AI, Privacy, and the Hidden Architecture of Harm from Inference
AI, Privacy, and the Hidden Architecture of Harm from Inference
https://www.techpolicy.press/ai-privacy-and-the-hidden-architecture-of-harm-from-inference/
Publish Date: 2026-06-17 16:23:00
Source Domain: www.techpolicy.press
This post is part of a series of student essays produced in collaboration with the Berkman Klein Center for Internet & Society at Harvard University. Read more in the series here.
Artificial intelligence is often framed as a story of innovation: faster systems, smarter predictions, and more efficient decision-making. Beneath this narrative, however, lies a profound transformation in the nature of personal information itself. Rather than merely storing data, foundation models learn latent statistical representations from vast quantities of human-generated information, transforming data into inferential capabilities. These capabilities enable AI systems to generate sensitive inferences about individuals from information that was never explicitly disclosed.
For instance, a model can aggregate purchasing behavior, social media activity, and conversational patterns to make reliable predictions about an individual’s mental health status, political affiliation, or income level. While an individual might have never volunteered to disclose that sensitive information, a more fundamental concern arises: seemingly mundane data about others can be aggregated to generate sensitive inferences about any individual in ways that are difficult for individuals to anticipate, monitor, or contest. This challenges conventional privacy frameworks built around the collection, storage, dissemination, and management of discrete, identifiable records per individual.
While recent federal data privacy proposals have expanded notions of covered data to include inferred data, they continue to conceptualize privacy harms as arising from identifiable pieces of information. When the relevant concern is no longer a piece of data that users can access, correct, or delete, but a model’s ability to generate sensitive inferences that users cannot reasonably foresee or control, existing conceptions of digital privacy begin to break down.
To ensure that future technological innovation develops within a…
