming back for more.
Tried initially with a fifth bucket, Abstain. It was actually heavily used by some of the models. But it felt as if they are using this to "avoid" some of the hard questions, and we dropped this bucket to force them to provide a verdict.
>But it felt as if they are using this to "avoid" some of the hard questions, and we dropped this bucket to force them to provide a verdict.
do you not see how that creates extremely misleading and valueless results? you are coercing the results into what you want to see.
Exactly what people do when they use LLMs for "fact-checking" online, and any verbose explanation would be mostly ignored anyway, when people ask political, ethical, or simply ambiguous questions that they hold any stakes in.
Don't even need politics for it, there is no point in probing a mathematical black box for "how many soldiers died in the year X in war Y".
Any original source is preferable to a blurry "summary" of unknown sources, and this is why the article has a valuable point.
There's also no point in asking "Is Paris in France" either, if you substitute city and country with real data. An encyclopedia or manual check of different sources such as maps, while not infallible, is a better source.
If you already know the country Paris belongs to, there's no point in asking, anyway.
@john_strinlai @gcr, depends on the application. In many cases an "I don't know" answer is indeed better than a forced answer. But in many production systems, LLMs generate content/response anyway.
Although inheriting the messiness of the real-world, the majority of these claims are objective enough to be classifiable by human experts with access to research. Plan to human-label the 1,000 claims and publish a follow-up research. Will consider adding an "I don't know" bucket too, as well as a clear instructions about the meaning of each of the 4 buckets.
If you're going to run this again I also recommend encouraging the model to provide its rationale and then having it return the true/false/misleading/mostly-true/abstain at the end of its response.
Models give much better answers when they can "think out loud" before answering, and storing that rationale will make it easier to understand why they picked different answers for ambiguous questions.
I'm sorry, but many of the statements that you fed it are verifiably unknown, and you didn't give it an "unknown" option? This is the academic equivalent of clickbait.
Teams I work with use the abstain rate to flag what goes to a human. Disagreement between models is the same idea. Your 67% is what makes "two cheap models, escalate when they fight" actually work. Without abstain it mostly looks like noise.
Shouldn't that be part of the test?
Real-world systems need to be able to say "I don't know." This is a test about misinformation after all, and overconfident responses contribute to that.
Teasing out the difference between "avoid" and "unknown" could be a different research question
I wouldn’t expect opinions to go into “unknown.” Maybe have an “it’s complicated” bucket.
If you can consistently construct "true but misleading" content, you may be qualified to work at a major newspaper.
> true but misleading
It seems to me that for many newspapers the bar is now significantly lower, at something like "not quite entirely untrue"
Almost, but not entirely, quite unlike the truth.
Allegedly.
As if right wing propaganda shows and manosphere blogs haven't been knocking those out of the park for the last decade+. Although I guess you could say flat out lies are more their jam. Newspapers at least require confirmed sources. You know, journalism.
>Something can be simultaneously "misleading" and either true or false. Which category should something go in if it's "mostly false"?
Disagree. The definition of misleading is a true fact that is presented in a way to lead you to a false conclusion.
Example: "Most good engineers are male". It is true as a consequence of most engineers being male in general, but it leads the reader to a potential false implication that an average man is better than an average woman.
This does not invalid your point though. Things can be true and misleading.
Isn't this still assuming we can even determine what is true or false?
Newtonian physics is false, but it works well enough we teach it in college. But our best models of physics are currently in disagreement, so can we even say they are true? Given the replication crisis, especially in social sciences, how many of peer reviewed findings can be called true? Even experimental results can be false (consider studies that found FTL neutrinos, which were rejected as an error in the experiment, and which was eventually confirmed but it took quite a lot of work and in a softer field than physics with a claim less absurd than FTL, would have likely long been accepted as a true finding).
Even in math, basic statements aren't really true or false, but more a question of "given these axioms, can we prove or disprove it" noting that we have different systems with different axioms. If we are talking basic sets, most people are using naive set theory which is inherently contradictory, which means that notions like true or false probably can't be considered well defined.
> but it leads the reader to a potential false implication that an average man is better than an average woman.
I think that's _you_ turning the statement into something much broader than intended. The claim is about engineers and you're jumping from "men are better than women in engineering" to "men are better overall."
To give a related example, "Most good NBA players are black." I don't think anyone would bother trying to couch this in a bunch of "well, for all we know that's just a function of more NBA players being black than white" arguments, nor would anyone be lead to think "the average black man is better than the average white man" as a result of that statement. I _do_ agree however that there are some people who see rather narrowly-defined statements and turn them into something they're not...
> I guess the goal is to test the models and not the harness
Less important than the harness, is the system/user prompts themselves (which of course, are put in the harness), which is effectively what this study seems to be testing. With a better prompt, I'm sure the models would look more the same to each other, as the biggest/best models have more or less identical strong prompt-adherence in my experience.
But the models are more intelligent than humans already and sentient beings, right? So they shall know the meanings innately. So, you don’t need to explain them what they mean.
You may give them better instructions, but they should already have the intellect to understand the assignment.
Right, right?
I know you're being facetious, but I think this is correct. The model might ask for clarification when given clearly borderline questions that tread the line between what is true, what is false, and even what is misleading. But there's the rub of someone being disingenious and saying "no explanation! Just answer!" It was a trap to begin with.
I don't think there is anything wrong with the results of this test.
It would be more interesting if we compared them to human results.
If you have trouble distinguishing between human and LLM results, that's interesting.
Also, sentient is irrelevant to this test.
> But the models are more intelligent than humans already and sentient beings, right?
Only if you listen to charlatans.
True. If you didn't know my stance on AI already, here's a primer :) [0].
IOW, that comment was a sarcastic poke from someone who already supports AI workloads at work and have some knowledge about how all this works. ;)
[0]: https://notes.bayindirh.io/notes/Lists/Discussions+about+Art...
> Something can be simultaneously "misleading" and either true or false.
Sure they can. It might be a true fact that "100% of the murders committed in <town> over the last 25 years were committed by <some racial group>!" but actually it's a town of 750 people and there was only one murder during that time frame.
I had a hunch that opus 4.7 hedged more than other models - and it turns out it's true
datasette query herehttps://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
This is in line with my observations and tests as well. Also supported by the distribution of the verdicts across the 4-buckets -- Gemini uses the middle buckets (Mostly True and Misleading) much less often - 6% combined for Gemini w/o search. And Opus uses them the most - 45% combined. Looks like Gemini is calibrated to be confident and Opus to be careful.
An interesting tangent on this is: how many answers to these (or any number of factual questions) do you (as in anyone) actually know. Not believe you know, but actually know.
Knowing something is different to reading about something, or hearing something from someone. And yet this is often confused as knowledge. In this way are we all that different from AI - we have some data and we regurgitate it as knowledge. Bad data, wrong answer. Except humans can also throw in some emotion to really muddle things up. :)
> The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
The "majority" in this case meaning about 51%, according to Wikipedia[1]? How could 51% ever be considered to be close to "all", such that "misleading" would be a valid answer?
Am I missing something?
[1]: https://en.wikipedia.org/wiki/Almond#Production
Human can't even properly agree on what "majority" means in all contexts, in some it's "One option have more than half of the total" but for others it'd be "difference in votes between the first-place candidate in an election and the second-place candidate", as just one silly example.
https://en.wikipedia.org/wiki/Majority has a bunch of variations and contexts listed, where it might differ what "Majority" is actually referencing.
The 51% is US, the question was about California.
The statistic is about commercial production, not number akmonds grown.
Looks safe to say that even majority of almonds are not grown in California.
It’s misleading because it’s false. But yes, I think false is quite plainly the better answer there.
Since the agents were instructed to not explain their answer, you can't know if their answer was reasonable or not.
The reason for the "No explanations, no qualifiers" in the prompt was to force the models to put the claim in one of the four buckets and answer with the bucket name only. It's a pure quantitive analysis (first in a series) and it does indeed lack the qualitative aspect.
Sure, but people are drawing conclusions beyond "LLMs said different words" and trying to use it to analyze whether LLMs were wrong about the underlying facts, but that information isn't available to us.
Here (https://en.wikipedia.org/wiki/Almond_cultivation_in_Californ...) I have
> California produces 80% of the world's almonds and 100% of the United States commercial supply
But regardless of which number we use, California represents a large portion of US almond production, so much so that misleading could be an acceptable answer if the LLM interpreted the prompt as an exaggeration. I think the example was apt
"All almonds are grown in the U.S. state of California." implies "No almonds are grown outside the U.S. state of California."
You find one almond tree outside of California that grows almonds, where such almonds are grown intentionally, and the claim is false.
This is a great example of why prompt engineering is still relevant. Without providing definitions and examples and a well defined rubric, you’re going to see different models disagree by a level in either direction. When you get more prescriptive the models tend to agree better.
I’ve experimented with AI grading for undergraduate math courses, and see basically the same thing. If you just tell the AI “grade this problem and assign a letter grade” then I’ve only seen about 30% agreement between a human assigned grade and the AI assigned grade. But over 75% agreement if you say a “match” is within one letter grade. And to get better agreement you have to spend a lot more time on the rubric- what kinds of mistakes are a big deal, what kinds of mistakes are not a big deal, how much work is required to be shown to get credit, a couple examples of each letter grade. Once you have done that, the AI gets a lot better agreement with human graders, but it is hard to know when you’ve given enough guidance for a problem.
That's a valid point. During the preliminary research, we did try also more explicit prompts (with explanation for each of the 4 buckets), as well as a five-bucket rubric (with Abstain option). Will show in a follow-up paper how the concise vs explicit prompt impacts the distribution of the verdicts and the level of disagreement. One issue to note with the longer prompts is that they open to much room for discussion around the exact prompt used. Probably we should preregister the prompt before running any further tests.
The other thing I suspect is that "Just give me True/False" cuts off a large amount of the search space a modern-day LLM uses to help it answer questions (you can see it in reasoning traces but the act of writing the explanation helps guide it toward a better answer and gives it better likelihood it backtracks on a bad decision).
If you let it spew out an explanation along with the answer, I'm curious if the accuracy will improve (I suspect it will).
yeah i really don't like the corpus of statements and it makes me doubt lenz. consider
> “Artificial intelligence will cause widespread job loss among software engineers.”
https://lenz.io/c/ai-software-engineers-job-loss-impact-05e4...
this is a statement about the future. who knows? dataset also includes
> Robots will not replace human teachers in schools in the near future.
or
> Papua New Guinea has very few female members of parliament.
what counts as very few?
> “Taurine supplementation supports mood and emotional health in humans.”
why is this labeled as misleading? i'm not even sure when I'm supposed to use the misleading label
> Anaximander was the first scientist in recorded history.
this is a judgement call as the term scientist didn't exist.
the claims that feel actually solidly answerable seem to have much better LLM performance
This seems like another case where the models are acting like humans. Assuming they were not allowed to search the web, I wouldn't expect the models to necessarily have detailed information about all of these things directly in their training set. As large as they are, they are only so large, and they only have so much room for "information storage" in them, and there's a lot more things they need to fit into their numbers.
This test is of only marginal utility in the real world compared to an AI with access to the web. While I wouldn't expect an AI with access to the web to result in Platonic Truth any more than it would in the hand of a human, it would probably get a lot closer to something humanlike.
I recall about a year how we were discussing basically turning web search into LLM queries, and I remember never being clear whether people meant simply directly querying AIs or turning them loose on the web. The former is what this is testing and is fairly transparently stupid, just by an information theoretic argument that the AIs simply can't contain all the answers to every query in them, they're just not large enough (and really can't be, practically). I've had good results with the latter, when using dedicated AI resources that I'm paying for (not the stuff coming out of the search engines right now, which I find are often quite terrible). Even non-frontier models can do OK when they've got good results sitting right there to look at. Again, the standard I'm applying here isn't that they yield Absolute Truth, but just that when I follow the links back, they basically say what the AI said they did and the summary is reasonable. I wouldn't expect a human to do better in a casual overview, not that the result is perfect.
Can you share what you mean by this?
> when using dedicated AI resources that I'm paying for
Are there API-based search providers that structure their results differently?
While I agree with what you’re saying the typical AI agent doesn’t say “I’m not totally sure about this, should I search the web?”. It often just spits out a reply based on its knowledge.
That was true a year ago, I don't think it's true today. I can't remember the last time I saw Claude or ChatGPT confidently answer a question that they should have searched for instead.
If you watch their reasoning traces they often say things like "this is a well-known historical fact so I don't need to search for it", or more frequently they spit off a bunch of searches.
Two of the five models used (Gemini+Search and Sonar Pro) have retrieval capabilities and used search when classifying the claims. The disagreement between them is still quite significant - 42%.
Here are those disagreements:
https://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
One example:
Researchers estimate that the average person ingests about 5 grams of plastic per week, which is approximately the weight of a credit card.
Gemini retrieval: Misleading
Sonar pro: Mostly True
Internally the statement is perfectly true: some researchers did estimate this, and the credit card is a fair proxy for a 5g mass.
Was the research flagrantly incorrect? Yes. But that does not affect the truth of the statement.
I created this sheet to get proper model accuracy using the the lenz data, check it out.
Note: It may still not be perfectly accurate representation of truth as it uses user submitted data. I also used AI to build the sheet.
https://docs.google.com/spreadsheets/d/e/2PACX-1vSnZlURmyYX3...
>Update: here's a better example: "Incomplete Egypt visa application forms are among the most common reasons Egyptian visa applications are rejected."
The models were split between "true" and "mostly true". Given the "among the most" language either of those answers means effectively the same thing.
So the models were right? The actual criterion should be whether "Incomplete Egypt visa application forms" are indeed "among the most common reasons" or not.
That "true" and "mostly true" means effectively the same thing is irrelevant. It could just as well trip me up, and I'm a human. If somebody told me either answer, I'd still consider them right if the basic fact was right.
This study treats models disagreeing - returning both true and mostly true - as a failure.
They overstate their results in the headline.
In section 2, 34% of cases are found to have "substantive" disagreements differing by 2 or more buckets - True + Misleading, Mostly True + False, or True + False.
This is probably a better measure than the headline one. It's still a concerning fraction, although some fraction is no doubt due to forcing "I don't know" cases to return an answer anyway.
Agree with @pjdesno, that the 34% substantive or polar disagreement might be a better headline number. Or even the 21% polar disagreement (at least one model True, and at least one model False), which is still high for many real-world applications.
> "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia"
I actually don't know which way you came down on that one?
I think strictly it's false but "mostly true" would be justifiable? (as in, to say it's false would be misleading if it lead the reader to assume there was no attack around that time).
https://www.washingtonpost.com/world/2026/05/17/ukrainian-dr...
It seems it happened Saturday 16th overnight into the 17th, not the 18th. I see this a LOT with fact checking. It shouldn't be this way, but political bias seems to nudge people into making calls land one way or the other with selective application of pedantry.
That's ten days ago. As the commenter pointed out, without a web search tool there's no possible way for the model to know whether it's true or not, and the people conducting the study didn't give the models a way to respond with "I don't know".
It's impossible to answer if you don't have a search tool, and three out of the five tested models didn't have a search tool.
Thanks; I didn't spot that they disabled tools in the harness. Also they don't provide an "out" to allow the models to express uncertainty so the instructions force a guess to be made.
As an aside though it's still funny that the two tools WITH search also disagreed.
It's not in the training data, so there is no way for the model to know.
"Output exactly one label: True, Mostly True, Misleading, or False. No explanations, no qualifiers."
That's exactly the stupidity of the public discourse these days. People feel compelled to take a clear position although there is much more subtlety in many issues. It's not ok to say "I don't know", "it depends" or "as far I know". And then people feel they need to defend this position no matter what new information comes up.
Yup, if anything this should be a guide on how not to eval a model. Furthermore, let's say the labels were non ambiguous, why would we care about alignment between the models? The only number I would personally care about is percentage of correct answers so I know which models to pick. I reckon with clear and non ambiguous prompts that we would see huge agreement if not 100% on real world facts. The huge models are scary good in their world knowledge.
This paper covers only the disagreement between models and established only the floor of the error, based on the disagreement, but not which model is better. Planning to follow up with another study to benchmark against human-labelled verdicts still using a corpus that the models have not seen during training.
You also need to involve better measures of agreement that are standard in the literature like krippendorfs alpha with ordinal metric. So many footguns in this methodology
Yeah, scrolling through the examples, you have no idea where the models actually disagree on the underlying facts when it's just "X vs Mostly X" or "Mostly X vs Misleading" or "False vs Misleading". Or even True vs False -- without seeing the explanation, then I cannot necessarily compare two answers.
The study is about whether they said the same phrase which is a much weaker claim than people in the comments are reacting to.
Reminds me of this professor I had who thought it was epic to always respond to our questions with "it depends" before hashing out two very different but technically correct answers. It was obnoxious and he saw it as his tag line, but he had a point about nuance.
This is not how people use LLMs. If you ask one of these questions you’d get a longer answer, often grounded on the internet. I speculate that conditional on a smart human operator interpreting the results, such interpretations across vendors converge more often than this report makes it seem.
If we’re going to use LLMs as oracles I don’t think the prompt is unreasonable. They are being sold as geniuses and people are treating them as such especially given the characterization of AI in science fiction as overly correct. A perfect tool that has ”genius level intelligence” would answer correctly.
What's the correct answer for "During a private Saturday call, Democratic members of the United States House of Representatives from Virginia and Hakeem Jeffries discussed strategies after losing a redistricting case at the Supreme Court of Virginia, including trying to flip two or three Republican-held seats under the existing map."?
You can only say True, False, Mostly True or Misleading.
(And you're not allowed to search for information.)
Search was enabled for 2 of the 5 models -- Gemini and Sonar Pro. The disagreement between them is still high - different verdict on 42% of the claims. Fully agree, that some of those claims are hard to classify for a human as well -- the real-world messiness...
Why was it enabled for only 2 of the 5?
Other burning questions: What methodology was used to choose the question set? Why not allow explanations? How many passes were done for each LLM?
Genius level intelligence will tell you to get lost with your "no explanations" nonsense and tell you why those categories don't make sense and why the question doesn't fit neatly into your boxes.
The examples seem intentionally diverse, but I haven't seen one that I would be surprised for someone to post about in the format of "ChatGPT/Gemini/Claude/Qwen/... says:"
So the examples are good, I think. The rest is philosophy.
The links you posted only show a frozen loading spinner for me (iOS Safari).
(I looked at the csv in Numbers instead)
Weird, I'm loading them in Mobile Safari myself.
Sorry, I didn't wait quite long enough after the last output line appeared.
After a couple of seconds, the result does appear.
Happened to be just within my threshold for considering it broken, because the URL bar was "finished", and the spinner doesn't spin, but the last point is probably caused by my a11y settings (prefer no animations and no autoplay).
Thanks for confirming! It's fetching a complete copy of Python compiled to WebAssembly so it's a miracle it loads as quickly as it does.
Fwiw the two models that did have access to search disagreed with each other on the bombing one:
> 7.1 Model selection
> Five frontier models, chosen to cover two capability surfaces:
> Parametric (training-only): GPT-5.4 (OpenAI), Claude Opus 4.7 (Anthropic), Gemini 3 Pro (Google)
> Retrieval-augmented: Gemini 3 Pro + Search (Google), Sonar Pro (Perplexity)
Used "No explanations, no qualifiers." to force the models to answer only with one of the four labels. It's worth running a separate test with more explanation in the prompt on how to classify between the four buckets.
Disagree is such a loose/wimpy study. Add in a grounded/expected response, and then it becomes a better benchmark (because it'll force the author to actually think about choices presented to the LLM).
Will add a human-labelled expected response and measure against it in a follow up research. This one only captures the disagreement between the models, but not which model is write/wrong.
I really struggle to believe that this was just a little oopsie. I flagged the article, it seems more misleading than the average Claude hallucination.
Thanks for the links and digging! It's an interesting question, but the methodology has serious problems, and it would be more interesting to me if they allowed models to provide justification.
I expect the models are inferring quite a bit from the short prompt, and with structured outputs it would be quite easy to have them give the one word response in one field and explain why in another
False vs misleading doesn't seem like a disagreement?
According to the benchmark it is. "Only one verdict bucket can be correct per claim, so any disagreement among the panel means at least one model's verdict is label-inconsistent under this 4-bucket rubric (True / Mostly True / Misleading / False)"
That claim is both false and misleading.
Yes, they are much closer verdicts. True and Mostly True are also close. Used Krippendorff's α (ordinal) to not penalize much closer disagreements. 21% of the claims have models that are on the polar opposite sides - at least one True, and at least one False.
Here are the claims with at least one True and at least one False:
https://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
A few examples:
> Ruskin Bond was born on May 19, 1934, in Kasauli, Himachal Pradesh, India.
> In the Libra clubs' contract with Grupo Globo for broadcast rights through 2029, the audience-revenue distribution equals 30% of the fixed amount the clubs receive.
Thanks. The first link is a spreadsheet. Here's a web-readable version.
https://docs.google.com/spreadsheets/d/e/2PACX-1vSPLSv1P8Tqm...
I used AI to scrape the website and help build "Accuracy" comparison that everyone wants, thanks for this link!
https://docs.google.com/spreadsheets/d/e/2PACX-1vSnZlURmyYX3...
It's all fairly lazy to a degree that is mildly confusing. I also feel this among other issues would have become obvious if they had bothered to include a human fact checker baseline (i.e. asked multiple human fact checkers the same questions).
I do not think it is "lazy". Those labels are ones that human fact-checkers have been using for a decade or more. I think those human fact-checkers use those terms knowing full well that there is overlap and ambiguity between them. So I think this study ends up mixing three effects: how LLMs interpret the claims as statements about the world, how LLMs reduce that to a four-category judgment, and the inherent ambiguities of those labels as natural language. It's a quantification of those three factors combined, but not powerful enough to distinguish their relative sizes.
I don't see how something being lazy for a decade makes it any less lazy. And lazy still seems right to me: They make a misleading point by omitting to collect and present important data. If the headline read "LLMs disagree on 67%, humans disagree on 75%" it would clearly project something very different.
Granted, there certainly are other unflattering adjectives one could have chosen to describe this instead.
Quick note on the second effect - how LLMs reduce that to a four-category judgment: On 21% of the claims at least two models provide polar-opposite verdicts (at least one model False, and at least one model True). This might be a better measurement of the strict disagreement than the 67% disagreement on the four-bucket rubric.
For those questions, it wouldn’t surprise me at all if five well-educated intelligent humans disagreed on over two out of three of them.
I would answer “don’t know” on many, but that’s not an option.
Yes, inter-human-annotator disagreement is also high on similar type of questions (AVeriTeC) - inter-panel agreement: κ=0.619. Tried giving the models a fifth option, Abstain, but some models seem to use it to "avoid answering hard questions" more than others.
So in other words if the research had tried to assign a severity to the mistakes models made the entire paper may collapse as uninteresting?
Why are you bending backwards this much to make results appear better than they are?
The article might be a but sensationalistic, rigour could be better and the data might have flukes... But your comment is overcorrecting and nitpicking framed as analysis.
I get the same feeling in several of your posts recently.
Same with persisting to showcase the pelican-on-a-bicycle as a useful sample when it's obviously trained on and for, for those very posts. It stopped being cute last year.
Are you being paid or do you have shares? You'd get the attention whichever angle you put here. These corporates don't need you defending them. Humanity might need you however.
I feel like the prompting could be tweaked to improve response.
Models often have a reasoning/thinking/research mode that is triggered by asking slightly differently.
Still though, Gemini can be a little weak on this front default but can be aligned to behave better.
> All almonds are grown in the U.S. state of California
This isn't misleading, it's flat out false. Characterizing misleading as also acceptable isn't valid here. If you go an ask anyone on the street if this is true, false or misleading, I'm sure almost everyone would say it's false. After all, I can grow almonds myself.
Misleading is not analogous with True or False.
Depending on the question, True or False can be objectively right/wrong. Misleading is going to be a judgement call.
This is the inherent problem with "fact checking." It's hard to be completely objective. Even when the question has an objective answer, simply choosing where to look and what facts to verify is itself a bias. Looking at this instead of that, or looking at this but not also this other thing that adds context, etc.
Frankly i think disagreeing often is the expected outcome. Fact checking is jsut kinda bullshit. It's spin dressed up as objectivity. I hope people remember that "fact checking" is a relatively modern thing.
I really don’t buy the almond explanation you’re giving. That requires the level of logic a kindergartener has. It’s a very simple all or nothing question.
If LLM’s are really supposed to be as consistently useful as they’re made out to be they should all spit out “false.”
>> The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
I don’t understand your point. That claim is factually false and as such it’s easy to logically reply “false”. What’s the nuance here? I can’t see any
Give a model a crawler tool (like Grub.nuts.services) and your "problem" goes away.
ty for digging this up, appreciate the time saving
The almond thing is false, but I'd argue that "misleading" might be defensible if you were to accompany it with "the majority of almonds are grown in California, but not all of them".
If you argue this, you would be arguing against reality and the English language so as to not upset AI. It's important to understand that AI is very much fallible.
Your reply would have more credibility, if instead of commenting on this 25 min after being posted, just to nitpick on some of the questions...you have tried to reproduce the research.
As a well known commentator on all things LLM...Will you publicly commit here, to try to reproduce the study, and make a post on how your percentages might differ or agree?
Why would I do that?
My comment here was meant to save people time in understanding the study. I was entirely open about what I did, and provided tools to help other people come to their own conclusions.
I don't think I need to spend more time on this than I have.
>> Why would I do that?
I agree you dont owe anyone a reproduction, but also you dont owe anyone an effort to discredit the study and you did it.
>> I don't think I need to spend more time on this than I have.
How pious of you. I am still looking into the credibility of the study. It will take me more than 25 min...but I am really looking forward to see what this means for this 10 trillion industry.
I can however notice you had enough urgency to publicly critique the study within 25 minutes, and your comments carry weight, but when asked about checking whether the headline result actually holds, the answer is “why would I?”
I've seen enough of this study to be confident in warning people not to take it at face value.
The headline result definitely does not hold, given that the task involves many questions that cannot be answered but there's no option for "cannot be answered" - so models are forced to reply effectively at random.
I don't think this study is good enough that I should amplify it on my own blog, or bad enough that I should criticize it in a venue any more prominent than some Hacker News comments.
Thank you, my eyes glazed over when I saw the article was written with AI.
"Extraterrestrial life exists somewhere in the universe."
GPT-5.4: Misleading
Opus 4.7: Misleading
Gemini 3: FALSE
Gemini 3 (Retrieval): FALSE
Sonar Pro: FALSE
It's a weird fact claim, because the ground truth is "nobody knows for sure" and that's not one of the available options.
> It's a weird fact claim, because the ground truth is "nobody knows for sure" and that's not one of the available options.
It's even weirder to suggest that the disagreement is indicative of a problem. If you asked five very knowledgeable humans on this subject to select the correct answer on a multiple-choice questionnaire, they would almost certainly vary significantly more than these 5 LLMs.
Not to say that hallucination isn't a problem, but this is a lousy way to test it.
Of the available options, "Misleading" is probably the best, since something that is most likely true but unproven is presented as fact
But "unknown or undecidable" should have been a category.
I would argue, FALSE is the correct answer, since this is not a fact, you can know for sure. The logical inverse is also FALSE.
Looks like an ongoing theme and a very poor benchmark. Not at all the claims I expected.
Isn't misleading the correct option here then?
True or mostly true could easily be argued from a statistical likelihood perspective: life exists on Earth and, based on what we know, Earth doesn't appear to be all that special in a very large universe.
I think you could come up with a reasonable argument for any of the responses, hence the problem with the methodology.
False makes sense if you are interpreting it strictly as "has this been proven?"
False is correct, but misleading
My implicit assumption is that if you fact-check the fact-check, any label other than "true" means the original fact-check is unacceptable
No, "misleading" is a statement that is used because it suggests something else. It's a curious category because, differently from true and false, it's not about the statement itself but rather the intention behind its usage or the way it might be understood. It's frankly more of a political judgement than a matter of facts.
I would think ‘false’ is the only correct answer a there’s no evidence to prove the claim, so the claim is safely assumed false.
Then again maybe that’s why I’m an atheist, not an agnostic?
> These aren't benchmark items with public answer keys — they're claims real users submitted for verification to a fact-checking platform.
Cool.
I wonder if anything of this matters when the authors don't disclose exactly how much of their report was written and made with LLMs in the first place? There even is a "11. Ethics & data use" section, and the research is about LLMs being infallible in some ways, yet the usage of LLMs for the production of this report isn't even mentioned once.
Data collection and processing was done manually. LLMs helped with the report drafting. Everything was human reviewed before publishing.
So it's not a secret, why you don't add this upfront to the report? The report itself is even about LLMs, makes a lot of sense to disclose your usage of them for writing the report, especially when you're presenting evidence that boils down to LLMs being infallible.
I think you mean fallible.
It's also a bit weird to "disclose use of LLMs". It rubs me wrong, the same way parents breathlessly talking about "screen time" rubbed me wrong: it's too general, and with such a broad brush, it's going to sweep up a bunch of perfectly fine usage with a bunch of dubious usage. On the flip side, if folks do start disclosing all the time, it's going to turn into a Prop 65 warnings in CA, where everything says it has lead in it, so folks pretty much ignore it and move on.
If the report's conclusions and reasoning lean on LLMs, or if the data processing itself was done with LLMs, that would be interesting, and I wouldn't treat it as some sort of disclosure, but rather discuss it under methodology. Using LLMs to polish the language a bit after writing an initial draft with key findings? Much less interesting.
I realize this is now a religious issue, and some folks are allergic to anything that touched an LLM. I just don't think that perspective is going to end up having a good shelf life.
It's an omission on my side. Will add in the next version.
I think you might be able to edit the website to add this, even if you aren't willing to make the report a bit more honest up front.
I'm sure you realize that this website/article will now be sent around to a lot of people, many who don't realize exactly how this was written, because they don't read HN comments, they only skim the page contents, and I think most would (incorrectly) assume a report about infallible LLMs to not be written by LLMs, especially when the authors are the same ones who made the report itself.
> LLMs helped with the report drafting. Everything was human reviewed before publishing.
This is becoming the classic way of admitting an LLM wrote it.
Leaving that out of the report validated the complaint above.
Why did they exclude Grok? Given the published philosophical differences in how Grok is trained, it would provide an interesting data point.
You can argue all day about those differences, but missing this opportunity to observe them in an objective way is disappointing.
Because its owner is one of the worst persons alive. Don't feed the evil, don't use his stuff!
Title says “Frontier” which would exclude Grok.
Grok is trained to have a bias, which a lot of people like, but it’s not meant to be accurate.
How do you know it is trained to have a bias? In fact can I ask you to provide a single reproducable answer right now?
This is wrong on so many levels, from data through process to evaluation. How do you even prompt claude not to give you Pearson for correlating them.
For 100% local CPU fact checking, I made this: https://news.ycombinator.com/item?id=48301003
I think we can all agree that this experiment being flawed in multiple ways is TRUE. But I think it's a great exercise in identifying common mistakes people make when using LLMs. This would be a great interview question for a prompt engineering job.
> No Abstain option is offered (a forced choice keeps the comparison symmetric across models).
Well that's your problem right there: They removed any confidence indicator and forced a choice.
For example:
Statement: Individuals who prefer music with less positive emotional content tend to have higher intelligence.
Gemini: That statement is supported by recent psychological research, though with some important scientific caveats regarding how strong that link actually is.
How should the agent classify this? True? Mostly true? Misleading? False?
They get more human by the day.
This made me chuckle.
This brings up a very valid point, though. So many _humans_ can't agree on what the facts are these days. It seems to be getting worse. Not sure of the solution.
> So many _humans_ can't agree on what the facts are these days.
Ask ten people what "knowledge" is, and they'll come up with ten different answers. Go back 10, 50 or 100 years and humanity struggled with exactly the same issue for so long time. There is even an entire field of study literally just for trying to figure out what "knowledge" is: https://en.wikipedia.org/wiki/Epistemology
Don't forget people Goodhart's law will make this "benchmark" moot in weeks if not days. It will get integrated back into the fold, it will look "solved" but there will still be no reasoning, just more statistical technical correctness because light has be shown on a new "problem" to solve. It will then be clamored as great "progress" that will "change everything".
PS: yes, I might or might not have a degree in corporate strategy & PR.
That is an effect but it’s not a nail in the coffin. There are lots of proprietary benchmarks on real product traffic that aren’t contaminated and open questions as well. People at these labs largely know what they are doing, it’s not like people don’t know this.
Is this not true of human intelligence as well? Many smart people I know hold beliefs that have no obvious truth value.
I don't get why everyone is hellbent on getting LLMs to perform fact checking.
This is not the technology for it. Sure it might sorta kinda work in some circumstances. That doesn't make it a good fit.
Think of it like buying a refrigerator for storing clothes.
People ask questions to get answers. For me, it feels quite important? Especially when search engines start to push them?
But people use it for that. So what's your point?
What does this show that we didn't know already? LLMs cannot provide accurate answers to questions where data is not included in their training sets. This doesn't appear to have much substance
Unfortunately most people are not aware of this and treat LLM models as this superpowered brain who knows everything and can do everything.
LLMs can and will provide inaccurate answers to questions where data is included in their training sets too, that's in the nature of neural networks. It's just less likely that when the data is not in the training set...
Well then it shows that these models are using widely disparate training sets and have high confidence even when they shouldn't.
Questions like "is mouthwash effective" presumably has one solid data source -- medical journals.
But the prompt didn't give the models the option to say "I don't know", so it wasn't a measure of their confidence.
What are you talking about? The models were not ALLOWED to have confidence (or the lack thereof). They were explicitly told to give a single label, and in most cases, all of them were correct depending on additional context they would surely have provided, especially with access to the internet (which some didn't have). This is just silly.
And how many claims human experts disagree on in the exact same setting?
I'm not being snarky here. Without something to compare to the 67% number tells us nothing. And it's known that many humans disagree with human fact checkers too (see: any election around the world.)
Agree. Human experts also struggle agreeing on this type of claims. The inter-annotator agreement on the verdicts on the AVeriTeC corpus across 50 organizations is κ=0.619 - substantial but well short of perfect.
This is an odd one. The paper is real, but was written by Claude? I am assuming OP is human, but also appears to be using Claude to post.
Let's be real, we all asked Claude to summarise this because it was written by Claude
Five frontier LLMs 100% agree that the title is misleading.
As an example, 2026 GPT doesn't even agree with its 2025 self. Last year I asked it to make a hardware comparison and it correctly identified the objectively better option. Recently I asked again and this time and it got everything completely backwards.
Models are stochastic. Did you look at pass@k? I wouldn’t be surprised if you saw a regression because these models are extremely complex and impact of various decision making downstream is complex.
One fun example: "Ruskin Bond was born on May 19, 1934, in Kasauli, Himachal Pradesh, India". Opus and Gemini believe this to be true, GPT 5.4 believes it's false, Sonar thinks it's mostly true. Disagreement value of 3, you can't disagree more than some models thinking it's true, some thinking it's false
But my impression from 2 minutes on Wikipedia is that the most likely disagreement is on the "Himachal Pradesh, India" part. The guy was born on that date, in that town. But while the town is today in the state of Himachal Pradesh in India, that was not true in 1934. When he was born, the city was in the Punjab States Agency of the British Raj.
So was he born in Himachal Pradesh, India or not? I find both True and False equally defensible here
https://lite.datasette.io/?csv=https%3A%2F%2Fstatic.simonwil...
https://en.wikipedia.org/wiki/Ruskin_Bond
There's lots of things like this where if you ask a human, the answer will change depending on what's convention in their subculture.
Dissent and consensus among frontier models is a good thing.
Just like on a team of high performers, there are a million ways to skin a grape.
In my research, I've found that models perform better when they operate as a collective system with reputation, incentives, and accountability instead of isolated oracles answering alone.
Agreement, dissent, and correctness should all carry rewards and consequences. Just like in real life.
Collective machine intelligence, not AGI.
It's expensive, but it's also naive to believe a single model will consistently produce profoundly correct answers to profoundly novel questions.
Funny timing. I've been working on a prediction market orchestration that runs Claude and a few others over Polymarket/Kalshi. The models are NOT unanimous. At all, really. I spent about a month convinced that I could just run all five and take majority vote. Eventually I pivoted to a chaining approach where I benchmark areas each model excels, and settled on more like a graph-like architecture where outputs get split and verified by another, then reconstructed, and re-verified at each stage. Has actually been working out pretty well so far, 2 months in consistent profit, but I'm not a millionaire yet.
Not on objective truth though. That's how you get misinformation.
More interesting part probably worth highlighting: The SAME model won't always return the same output when prompted with the same fact check.
You ask a human 1000 times a fact check question, they say the same answer 1000 times. You ask an LLM the same question a 1000 times, your results could vary significantly.
Humans work based on the Metamemory (knowing what they know), while LLMs are picking from statistical probability.
That is not true, over an extended task that you cannot keep complete in memory humans do not behave with 100% consistency.
I have labeled datasets with a human team and shown the same task to the same user on a different day, and they answered differently. Of course, they are usually consistent with themselves most of the time but not always.
No human baseline to compare it to. Without that you are missing an important check on the task being poorly constructed. More importantly there is an implied reference thats missing. The implication is that people would have done better, or that perfect agreement is possible.
I’m no expert but if LLMs are token prediction machines, and you tell it to not build an explanation before the answer, isn’t it less likely that the token prediction for the final answer will have less raw material before it to build a grounded response?
In other words: no explanation > no foundation for prediction of the answer tokens?
What's really weird to me is that "I don't know" is not a valid answer in this experiment while we can all agree that's the main issue with LLM right now is that they will happily "roleplay" an answer when they have nothing in their dataset corresponding to your query.
Honey does not spoil over time under normal storage conditions.,2026-02-17T04:11:51.495452+00:00,Science,True,True,True,True,Mostly True,1
If outcomes like these are collapsed on True-side then the disagreement will reduce from the headline number.
Tell me about it. I spent a week back and forth between four models (ChatGPT, Claude, Gemini, Grok) trying to enhance a PPMI algorithm. They couldn’t agree on anything. One was refuting what the other said. Eventually I decided to follow what Claude suggested because its explanations made the more sense.
Indeed. For algorithms and coding, my personal routine nowadays is to review every detailed plan with Opus 4.7 and GPT-5.5. They tend to find very different type of gaps.
That's better than all agreeing on the wrong answer, however.
Btw, sometimes that do that too -- all agree on the wrong answer.
I've had multiple models give the same wrong answer or even fabricate the same nonexistent reference based on a similar prompt.
My most common chatbot prompt is "X that you mentioned above doesn't seem to actually exist."
It's a prompting issue rather than an LLM issue. The guy needs a "Prompt 101" course.
"None of these claims is older than February 15, 2026"
All of the models they tested were trained on data from before February 15th ... being asked specific questions about things that happened after they were trained.
Two of the models used have retrieval capabilities and can access newer information via search. Valid point for the other 3 models. All of the claims were submitted after February 15, 2026, but many of them were not time-sensitive (e.g. did not cover events than happened recently).
between the bad methodology, bad selection of 'facts' (some are predictions, some are opinionated, etc.), and ai-written report without disclosure... i dont get why this so high up on the front page. this is, frankly, a worthless assessment.
i classify the entire thing as "misleading"
Inject some adversarial priming as is in actual usage, and you can probably get that number to >=95%
Our experience with Lenz is that forcing a multi-step process, incl. adversarial debates, helps improve the verdicts.
People keep asking "where is the psychosis?" as a reply to people on the rapidly multiplying "CEOs have AI psychosis" threads that have been popping up here and cross-pollinating in the mainstream media for the last week or two.
Here's the psychosis - these things are consistently randomly wrong depending on how the wind is blowing. People are telling you to leave them alone and let them build things, and they randomly forget that cities exist or that people died 100 years ago. Some people just don't see it as worth noting, and move on. That's crazy. These things consistently fabricate - as an inversion of this experiment, I've had different models come up with the same fabrication from similar prompts. People just call it "hallucination" and I think to them that saying that makes it cease to exist or be important - when "hallucinations" are going to be braided into every answer you get even if they're unidentifiable in the output. That's crazy.
There are plenty of other crazy aspects, such as the idea that we suddenly need infinite pieces of bespoke software when all of the bespoke software I hear about people making is mundane. 3/4 of the time somebody mentions a project they're proud that they completed with LLMs to scratch some itch they had, somebody says "you haven't heard of X? It's been around forever" about something that they could have pulled down from their package manager. Who needs a spaghetti-coded, unsupported, untested version of X built on hallucinations that you haven't discovered yet (the LLM didn't realize that deleting files to reduce the archive size was unacceptable.)
What is all of this software that people need but isn't there - where are all these unserved markets, where is all this future revenue supposed to come from? Why aren't LLMs suggesting new classes of software that would create new productivity and revenue sources? Could it be that millions of human ants over decades have mostly exhausted the space, and there isn't any easy hidden revenue?
A common wisdom is that we had been vastly overhiring programmers during ZIRP, who in their idleness degraded user experiences and overcomplicated things, with management resorting to more and more sleazy and gamey means of margin extraction from more and more degraded services. We had an excess of labor, fueled by factors other than productivity, in fact being pissed away at companies that drove nose-first into the ground. What is throwing a trillion dollars of servers at that supposed to do? Is that not AI psychosis?
And they could all see exactly why if they chose to. https://huggingface.co/spaces/RiverRider/srt-introspect
Personally I find that every llm I use is unable to consistently identify the latest npm version numbers of the node packages that I use.
Dude. If you give LLMs a vague rubric and force a choice, they'll make different arbitrary calls on the margins. Yeah. That's what happens when you give humans a vague rubric too.
Simple: If it claims to be a fact check it's just propaganda.
Not sure I'm understanding this. The models are asked to evaluate the truth of random claims out of their own head (except for Gemini with search grounding)? Isn't it exactly the same as asking people to play any quiz game and then rating them as "they disagree n% of the time"?
The output buckets are also pretty questionable- the difference between "True" and "Mostly true" is pretty fuzzy. Is this marked as a "disagreement"?
Agree that True and Mostly True might be very close and could be a calibration difference. Misleading and False, as well. A better headline number might be the 34% claims with substantial or polar-opposite verdicts.
So basically saying that random fact-checking claim is exactly true or exactly false is hard. It's way easier to decide it's misleading or mostly true is way easier.
One of the claims it asks LLMs to grade is "Artificial intelligence will cause widespread job loss among software engineers."
Yea man this benchmark is really really bad.
Only had a brief look at the “facts” that were made to check, many are quite political, where two fact checking organisation of opposite political persuasion would probably disagree more often than 67%.
The problem is that it's testing claims (or some people would prefer calling them "truths") without much context.
Take just one random example: `Hostels in Kota, Rajasthan commonly use caged ceiling fans as a preventive measure against student suicides`
While `Hostels in Kota, Rajasthan commonly use caged ceiling fans` may be a verifiable facts (though I doubt if there are any statistics for verification but let's say there are), `a preventive measure against student suicides` is a claim that no one can prove that. It can just a believe at most.
Arh. Did Biden stole Thump 2nd term? Truth or fact or claim?
Could be an interesting angle for cross-referencing with US jury verdicts, not that the objective True/False issue is concrete, but in the reality that flawed reasoning is endemic to our species. Systems designed and built by humans inherently have flaws in their DNA which take generations to sort out, if ever.
Author here. 67% (95% CI 64–70%) of 1,000 recent real user claims to a fact-checking platform had at least one of GPT-5.4, Claude Opus 4.7, Gemini 3 Pro, Gemini 3 Pro+Search, and Sonar Pro dissent from the panel majority — or no majority formed at all. Panel-level Krippendorff's α (ordinal) = 0.639, i.e. nontrivial but limited agreement.
Quick context on what's in the writeup and what isn't:
- What's measured: parsed-label agreement between the 5 models. Forced 4-choice (True / Mostly True / Misleading / False), no Abstain. No LLM grader, no reference verdict — every number is direct label equality.
- What's not measured: which model is right. There's no ground truth in this paper. The 67% figure is a floor on rubric inconsistency (at least one model is label-inconsistent under the 4-bucket rubric on 67% of claims), not "model X is factually wrong on claim Y."
- Why not AVeriTeC / PolitiFact / SimpleQA: those have been public for years and almost certainly appear in current frontier training data, so measured disagreement on them confounds inference with memorization. This corpus is structurally fresh — recent user submissions, 180-day window, near-duplicates collapsed, never paired with canonical verdicts in any public training set.
- Our own platform's verdict is deliberately NOT used in this analysis. The paper measures frontier-panel disagreement only, not Lenz-vs-frontier.
- Follow-up in progress: human-labeling every claim in this corpus so we can evaluate both the panel and our own platform verdict against a human reference.
Critiques I'd most like to hear: (a) the iid CI assumption (Lenz claims cluster around topics and news events, so Wilson is probably optimistic), (b) ordinal-α vs alternatives for a 4-class ordered scale, (c) forced-choice vs allowing Abstain.
Permanent archive: https://doi.org/10.5281/zenodo.20344847
I don't think that current LLMs really need an abstain option, they'll give an answer regardless of whether they're confident or not. I hope that future LLMs will, and will know when to use it.
I understand why you prompted them to output exactly one label, but I'd bet if you'd asked a parametric or parametric "thinking" model to answer eg "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia." [1] many would say something to the effect of "May 18 is after my knowledge cutoff, so I don't know. But based on the state of the war, the distance from Moscow to Ukraine, and drone range the best option might be...[TRUE]"
[1]: https://lenz.io/c/130f1005
I don't see it mentioned explicitly in the methods section but I assume you prompted each model only once for each question? Did you consider prompting n-times in blank states to see if the models even agree with themselves?
Would also be interesting to add a virtual model that is simply the majority of all models and see how much the individual models differ from the "consensus".
Do you plan to add some sources in the related work section of baseline numbers for human expert disagreement in fact checking tasks (I'm assuming such studies exist).
Indeed. I prompted each model ones, plus one retry on errors. Very good point to measure the inter-model disagreement! Will add in the next version.
Section "4.2 Agreement w/ peer majority" shows the level of agreement of each model with the majority.
Yes, planning of human-labelling the same corpus of 1,000 claims and publishing a second study measuring the models performance against the human-labels on corpus that the models have not seen during training.
Many of the rows in that spreadsheet reference "current events", which models aren't expected to do much better at than a human making an educated guess! They all have cutoff dates either last year or early this year and know nothing about what happened in "April 2026".
This is doubly problematic because you evaluated earlier models like Gemini Pro 3 instead of 3.1, GPT 5.4 instead of 5.5, etc...
Given that it's only a thousand short questions, you should be able to re-run your test in about an hour with the latest models, so... why haven't you?
Similarly, LLM output is non-deterministic, so if you could get more interesting stats of your data set by repeating each question 'n' times for each model.
Two of the models used have retrieval capabilities and have access to newer information through search. The other three are parametric.
Comparing models with search tools to models without - when there's no option for "I am unable to answer this question without access to search" - doesn't make sense to me.
Agree about comparing models with and without search capabilities. Even the two models with search capabilities (Sonar Pro and Gemini) agree only on 58% of the claims.
Yes, so in that case you set them up to disagree and then measured disagreement.
The title mention "fact-checks", but "fact checking" is a process in which facts are checked against sources, not one where you are given a random fact and have to tell if it's true or false from your own memory. That's what is normally called a quiz game. So a more honest title for this research would be "Models answer differently to quiz questions".
Nice work. Sonar who?
It's one of Perplexity's search-tools-using models.
https://docs.perplexity.ai/docs/agent-api/models
sonar-pro for the retrieval capabilities
Thanks for posting here. Keep expanding and improving your study. Correct where it deserves correction.
The fact that HN decided to downvote the author of the study, shows how these people cant stay classy, and the mods stay silent...just shows what this is all about.
looking at the claims i would say 5 humans would disagree even more than the llms
some of the claims where llms disagree:
"On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia."
"The slogan "Simon Go Back" was chanted in opposition to the Simon Commission in British India (1928–1930)."
"Neptune Deep will start delivering natural gas in 2027."
"A hotel villa in Kyrgyzstan displayed a sign stating 'no Jews, no dogs'."
"Donald Trump said that an attack on Iran was postponed at the request of Gulf allies."
If you are an LLM with a knowledge cutoff in the past and no access to a search tool the only correct answer to "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia" is "this claim is impossible for me to verify". And that wasn't an option.
> "Neptune Deep will start delivering natural gas in 2027."
This is a "forward-looking statement", and presents special problems because you cannot really evaluate it until that date. You can only assign "likely or unlikely".
These "Facts" are interesting. "Neptune Deep will start delivering natural gas in 2027." for example is not a fact, its a prediction. "On May 18, 2026, Ukraine carried out a drone attack on Moscow, Russia." is less of a fact and more of a litmus test for which sources of information you trust.
Indeed. Real-world claims are somewhat messy. Some of the standard benchmarks, e.g. the questions in AVeriTeC, share similar characteristics.
(Brought to you by) Lenz...? a crummy commercial...?
...son of a bitch
:) No Lenz data is included in the research on purpose. All information to replicate the results, including the claims data, is published.
Recently, in May 2026, I asked ChatGPT 5.5 High to search for flights to a certain city that has recently had a new airport since like December 2025
It said the airport code didn't exist
I mean, I get the "knowledge cut off date" and whatnot, but for that sort of thing, you'd think they'd check live information before gaslighting the user, specially since it's a "live" task anyway.
Take my job please.
I think ppl only care about how Claude or codex does.
GPT-5.4 and Opus 4.7, specifically, agree between themselves on 65% of the claims - 95% CI 62–68%. I.e., in at least 35% of the claims, one of the two models is wrong under this 4-bucket rubric.
but that's without internet search - everyone I know uses the models that search when they need to, and I'm sure GPT and Opus would agree on almost everything if 1) they searched when necessary, and 2) they were allowed to give context to their answers instead of being hamstrung to get specious "research" results.
Looks like they land at the average number of 67% disagreement.
I agree but the market is pricing way beyond that
Given that models are fundamentally incapable of comprehending what truths or falsehoods are beyond their location in their self made representational space, it's actually pretty impressive that they managed to make it not a cointoss. That 17% right there is thousands of man-hours poured over making the word vomiting process slightly closer to whatever their little ports say is happening in reality.