Meta ordered to pay New Mexico $25 billion over Cambridge Analytica
AspireAI · CNBC
On October 21, 2026, a New Mexico district court judge ordered Meta to pay $25 billion to the state, plus interest that accrues during the appeal, in the Cambridge Analytica data-privacy case. A jury in Santa Fe found in September that Meta was liable for misleading state residents about how it safeguards and collects user data, and returned a verdict of more than 43 million violations of New Mexico's Unfair Practices Act. With a maximum civil penalty of up to $5,000 per violation, the calculated award reached $219.5 billion; the judge reduced it to $25 billion. The state had sought $35 billion to $40 billion, while Meta's lawyers argued for no more than $3.4 billion. The underlying facts are a decade old: Cambridge Analytica improperly obtained information tied to roughly 87 million Facebook users through a third-party personality quiz for targeted political advertising around the 2016 election, and Facebook disclosed the issue in 2018. Jurors assessed 34 separate times whether Facebook misled users about privacy and content moderation. Judge Francis Mathew wrote that Facebook's misrepresentations about users' control of their data 'were not isolated incidents, but rather a systematic pattern of deception at the core of Facebook's business model.' Meta's chief legal officer, C.J. Mahoney, called the ruling 'a tax on speech' and said the company will appeal, noting interest would accrue during the appeal and the violation count could be reduced. New Mexico is the only state that took the case to trial; 48 other states settled related claims in August for roughly $18 billion, a deal that also resolved their outstanding Cambridge Analytica claims. Florida also declined that settlement. What is genuinely new here is not the scandal but the price attached to it, arriving as courts elsewhere weigh AI-related liability.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.