What future research directions aim to improve privacy in AI systems?

Machine unlearning, federated learning, better privacy auditing, and formal verification of prompt safety.

* One federated-learning round: only model updates leave the device; raw data stays put. *

* Federated learning keeps raw data on-device. — MarcT0K, CC BY-SA 4.0, via Wikimedia Commons. *

The active research frontier:

• Machine unlearning — remove specific data from a trained model without full retraining (e.g. to honor a deletion request or scrub leaked PII).
• Federated learning — train models without centralizing sensitive data; only model updates leave the device.
• Better privacy auditing — tools to detect and measure privacy leakage at scale.
• Formal verification — mathematical proofs for prompt safety (very early stage).

The honest summary ("are we doomed?"): no perfect solution exists, training-time privacy has huge utility costs, and inference-time safeguards can be bypassed — but we understand the threat landscape, have partial mitigations, deploy carefully with defense-in-depth, and research is active.

Tip: "Machine unlearning" is the holy grail for the right-to-be-forgotten era — retraining a frontier model from scratch to delete one person's data is infeasible, so we need to surgically forget.

Go deeper:

• Machine unlearning (Wikipedia) — the "surgically forget" frontier (the right-to-be-forgotten holy grail).
• Federated learning (Wikipedia) — the decentralized-training direction among future approaches.