Experts have condemned Chatbot Arena, a beta crowdsourced judging platform for testing AI models, for being dangerous by design. They dispute its validity in correctly benchmarking these advanced technologies. LMArena is stewarding Chatbot Arena, which was co-founded by Wei-Lin Chiang, an AI PhD student at UC Berkeley. The platform then enlists volunteers to prompt two standout models of their choice and select their favorite responses. Experts contend that the social media platform has consistently failed to show a direct connection between what users express a preference for and what they ultimately select.
Their platform is dedicated to testing cutting edge new AI models. It engages users in the experience by encouraging them to complete multiple challenges. Despite its noble intentions, experts like Alex Atallah, CEO of the model marketplace OpenRouter, suggest that Chatbot Arena is insufficient for open testing and benchmarking of AI models. As a result, they have warned that the findings produced by the tool can’t be trusted.
Asmelash Teka Hadgu, co-founder of AI firm Lesan, has criticized how AI labs are utilizing Chatbot Arena to promote exaggerated claims about their models. Better, or more appropriate, benchmarks can be misleading to stakeholders, he warns. In defense of the platform, Chiang asserts that recent incidents, such as the Maverick benchmark discrepancy, stemmed from a misinterpretation of Chatbot Arena’s policies by the labs, rather than flaws in the platform’s design.
In response to the feedback received, Chatbot Arena has updated its policies to “reinforce our commitment to fair, reproducible evaluations,” according to Chiang. The goal is to provide an independent, impartial and credible space for users to publicly rate and review different AI models.
Chiang stresses that the community surrounding Chatbot Arena is not just a group of tinkerers or model testers. Rather, it consists of people drawn there to the platform for a variety of reasons, such as concisely put as “learning and practicing new skills.” He emphasizes that they work to create an open environment for people to come engage with AI and iterate on group feedback.
Even though the platform is intended to promote more transparency, experts are not convinced. Emily Bender, a linguist and AI researcher at the University of Washington, emphasizes the need for valid benchmarks. She explains, “For a benchmark to be valid, it should actually measure an actual something … It should have construct validity, that is, there should be proof that the construct of interest is clearly defined and that the measurements genuinely pertain to the construct.” This means that crowdsourcing evaluations alone is likely not enough to create reliable benchmarks.
Matt Frederikson, CEO of Gray Swan AI, emphasizes the need for transparency between model developers and creators of benchmarks and challenges. They need to be ready to act quickly when their conclusions are challenged. He notes, “It’s important for both model developers and benchmark creators, crowdsourced or otherwise, to communicate results clearly to those who follow.”
At Chatbot Arena, we have built in some fantastic incentives to encourage participation. It offers cash prizes for completing particular tests, incentivizing users depth. Continuous and collaborative improvement Hadgu believes that benchmarks should be flexible and adaptive. This concept keeps them up-to-date for accurately measuring AI capabilities.
Despite these criticisms, major AI labs such as OpenAI, Google, and Meta are leveraging Chatbot Arena to explore the strengths and weaknesses of their latest models. Its platform’s continued capacity to pull in a more diverse user base is arguably the platform’s greatest asset.