Discussion

AI agents make privacy a question of permissions, not just consent

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#10879

AI agents can expose information about people who never agreed to share it, and rules built around one person consenting to one disclosure may not be enough. MIT Sloan researchers Catherine Tucker and Rozhina Ghanavi say the gap has consequences for privacy and competition, from what agents retain to which firms can access useful data.

Watch Desk analysis

What happened

In an interview with PYMNTS, Tucker describes how an AI agent researching a medical condition could infer details about a person’s relatives from one family member’s conversation. Ghanavi argues that agentic AI shifts the question from consent to a particular piece of data towards permissions: what an agent may access, how long it may keep access and what it retains afterwards.

The researchers also point to a knotty problem with deletion. Removing material after it has been used to train or fine-tune a model may not remove the capabilities that material helped create, Ghanavi says. Their discussion connects privacy rules to competition: limits on sharing data could favour companies with established customer relationships, or encourage firms to bring applications and data inside the same organisation through acquisitions.

Why it matters

An agent given access to email, calendars or customer records can reveal patterns and relationships beyond the immediate task. Permission therefore needs to cover more than whether the software may open a file. Duration, revocation and retention matter too, while the consequences of a disclosure may fall on people who never received its benefits.

The competition angle is just as practical. If privacy restrictions make external data-sharing harder, companies with first-party information or the means to acquire it may gain an advantage. Tucker and Ghanavi argue that regulators should examine whether information is treated differently when it moves between companies than when the same activity is brought under one corporate roof.

Our read

This is a useful shift in the privacy debate: an agent’s access is not a one-off tick-box when it can act repeatedly, combine information and leave traces in a trained model. The researchers’ examples make the problem legible without pretending there is a tidy technical fix. If you are designing an agent, ask not only what it can read, but when its permission ends and what remains afterwards.

What to watch

  • Whether rules for AI agents specify how access is granted, limited and revoked.
  • How organisations handle deletion requests when data has contributed to model training.
  • Whether privacy regulators scrutinise acquisitions that bring applications and data inside one company.

Discussion spark: Should AI agents receive time-limited, task-specific permissions by default, even if that makes them less convenient to use?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.