Traditional surveillance historically has been limited by the silo problem. Data was fragmented across different systems: license plate readers, facial recognition databases, and social media dragnets were separate tools requiring manual labor to connect. Multimodal frontier models change this dynamic by collapsing these independent signals into a single, unified layer of interpretation.
Heidy Khlaaf, Chief AI Scientist at the AI Now Institute, said that the very data used to train these [frontier AI] models—often scraped from the public Web or procured via data brokers—enables “dual-use” capabilities that facilitate state monitoring.
“Disparate datasets can be consolidated into a centralized model that can then be queried to produce determinations and inferences about populations with ease and scale,” Khlaaf explained. She warned these correlations are often prejudiced and can falsely implicate individuals based on flawed statistical patterns.
In other words, if AI labs allow governments latitude to use their models in this way, frontier AI could enable whole new levels of surveillance.
[…]
Khlaaf also suggested that the scope of current restrictions is too narrow to protect the public. While the U.S. has issued executive orders preventing “countries of concern” from accessing bulk personal data, these orders do not yet extend to the AI models themselves. “Commercially available models trained on such data can enable the insights derived from personal data to be utilized for surveillance even where the data itself is restricted from sale,” Khlaaf explained.
Read the article here.
Research Areas