Berkman Klein fellows: bake AI privacy into the architecture

Purpose limits belong in the code, not the contract. That is the argument Isabel Hahn and Aaron Alva, both Fellows at the Berkman Klein Center for Internet & Society at Harvard University, make in a TechPolicy.Press essay laying out three principles they say should govern AI compliance: purpose limitations, accountable deployment, and transparency with use control.

"Purpose limitations must become architected into AI services, and not merely contractual," the authors write, calling for technical constraints on which tools an agent can call, which data sets it can retrieve, and whether outputs can be used for training. Their framing of agents is blunt: "That shifts AI into a category closer to delegated action, where the user is not controlling nor even observing each action."

The concrete asks are product-level. For services that store memory, Hahn and Alva want "memory dashboards that surface what the AI system remembers and allow users to edit or delete it," plus permission receipts explaining what data an agent accessed and pre-action confirmations before an agent sends a message. They point to FTC settlements against Rite Aid on pre-deployment testing, Drizly on data minimization, GM on connected-services disclosure, and Amazon Alexa on deletion rights, alongside guidance from privacy regulators in the UK, EU, Hong Kong, Singapore and Australia.

The piece closes on the alignment claim: "Unlike the internet advertising economy, where privacy and profit have often been in tension, AI presents an opportunity to align them."