Mayer's Dazzle bets photos over email for personal AI memory

Mayer hasn't disclosed what training data Dazzle uses, how it handles the massive privacy surface of processing photo libraries, or whether it's actually trained on visual data or just using images as a retrieval layer on top of existing text models. The 'visual-first' framing may obscure that the underlying inference is still language-based.

This story sits adjacent to but separate from the facial recognition debate we covered last week (the Flock competitor adding face-matching). Both involve extracting identity or behavioral signals from images, but Dazzle's pitch is about personal convenience while the surveillance vendor's move is about law enforcement capability. The real connection: as visual data becomes a primary input for AI systems, the privacy and consent questions that plague police cameras will inevitably follow Dazzle into consumer products. The difference is Dazzle users chose to upload their photos, which may feel like consent but likely isn't informed about how granular the extraction will be.

If Dazzle publishes its data retention policy and third-party access rules within 60 days, that suggests real privacy engineering. If it stays vague or mirrors typical tech company boilerplate, the visual-first positioning is marketing cover for the same data practices that made email-based assistants controversial. Also watch whether any state attorneys general request information about Dazzle's photo storage and deletion policies before launch.