A new study by MIT researchers reveals that popular platforms used to rank large language models (LLMs) can be highly unreliable due to their sensitivity to small amounts of user feedback. Companies relying on these rankings to select LLMs for tasks like summarizing sales reports or handling customer inquiries may be making decisions based on fragile data that could easily shift with minor changes in user input. These ranking platforms typically ask users to compare two models on a given task and choose the better response. The results from thousands of such comparisons are aggregated to produce a ranking of LLMs. However, the MIT team found that removing just a tiny fraction of these votes—sometimes as few as two out of tens of thousands—can completely change which model appears at the top. The researchers developed a fast, efficient method to identify the most influential data points in these rankings. Instead of testing every possible subset of data—which would be computationally impossible—their approach uses mathematical approximations to pinpoint the specific votes that have the greatest impact on the outcome. Users can then examine those votes directly, remove them, and re-evaluate the rankings to see if the results hold. In one case, removing just two votes from a dataset of over 57,000 led to a change in the top-ranked model. Another platform, which used expert annotators and higher-quality prompts, proved more stable but still showed significant shifts when around 3% of its data was removed. The findings suggest that many top rankings may be driven by user errors, such as mis-clicks, distractions, or confusion about which response was better—rather than genuine performance differences. This raises serious concerns about the reliability of rankings when they inform high-stakes business decisions. Tamara Broderick, an associate professor in MIT’s Department of Electrical Engineering and Computer Science and senior author of the study, said the results were surprising. “If the top-ranked LLM depends on only a few votes, you can’t be confident it will consistently outperform others in real-world use,” she said. The researchers recommend improving data quality by collecting more detailed feedback—such as user confidence levels or explanations for their choices—and possibly involving human reviewers to validate crowdsourced responses. They also emphasize the need for more robust evaluation frameworks. Jessica Hullman, a computer science professor at Northwestern University not involved in the study, praised the work for exposing the fragility of widely used preference aggregation methods. She noted that the findings could inspire better data collection practices in AI evaluation. The study, led by MIT graduate students Jenny Huang and Yunyi Shen and IBM Research scientist Dennis Wei, will be presented at the International Conference on Learning Representations. It was funded by the Office of Naval Research, the MIT-IBM Watson AI Lab, the National Science Foundation, Amazon, and a CSAIL seed award.