Share with your CISO
Apple, Google, and a growing range of AI hardware makers are quietly redefining what counts as a recording, and the legal and compliance implications haven’t caught up. Apple’s rumored smart home camera would process visual data into text descriptions without storing footage, and its Audio Intelligence features on Apple Watch already delete raw audio after on-device processing. Google is exploring the same architecture for smart glasses. The central question Victoria Song raises is whether derived outputs like transcripts and summaries count as recordings under any existing legal framework.
What this means for your business
Every enterprise workplace consent policy, every two-party recording law compliance checklist, and every data retention schedule your legal team has signed off on was written with a binary assumption: a device either records or it doesn’t. That assumption is now structurally broken. An employee wearing Apple Watch Series 12 with Siri Recap enabled is generating conversation summaries throughout the workday. Whether that triggers your recording consent obligations depends entirely on a definition that nobody in your jurisdiction has officially updated.
The “no raw audio saved” framing that Apple uses is technically accurate and legally untested at the same time. The derived output, a summary of what was discussed in a meeting, is preserved and reviewable. Courts applying wiretapping statutes or state all-party consent laws have never had to rule on whether an AI-generated summary of a conversation is itself a record of that conversation. The first case to test this will almost certainly involve an employee, a workplace dispute, and a smartwatch. Your policy needs a position before the case law does, not after.
The harder problem for enterprise security leaders is that the hardware is becoming invisible as a signal. A Vocci Ring with a microphone and an Oura Ring without one are indistinguishable at a glance. Meta’s privacy light uses white, not the universally recognized red, and it’s easily obscured. If your sensitive facilities or boardrooms have a “no recording devices” policy, you need to reckon with the fact that enforcement now depends on either trust or detection technology that doesn’t yet exist at scale. The budget question isn’t whether to update the policy; it’s whether the policy is enforceable at all without a new technical control layer.
Concept deep-dive: On-device AI processing with derived outputs
Traditional recording captures and stores raw data, audio or video, that can be replayed. On-device AI processing instead runs the raw data through a model locally, discards the original, and keeps only the model’s output, a transcript, a summary, a label. Think of it as the difference between photocopying a document and writing a note about what the document said. The legal and privacy frameworks built around the photocopy don’t automatically apply to the note, which is exactly the ambiguity these products are designed to exploit.
Based on reporting from We can’t just change the definition of ‘recording’, originally published 2026-10-06 12:29:00.
