GRASP trains anonymizers to reduce what models infer from personal text

A new arXiv paper proposes GRASP, a method for improving language-model anonymizers that rewrite text to reduce sensitive personal inferences. The authors focus on attributes such as age, location, and occupation that can be guessed from ordinary writing.

The privacy problem is subtle. Removing explicit names or addresses may not be enough if a model can infer personal details from style, references, or context. Existing adversarial anonymization methods can use a strong model as both rewriter and attacker, but that may require sending private text to a third party.

GRASP explores reinforcement learning for anonymizers, aiming to make privacy-preserving rewriting more practical. The paper is not a guarantee that rewritten text becomes safe, and anonymization remains difficult to verify. But it highlights a growing need: as language models become better at inference, privacy tools must protect against what text implies, not only what it states directly.