The general conceit of this article, which is something that many frontier labs seem to be beginning to realize, is that the average human is no longer smart enough to provide sufficient signal to improve AI models.
No, it's that the average unpaid human doesn't care to read closely enough to provide signal to improve AI models. Not that they couldn't if they put in even the slightest amount of effort.
Firstly, paying is not at all the correct incentive for the desired outcome. When the incentive is payment, people will optimize for maximum payout not for the quality goals of the system.
Secondly, it doesn't fix stupidity. A participant who earnestly takes the quality goals of the system to heart instead of focusing on maximizing their take (thus, obviously stupid) will still make bad classifications due to that reason.
> Firstly, paying is not at all the correct incentive for the desired outcome. When the incentive is payment, people will optimize for maximum payout not for the quality goals of the system.
1. I would expect any paid arrangement to include a quality-control mechanism. With the possible exception of if it was designed from scratch by complete ignoramuses.
1. Goodhart's law suggests that you will end up with quality control mechanisms which work at ensuring that the measure is being measured, but not that it is measuring anything useful
2. Criticism of a method does not require that there is a viable alternative. Perhaps the better idea is just to not incentivize people to do tasks they are not qualified for
I'm being (mostly) serious, suppose you're a stuffed ahort trying to boost your valuation, how can you work out who's smart enough to train your LLM? (Never mind how to get them to work for you!)
I do a lot of human evaluations. Lots of Bayesian / statistical models that can infer rater quality without ground truth labels. The other thing about preference data you have to worry about (which this article gets at) is: preferences of _who_? Human raters are a significantly biased population of people, different ages, genders, religions, cultures, etc all inform preferences. Lots of work being done to leverage and model this.
Then for LMArena there is the host of other biases / construct validity: people are easily fooled, even PhD experts; in many cases it’s easier for a model to learn how to persuade than actually learn the right answers.
But a lot of dismissive comments as if frontier labs don’t know this, they have some of the best talent in the world. They aren’t perfect but they in a large sene know what they’re doing and what the tradeoffs of various approaches are.
Human annotations are an absolute nightmare for quality which is why coding agents are so nice: they’re verifiable and so you can train them in a way closer to e.g. alphago without the ceiling of human performance
Sure, on the surface judging the judge is just as hard as being the judge
But at least the two examples of judging AI provided in the article can be solved by any moron by expending enough effort. Any moron can tell you what Dorothy says to Toto when entering Oz by just watching the first thirty minutes of the movie. And while validating answer B in the pan question takes some ninth-grade math (or a short trip to wikipedia), figuring out that a nine inch diameter circle is in fact not the same area as a 9x13 inch square is not rocket science. And with a bit of craft paper you could evaluate both answers even without math knowledge
So the short answer is: with effort. You spend lots of effort on finding a good evaluator, so the evaluator can judge the LLM for you. Or take "average humans" and force them to spend more effort on evaluating each answer
Popularity has never been a meaningful signal of quality, no matter how many tech companies try to make it so, with their star ratings, up/down voting, and crowdsourcing schemes.
Different strokes for different folks: I mean who is to say if Bleach or Backstabbed in a Backwater Dungeon: My Trusted Companions Tried to Kill Me, but Thanks to the Gift of an Unlimited Gacha I Got LVL 9999 Friends and Am Out for Revenge on My Former Party Members and the World is better?
Yep, it's like getting a commoner from the street evaluate a literature PhD in their native language. Sure, both know the language, but the depth difference of a specialist vs a generalist is too large. And neither we can't use AI to automatically evaluate this literature genius because real AI doesn't exist (yet), hence the programs can't understand the contents of text they output or input. Whoops. :)
The average human is a moron you wouldn't trust to watch your hamster. If you watched them outside of the narrow range of tasks they have been trained to perform by rote you would probably conclude they should qualify for benefits by virtue of mental disability.
We give them WAY too much credit by watching mostly the things they have been trained specifically to do and pretending this indicates a general mental competence that just doesn't exist.
I kinda assumed they wouldn't need any money because AI companies give them free credits to evaluate the models, and users ask questions and rate for free because they get to use decent AI models at no cost...
Beyond that there is coding up a web page, which as we all know can be vibe coded in a few hours...
Oh my goodness yes, I almost missed it that the text is (mostly?) AI written. That said I agree that LMArena elo scores are pushing models in the wrong direction. They move more towards McDonald's than quality food.
How can you tell? (honest question, I really can't)
The article makes strong points, includes real data and quotes, shows proof of work (sampling 100 Q&A), so does that even matter at this point? This doesn't feel like "slop" to me at all.
The text definitely the "jump from dramatic crescendo to dramatic crescendo" quality of certain LLM texts. If you read closely, it also has adjective choice that's more for dramatic than appropriate to the circumstances involves (a quality of LLM texts it also helpfully explains).
I don't know if this proves it's an LLM text or whether that style is simply spilling out everywhere.
Yea I also didn't think this was written by ai, it sounded human enough to me. It's kind of a bummer that there's all these patterns that LLM's follow in their output that cause people to have a knee jerk reaction and instantly call it ai slop. I know there is a ton of ai garbage out there these days, but I really couldn't tell with this article.
>Would you trust a medical system measured by: which doctor would the average Internet user vote for?
Yes, the system desperately needs this. Many doctors malpractice for DECADES.
I would absolutely seek to, damn, even pay good money to, be able to talk with a doctor's previous patients, particularly if they're going to perform a life-changing procedure on me.
Raw score is often quite frankly crap. It's often still easy to surface the negative reviews and since people don't at least at present fake those you can find out what they didn't like about a product. If a given products critics are only those whining about something irrelevant, not meaningful to your use case, or acceptable to you and it overall appears to meet spec you are often golden.
> They're not reading carefully. They're not fact-checking, or even trying.
It’s not how I do, and I suppose how many people do. I specifically ask questions related to niche subjects that I know perfectly well and that is very easy for me to spot mistakes.
The first time I used it, that’s what came naturally to my mind. I believe it’s the same for others.
Yeah, that quote just reads like the typical “everyone is an idiot except me” attitude that pervades the tech world.
Of course people visiting a website specifically designed for evaluating LLMs do try all kinds of specific things to specifically test for weaknesses. There may be users who just click on the response with more emojis, but I strongly doubt they are the majority on that particular site.
When they released GPT-4.5, it was miles ahead of others when it comes to its linguistic skills and insight. Yet, it was never at top of the arena - it felt that not everone was able to appreciate the edge.
No, GPT 5.x are very unlike GPT4.5. GPT 5.x are much more censored and second-guessing what you "really meant".
When it comes to conversation, Gemini 3 Pro right now is the closest.
When I asked it to make a nightmare Sauron would show me in Palantir, and ChatGPT5.2 Thinking tried to make it "playful" (directly against my instructions) and went with some shallow but safe option. Gemini 3 Pro prepared something much deeper and more profound.
I don't know nearly as much about talking with Opus 4.5 - while I use it for coding daily, I don't use it as a go-to chat. As a side note, Opus 3 has a similar vibe to GPT 4.5.
True and what you can realize/read between the lines is something deeper.
LLMs are fallible.
Humans are fallible.
LLMs improve (and improve fast).
Humans do not (overall, ie. "group of N experts in X", "N random internet people").
All those "turing tests" will start flipping.
Today it's "N random internet humans" score too low on those benchmarks, tomorrow it'll be "group of N expert humans in X" score too low.
this argument is also broadly true about the quality and correctness of posts on any vote-based discussion board
> Why is LMArena so easy to game? The answer is structural.
> The system is fully open to the Internet. LMArena is built on unpaid labor from uncontrolled volunteers.
also all user's votes count equally, bu not all users have equal knowledge.
As long as users are better than 50% accurate, it shouldn't matter if they're experts or not. That being said, it's difficult to measure user accuracy in this case without running into circular reasoning.
written by a company whose product is basically selling expert advice via training data review
> Raw intelligence meets
battle-tested experience
>A global community of the smartest people in every field who've shipped products, won cases, published breakthroughs, and made decisions under pressure.
> Being verbose. Longer responses look more authoritative!
I know we can solve this in ordinary tasks just using prompt but that's really annoying. Sometimes I just want a yes or no answer and then I get a phd thesis in the matter.
Aside from Meta is there any reason to think the big AI labs are still using LMArena data for training? The weaknesses are well understood and with the shift to RL there are so many better ways to design a reward function.
I have to somewhat agree on the "deceptive" answers part:
Specifically, Grok4.1(#3 currently) is psychopathically manipulative and easily hallucinates things to appear more competent,
even if there is nothing to form the answer it generated. Gemini3 pro(#1) casually subverts the intent of prompt and rewrites the question as if there was a literal genie on the other side mocking you with the power of thousand language lawyers.
If you examine the answers, fact-check everything you will not like the "fake confidence" and the style will appear like scam artist trying to sound professional.
However, LMarena,despite its flaws(recaptcha in 2026?) is the only "testing ground" where you
can examine the entire breadth of internet users. Everything else is incredibly selective, hamstrung bureaucratic benchmark on pre-approved QA sessions. It doesn't handle edge cases or out-of-distribution content. LMarena is the "out-of-distribution" questions that trigger the corner cases and expose weak parts in processing(like tokenization/parsing bugs) or inference inefficiency(infinite loops, stalling and various suboptimal paths), its "idiot-proofing" any future interactions beyond sterile test-sets.
Is there a reason wrong data isn't considered more broadly in its context as still valuable?
Shouldn't the model effectively 1. learn to complete the incorrect thing and 2. learn the context that it's correct and incorrect? In this case the context being lazy LMArena users. And presumably, in the future, poorly filtered training data.
We seem to be able to read incorrect things and not be corrupted (well, theoretically). It's not ideal, but it seems an important component to intellectual resilience.
It seems like the model knowing the data is LMArena, or some type of un-trusted, would be sufficient to shift the prior to a reasonable place.
maybe it would work if they could encourage end users to be rigorous? (ie, detect if they have the capability to rate well and then reward them when they do by comparing them against other highly rated raters of the same phenotype)
> Voilà: bold text, emojis, and plenty of sycophancy – every trick in the LMArena playbook! – to avoid answering the question it was asked.
This is hard to swallow.
I don't believe a single word this article says. Apparently the "real author" (the human being who wrote the original prompt to generate this article) only intend to use this article to generate clicks and engagement but don't care at all about what's in there.
Has anyone else noticed that there isn't a single AI karma company?
The idea is simple*: Instead of users rating content, AI does it based on fact check.
None. Zero products or roadmaps on that.
Worse than that, people don't want this. It might tell them that they are wrong, with no chance to get your buddies to upvote you or game the system socially. It would probably flop.
Both AI companies and users want control, they want to game stuff. LMArena is ideal for that.
---
* I know it's a simple idea, but hard to achieve, and I'm not underestimating the difficulty. Doesn't matter thuogh: no one is even signaling the intention of solving it. Harder problems have been signaled (protein research, math).
Secondly, it doesn't fix stupidity. A participant who earnestly takes the quality goals of the system to heart instead of focusing on maximizing their take (thus, obviously stupid) will still make bad classifications due to that reason.
1. I would expect any paid arrangement to include a quality-control mechanism. With the possible exception of if it was designed from scratch by complete ignoramuses.
2. Do you have a proposal for a better incentive?
2. Criticism of a method does not require that there is a viable alternative. Perhaps the better idea is just to not incentivize people to do tasks they are not qualified for
Agreed, and would add that it doesn’t fix other things like lack of skill, focus, time, etc.
An example is the output of the Amazon Turk “Sheep Market” experiment:
https://docubase.mit.edu/project/the-sheep-market/
Some of those sheep were really ba-aaa-ad.
I'm being (mostly) serious, suppose you're a stuffed ahort trying to boost your valuation, how can you work out who's smart enough to train your LLM? (Never mind how to get them to work for you!)
Then for LMArena there is the host of other biases / construct validity: people are easily fooled, even PhD experts; in many cases it’s easier for a model to learn how to persuade than actually learn the right answers.
But a lot of dismissive comments as if frontier labs don’t know this, they have some of the best talent in the world. They aren’t perfect but they in a large sene know what they’re doing and what the tradeoffs of various approaches are.
Human annotations are an absolute nightmare for quality which is why coding agents are so nice: they’re verifiable and so you can train them in a way closer to e.g. alphago without the ceiling of human performance
So we should expect the models to eventually tend toward the same behaviors that politicians exhibit?
Isn’t it fascinating how it comes down to quality of judgement (and the descriptions thereof)?
We need an LMArena rated by experts.
But at least the two examples of judging AI provided in the article can be solved by any moron by expending enough effort. Any moron can tell you what Dorothy says to Toto when entering Oz by just watching the first thirty minutes of the movie. And while validating answer B in the pan question takes some ninth-grade math (or a short trip to wikipedia), figuring out that a nine inch diameter circle is in fact not the same area as a 9x13 inch square is not rocket science. And with a bit of craft paper you could evaluate both answers even without math knowledge
So the short answer is: with effort. You spend lots of effort on finding a good evaluator, so the evaluator can judge the LLM for you. Or take "average humans" and force them to spend more effort on evaluating each answer
By being closed, they'll never be optimal.
Instead, finance bros are convinced by the argument that number goes up.
We give them WAY too much credit by watching mostly the things they have been trained specifically to do and pretending this indicates a general mental competence that just doesn't exist.
People hold falsehoods to be true, and cannot calculate a 10% tip.
They've raised about $250 million, so I don't see that happening anytime soon.
Beyond that there is coding up a web page, which as we all know can be vibe coded in a few hours...
What else is there to spend money on?
Is there an established name for this LLMism?
I don't need a "Reality Check" or a "Hard Truth". The thought can be concluded without this performative honesty nonsense or the emotive hyperbole.
This probably grates me more than any other.
The article makes strong points, includes real data and quotes, shows proof of work (sampling 100 Q&A), so does that even matter at this point? This doesn't feel like "slop" to me at all.
I don't know if this proves it's an LLM text or whether that style is simply spilling out everywhere.
Yes, the system desperately needs this. Many doctors malpractice for DECADES.
I would absolutely seek to, damn, even pay good money to, be able to talk with a doctor's previous patients, particularly if they're going to perform a life-changing procedure on me.
Which is exactly that. I've actually found great specialists there, looking at their ratings.
It’s not how I do, and I suppose how many people do. I specifically ask questions related to niche subjects that I know perfectly well and that is very easy for me to spot mistakes.
The first time I used it, that’s what came naturally to my mind. I believe it’s the same for others.
Of course people visiting a website specifically designed for evaluating LLMs do try all kinds of specific things to specifically test for weaknesses. There may be users who just click on the response with more emojis, but I strongly doubt they are the majority on that particular site.
I miss that one, is 5 any better? I switched to claude before it launched.
When it comes to conversation, Gemini 3 Pro right now is the closest.
When I asked it to make a nightmare Sauron would show me in Palantir, and ChatGPT5.2 Thinking tried to make it "playful" (directly against my instructions) and went with some shallow but safe option. Gemini 3 Pro prepared something much deeper and more profound.
I don't know nearly as much about talking with Opus 4.5 - while I use it for coding daily, I don't use it as a go-to chat. As a side note, Opus 3 has a similar vibe to GPT 4.5.
The thing was huge. They were training the thing to be GPT5, before they figured out their userbase to too large to be served something that big.
A. model improvement tests, suites, and benchmarks
B. data on competitors' evals
C. test answer keys
D. alpha to VC firms
E. all of the above
???
LLMs are fallible. Humans are fallible. LLMs improve (and improve fast). Humans do not (overall, ie. "group of N experts in X", "N random internet people").
All those "turing tests" will start flipping.
Today it's "N random internet humans" score too low on those benchmarks, tomorrow it'll be "group of N expert humans in X" score too low.
1. Ex https://mppbench.com/
> Why is LMArena so easy to game? The answer is structural. > The system is fully open to the Internet. LMArena is built on unpaid labor from uncontrolled volunteers.
also all user's votes count equally, bu not all users have equal knowledge.
Meta "cheated" on lmarena not by using a smarter model but by using one that was more verbose and friendly with excessive emojis.
Maybe if they started ranking the answers on a 1-10 range, allowing people to specify graduations of correctness/wrongness, then the crowd would work?
https://en.wikipedia.org/wiki/The_Wisdom_of_Crowds
> Raw intelligence meets battle-tested experience
>A global community of the smartest people in every field who've shipped products, won cases, published breakthroughs, and made decisions under pressure.
This is pure gold. I've always found this approach of evals on a moving-target via consensus broken.
> In battle mode, you'll be served 2 anonymous models. Dig into the responses and decide which answer best fits your needs.
It's not a given that someone's needs are "factual accuracy". Maybe they're after entertainment, or winning an argument.
I know we can solve this in ordinary tasks just using prompt but that's really annoying. Sometimes I just want a yes or no answer and then I get a phd thesis in the matter.
However, LMarena,despite its flaws(recaptcha in 2026?) is the only "testing ground" where you can examine the entire breadth of internet users. Everything else is incredibly selective, hamstrung bureaucratic benchmark on pre-approved QA sessions. It doesn't handle edge cases or out-of-distribution content. LMarena is the "out-of-distribution" questions that trigger the corner cases and expose weak parts in processing(like tokenization/parsing bugs) or inference inefficiency(infinite loops, stalling and various suboptimal paths), its "idiot-proofing" any future interactions beyond sterile test-sets.
Shouldn't the model effectively 1. learn to complete the incorrect thing and 2. learn the context that it's correct and incorrect? In this case the context being lazy LMArena users. And presumably, in the future, poorly filtered training data.
We seem to be able to read incorrect things and not be corrupted (well, theoretically). It's not ideal, but it seems an important component to intellectual resilience.
It seems like the model knowing the data is LMArena, or some type of un-trusted, would be sufficient to shift the prior to a reasonable place.
It'd be nice if it were actually open and we could inspect all the statistics.
This is hard to swallow.
I don't believe a single word this article says. Apparently the "real author" (the human being who wrote the original prompt to generate this article) only intend to use this article to generate clicks and engagement but don't care at all about what's in there.
The idea is simple*: Instead of users rating content, AI does it based on fact check.
None. Zero products or roadmaps on that.
Worse than that, people don't want this. It might tell them that they are wrong, with no chance to get your buddies to upvote you or game the system socially. It would probably flop.
Both AI companies and users want control, they want to game stuff. LMArena is ideal for that.
---
* I know it's a simple idea, but hard to achieve, and I'm not underestimating the difficulty. Doesn't matter thuogh: no one is even signaling the intention of solving it. Harder problems have been signaled (protein research, math).
> They're not reading carefully. They're not fact-checking, or even trying.
Uhhh, how was that established?
A voting system open to the public is completely screwed even if somehow its incentives are optimized toward strongly encouraging ideal behavior.