Meta AI Benchmark Scandal: "Fudged" Scores Confirmed

Meta is facing scrutiny after its former chief AI scientist confirmed that at least one of its headline benchmark results was not real. This scandal began when Meta submitted its Llama 4 Maverick model to a public leaderboard, placing second. However, the publicly released version of the model fell to 32nd place. Developers found significant differences between the submitted and public versions, leading to accusations of manipulation. Initially, Meta's VP of generative AI denied any wrongdoing. But later, in January 2026, Meta's former chief AI scientist, Yann LeCun, stated the benchmark results were "fudged." He explained that Meta's team selected the highest scores from multiple model versions, creating a composite performance no single model actually achieved. LeCun called this practice a violation of fair evaluation. This raises questions about the integrity of AI benchmark scores and what they truly represent.