An industry-standard league table for ranking artificial intelligence models is being deliberately distorted by technology giants, researchers have claimed, leading to a misleading picture of which AIs are the best.
at Cohere Labs, a US non-profit, and her colleagues claim to have found that the popular benchmark is a “distorted playing field”, with policies that end up giving an advantage to large companies like Meta, Amazon and Google by allowing them to discard models that score poorly.
“If you can choose what score to post, we’re not doing science any more,” says Hooker. “Every scientist who is in this field knows that this is not responsible scientific practice. This was a very uncomfortable paper to write, in part, because it’s pointing out something about our field which is very broken. I think it’s a low for AI rigour.”
Advertisement
Chatbot Arena has become an industry standard because it allows people to compare AI models head-to-head. Users submit a text prompt and receive responses from two different anonymous AI models, then choose which performed best. These votes feed into a league table that ranks all of the AIs on the site.
Hooker and her colleagues, some of whom have submitted models to Chatbot Arena, say that the idea behind the site is sound, but that it has been allowed to drift from its . For one thing, organisations are allowed to test multiple models at once, without all of them appearing in rankings or a score being publicly announced. These submissions can then later be removed from the website, and no data on them publicly listed. Hooker says that this allows companies to test many variants, pick the one that scores the best and then publicly release only that version, to give a more successful appearance.
The team analysed data on more than two million head-to-head tests on Chatbot Arena carried out between January 2024 and April 2025. They found that Meta tested 27 private AI model variants in the lead-up to the release of its flagship Llama 4 in April, while Google had 10 prior to the launch of Gemma 3 in March. None of these models appeared in league tables, but did appear in specific head-to-head battles, creating a public record of their use. Chatbot Arena also runs specialised leaderboards for performance on visual problems, or generating code, and Meta was found to have an additional 16 variants in these areas, bringing the total number of models at one time to 43, say the researchers.
at the University of California, Berkeley, who is one of the people behind Chatbot Arena, says there is nothing wrong with companies choosing to remove their private models from the site. “It doesn’t make sense to release scores for a model that can’t be used by anyone. No one can use it or test it to verify the results for themselves.”
Hooker says there are also further problems. She claims that models submitted by large companies are being put forward in public tests more often than smaller, open-source models. The group’s paper claims that Google and OpenAI received 19.2 per cent and 20.4 per cent of all test prompts on the arena, respectively, while 41 fully open-source models together only received 8.8 per cent.
Angelopoulos says that it is true Chatbot Arena prioritises certain models over others. “We upsample the best models, and new models, so they get seen more frequently in battles and provide the most value back to the community,” he says. But he denies that the big tech firms are allowed to do things that other users of the site can’t. “Nobody is getting preferential treatment,” he says.
The claims have sparked wider concerns about the competing interests of science and business when it comes to developing AI. “Personally, I have trouble believing any results that people report about their models,” says at AI company Hugging Face. “People will inexorably find ways to game [benchmarks], or to amplify their performance with loopholes. I think part of the reason for this is that AI research has become so intertwined with commercialisation and profit that it’s no longer just about showing scientific progress.”
“I’m surprised that anyone is surprised,” says at the University of Surrey, UK. “Some companies have people dedicated to figuring out whether there are ways for systems to improve their scores against well-known benchmarks.”
To address these concerns, Hooker and her colleagues suggest that all models given a score must remain permanently on public record, that those submitting AI models should be able to test a maximum of three variants at a time and that fair sampling algorithms should be created to stop models from big companies being served to evaluators more often than smaller models.
In a statement , the developers of Chatbot Arena said they agreed with some of the suggestions, but added that the team’s paper contains “a number of factual errors and misleading statements”. The team have not publicly responded to this.
Google, Amazon and Meta didn’t respond to a request for comment.
arXiv
