Does CCPA apply to voice assistants and smart devices?

Yes, where a covered business collects personal information from California residents through the device. Audio recordings, transcripts, wake-word activations, device identifiers, and inferences drawn from any of these are personal information. There is no exemption for data collected through hardware rather than a website, and nothing about the interface being a microphone changes the underlying obligations.

Is a voice recording sensitive personal information under CCPA?

It can be. When voice data is processed to identify a specific individual — speaker recognition, voiceprint enrollment, or matching a speaker across sessions — it functions as biometric information used for identification, which CCPA treats as sensitive personal information carrying additional limitation rights. Storing raw audio for transcription alone is a weaker case, but the classification follows the processing purpose, and many products quietly cross that line when they add per-speaker personalization.

What about people who never set up the device?

They are consumers with rights too. CCPA rights attach to the individual whose personal information is collected, not to the account holder. A houseguest, a child, a customer at a counter, or a coworker in the room captured by an ambient device is a person about whom you hold personal information, with the same rights to know, delete, and opt out — and no relationship through which you can practically deliver notice.

Can we use voice recordings to train our AI models?

Only in ways consistent with the purpose disclosed at collection. Training a model on customer audio is generally a distinct purpose from providing the requested feature, and CCPA's purpose-limitation principle means it should be disclosed and, in practice, offered as a choice. Retroactively repurposing an archive of recordings collected for transcription into a training corpus is the specific pattern that has drawn the most attention.

CCPA and AI Voice Assistants in Connected Devices 2026: The Bystander Problem

Ambient Collection Breaks the Consent Model

Every privacy framework in wide use assumes a moment where a person encounters your product, is told what happens to their data, and proceeds. Voice assistants, smart cameras with audio, AI notetakers, and wearable recorders have no such moment for most of the people they capture. The person who bought the device consented. The people around it did not, and often do not know the device is there.

CCPA does not resolve this tension for you, and it does not exempt you from it either. The obligations attach to the collection of personal information about California residents; the absence of a practical channel to reach those residents is your design problem, not a defense. This is why compliance for ambient products lives in hardware and interaction design — an indicator light, an audible cue, a physical mute — rather than in a privacy policy nobody in the room will ever load.

What You Are Actually Collecting

Teams building voice features tend to inventory "audio" and stop. The data map is wider than that, and each layer carries its own obligations:

Raw audio — including the seconds buffered before the wake word and the false activations nobody intended.
Transcripts — often retained far longer than audio, and far easier to search, subpoena, and breach.
Voiceprints and speaker embeddings — the layer that turns audio into biometric identification.
Inferences — household composition, presence patterns, language, approximate age, emotional state, and health signals audible in a voice.
Device and environment metadata — identifiers, network, room-level location, and activation timestamps that reconstruct a schedule.

The inference layer is where products most often exceed what they disclosed. A feature that adapts responses by inferred age or mood is processing a category of information the customer was never told about, derived from data they were told was used to answer questions.

Where Voice Becomes Sensitive Data

The identification line

Transcribing speech is processing audio. Recognizing who is speaking is biometric identification, which CCPA treats as sensitive personal information with additional limitation rights. Many products cross this line the moment they ship per-speaker personalization or voice-based authentication.

Content that is itself sensitive

People say their prescriptions, their immigration status, their religion, and their finances out loud in their own homes. Audio captured in ambient settings will contain sensitive categories regardless of the product's purpose, which raises the stakes on retention and access controls independent of how the data is classified.

Children in the room

Household and retail devices capture minors routinely and without any signup path. Heightened rules for minors' data do not require you to have known a child was present, and 'we do not offer a child account' is not a data-flow control.

Overlap with biometric statutes

Voiceprint processing can also trigger dedicated biometric privacy laws with their own consent and retention requirements — and, in some states, a private right of action. CCPA compliance alone does not cover that exposure.

The Deletion Request You Cannot Fulfill

Ambient products routinely fail deletion at the architecture level. Audio is deleted while transcripts persist in a separate store. Transcripts are deleted while the derived embeddings stay in a recommendation index. Everything user-facing is deleted while an unlabeled copy sits in a training dataset that was snapshotted last quarter and has no linkage back to the requester. The right to delete is not satisfied by deleting the copy the support tool can see. Before a request ever arrives, someone has to be able to trace one utterance through every downstream store — and if nobody can draw that map, the deletion pipeline does not work no matter what the confirmation email says.

Design Controls for Ambient AI

For voice assistants, AI notetakers, smart cameras with audio, and wearables — the compliance lives in the product, not the policy page.

Give the device an unmistakable recording indicator that does not depend on an app being open— Start here

Provide a hardware mute that cuts the microphone, not a software toggle you can override— Essential

Process on-device where the feature allows it — data never sent is data never at issue— Architecture

Separate voiceprint enrollment from transcription, and make enrollment an explicit opt-in— Essential

Set a default retention window measured in days and delete false activations immediately— Essential

Map every downstream store an utterance reaches, then test deletion against that map— Essential

Keep model training as a separate, disclosed, declinable purpose — never a silent default— Purpose limits

Offer a rights path that does not require an account, for the people you captured without one— Bystanders

Publish inference categories, including the ones derived rather than collected— Notice

Your disclosures still live on a website someone has to be able to use

RatedWithAI scans your web properties for the compliance and accessibility gaps that turn into complaints and legal exposure. Start with a free scan.

Scan Your Site for Free →

Frequently Asked Questions

Does on-device processing remove CCPA obligations?

It substantially reduces them by shrinking what the business collects, which is the right instinct. It does not eliminate them: if any derived signal, transcript, or telemetry reaches your servers, that is collection, and the notice and rights obligations follow the data you actually receive.

How do we handle a deletion request from someone who was never a customer?

Build a rights path that does not assume an account. That usually means a public request form, a verification method proportionate to the sensitivity of the data, and a search capability keyed on something the requester can supply — a device they were near, a time window, or a voice sample they consent to provide for matching.

Is a beep or indicator light enough notice for people in the room?

It is necessary and not automatically sufficient, but it is far more effective than any text-based approach for people with no screen. The realistic standard is layered: an unmistakable in-the-moment cue, plus accessible detail for anyone who follows up, plus meaningful defaults for everything a bystander cannot control.

Do we need consent before enrolling a voiceprint?

Treat it as required. CCPA gives consumers rights to limit the use of sensitive personal information, and separate biometric statutes in several states impose affirmative written consent before collecting a voiceprint at all, with penalties per violation. Enrollment should be an explicit, separate, reversible choice.

What is the highest-risk pattern in this product category?

Retroactive repurposing — taking an archive of recordings collected for one feature and using it to train models. It converts a disclosed collection into an undisclosed purpose, applies to data from people who cannot be re-contacted, and is nearly impossible to unwind once the model is trained.

The consent requirements that attach to a voiceprint.

CCPA Sensitive Data and AI Inferences

When derived data becomes a regulated category.

CCPA Notice at Collection for AI Features

Where the disclosure belongs when there is a screen.

CCPA Data Deletion and AI Models

What deletion means once the data is in a trained model.