A wearable AI product can ask its customer to obey recording laws in a few sentences. Turning that obligation into a workable system takes considerably more effort.
For builders and operators, that gap is the central issue. A company cannot responsibly introduce persistent AI memory by treating consent as an onboarding event and everything after it as somebody else’s problem. The product’s defaults determine how much information is captured, where it goes and how difficult it becomes to retrieve or remove.
The useful policy question is how to preserve legitimate memory assistance while limiting the power to monitor other people. That requires attention to the full data lifecycle, with obligations shared among the wearer, the vendor and any organization deploying the system.
Separate the decisions that a single checkbox hides
Start with capture. Is the system listening only during an intentionally started conversation, or collecting throughout the day? Can it ingest people who are not participating? Does it make its active state understandable to those nearby?
Processing is a separate decision. Audio might be handled locally, transmitted for transcription or exposed to another service. A fleeting buffer and a remotely processed stream have different risk profiles, even before either produces a saved record.
Retention introduces another set of choices. Deleting raw audio can reduce exposure, but a transcript can preserve the sensitive substance. Summaries, speaker profiles and searchable representations can carry information forward after the original recording disappears. A retention policy needs to cover the derived records, not just the largest file.
Access determines who can use the record. The wearer, an employer’s administrator, a vendor’s personnel and an integration should not be treated as interchangeable. Sharing with an assistant or business application can create a new copy beyond the original product’s controls.
Finally, reuse changes the purpose. A reminder for one person can become an employee evaluation, a marketing input or material for model training. The architecture should make those changes explicit and require an appropriate basis for each one.
These distinctions matter operationally. They do not erase the legal question at capture. A locally processed or quickly deleted interception can still be unlawful.
Keep current law distinct from the architecture we want
The United States does not have one uniform recording-consent rule. Federal law generally permits interception with one participant’s consent, subject to exceptions. Its definition of protected oral communication includes a justified expectation against interception. State requirements differ, and a federal exception does not automatically resolve them. (18 USC 2511; 18 USC 2510)
Even “all-party consent” is an incomplete implementation specification. Washington’s statute, for example, provides a route through a reasonably effective announcement to all parties, with the announcement recorded when the conversation is recorded. That is more specific than requiring a signed waiver from everyone nearby, and it should not be generalized into a nationwide announcement rule. (RCW 9.73.030)
A rollout therefore needs advice tied to its jurisdictions and actual use. Workplace policy, confidentiality obligations and downstream data protection also deserve review. A consumer device’s onboarding flow cannot answer those organizational questions.
Read vendor promises at the level of the actual data
Bee says audio is deleted after processing and interaction data is stored in a Private Compute Environment. It also says data sent to a connected service leaves that environment and may remain with that service. Those boundaries matter more than a broad description of the product as private. (Bee terms)
Omi’s current policy, effective October 31, 2026, says conversation audio may be stored encrypted for functions including playback and transcript-quality improvement, with settings controls where applicable. It also describes cloud storage of transcripts and summaries. An older statement that audio is always discarded immediately would be an unsafe basis for a deployment decision. These are the vendors’ published representations, not independent verification of any particular device or configuration. (Omi privacy policy)
Procurement should require a data-flow description, configuration evidence and testable deletion behavior. Ask what survives a conversation’s deletion, what happens to integration copies and which settings can change after an update.
Make consent workable before expanding deployment
For a bounded meeting, the operating procedure can be straightforward: explain the recording and intended use before starting, obtain whatever consent the applicable law requires, offer an unrecorded alternative and make pause controls accessible. A late arrival should trigger another check rather than silently inheriting somebody else’s decision.
Continuous ambient capture is harder. A small light may tell someone that a device is active without explaining transcription, retention or access. A person passing through a room may never see it. An employee may understand the notice perfectly and still feel unable to refuse.
The product should therefore support deliberate recording sessions and conservative defaults. Sensitive settings should be excluded unless there is a separately justified, lawful use. A wearable designed as a personal aid should not quietly become an organization’s employee-monitoring infrastructure.
Accessibility needs also belong in the design process. Some people may benefit substantially from memory assistance. Accommodations should be developed with privacy safeguards and usable alternatives, rather than assuming either a universal exemption or an automatic ban.
Legislation is addressing visibility but the problem continues after the light turns off
As of November 2, 2026, New Jersey’s A5478 proposes continuous visible recording indicators, instructions and measures against defeating the indicator. It advanced from the Assembly Consumer Affairs Committee in October but has not become law. Pennsylvania’s HB2603, amended in October, addresses wearable-device recording indicators and distinguishes incidental unintended public capture and ephemeral processing within its proposed definitions. Those distinctions are not a blanket exemption from wiretap law. (New Jersey committee statement; Pennsylvania bill record)
California’s SB1130 was vetoed on September 30. These developments show active debate, not a settled national standard. (Governor’s announcement)
Indicators can improve awareness. They cannot, by themselves, establish the legality of every capture or govern every later use. Policy needs to connect visible notice to restrictions that remain effective when the data reaches a server or another application.
Litigation is testing that separation. In an August 2026 ruling in In re Otter.ai Privacy Litigation, a federal court allowed several recording and voiceprint-related claims to proceed beyond dismissal. The decision addressed the sufficiency of allegations, not whether Otter had ultimately violated the law. It is a warning against assuming that a participant’s decision to use a tool resolves every question about the vendor’s alleged collection and reuse. (Court order)
Give personal memory a narrow and accountable place
A useful reform framework would distinguish limited personal memory from persistent monitoring and onward publication. It should reward minimal collection, short retention and strong access controls while imposing stricter obligations for sensitive conversations, identification across contexts and unrelated secondary uses. Any reform should preserve carefully drawn public-interest protections for documenting misconduct. A universal veto on recording could protect the powerful as easily as the vulnerable.
Institutions should face comparable questions about necessity, access and retention. The comparison does not require pretending that a license-plate camera and a conversation recorder collect the same thing. It requires asking each operator to justify the power its system creates.
For a business deciding what to deploy, the practical test is whether its promises survive ordinary use: someone declines, another person joins, an integration is enabled, an employee leaves or a record is deleted. If the system cannot handle those events predictably, the consent screen has done very little governance work.
Personal AI memory can be useful. Earning a place in everyday life requires controls that protect the people who did not buy the device as well as the person who did.
Companion essay: Everyone Is Recording. Who Gets Permission to Remember?.
Sources and status
Status checked 2 November 2026. These are selected US examples, not a 50-state compliance guide. The essays advocate reform; they do not describe a new personal-memory exemption. Product statements below are vendor representations, not device testing.
Current law
Federal law: 18 USC 2511 subsection 2d supplies the general participant or one-party-consent exception, with the criminal or tortious purpose limitation. 18 USC 2510 subsection 2 defines protected oral communication through a justified expectation against interception. Neither is a nationwide override of stricter state requirements.
State examples: Florida 934.03 subsection 2d and 934.02 subsection 2 illustrate all-party prior consent and the oral-communication definition. Washington RCW 9.73.030 protects private conversations and specifies an announcement route. These examples do not replace a jurisdiction-specific assessment.
Legislative proposals
New Jersey A5478: Introduced September 14, 2026; reported favorably by Assembly Consumer Affairs on October 5. Pending, not enacted. Addresses indicators, resistance to disabling them and purchaser instructions. Bill text and committee statement; official status table.
Pennsylvania HB2603 PN4012: Reported as amended, received first consideration and was laid on the table October 7, 2026. Pending, not enacted. Its proposed definitions distinguish incidental unintended public capture and ephemeral processing; those distinctions do not create a general wiretap exemption. Current text and official history.
California SB1130: Vetoed September 30, 2026. The wearable-device proposal is not enacted law. Governor’s announcement.
Litigation
In re Otter.ai Privacy Litigation, 25-cv-06911-EKL: The August 13, 2026 order, docket 68, allowed ECPA, California CIPA and Illinois BIPA claims to proceed. Other claims, including the Washington claim, were dismissed with leave to amend. This was a pleading-stage ruling, not a finding of liability. Court order.
In re Meta AI Glasses Litigation, 26-cv-01897-EMC: A September 15, 2026 order, docket 112, addressed pseudonym requests in litigation involving alleged intimate captures. This is a direct wearable example, but that procedural order does not establish the truth of the allegations or liability. Court order hosted by Justia.
Products and the institutional comparison
Bee terms, updated October 15, 2026: Required consents remain the user’s responsibility; indicators may not be obscured. Audio deletion, retained interaction data and connected-service boundaries are separate provisions.
Omi terms section 13 assigns responsibility for required consents or notices to the user. The current privacy policy, effective October 31, 2026, permits encrypted audio storage for specified functionality. Do not substitute an older blanket claim that audio is never retained.
Maricopa’s Flock FAQ describes that city’s license-plate cameras as not recording audio. The comparison in the personal essay concerns accountability for surveillance, not identical collection methods.
Independent analysis and policy commentary, not legal advice.