AI Wearables Expand Meeting Recording Risks Beyond Video Calls

AI meeting records no longer depend on a bot appearing in a video call. A watch, pendant or phone can capture a conversation without joining a meeting platform, shifting the problem from software adoption to the physical presence of recording devices. For companies, that changes what a useful policy must cover: the microphone may already be in the room, and attendees may not know it’s there.

Organizations have sound reasons to adopt AI note takers. Teams handling many customer meetings can use searchable summaries to follow commitments and understand client needs. At AI focused startups, a new account manager may be able to review the recorded history of an account and get context from conversations that happened before they joined. That can help teams serve customers with less dependence on informal handoffs.

But convenience doesn’t settle questions about consent, security or the accuracy of the resulting record. Some companies have seen hundreds of accounts adopt note taking tools within weeks, including vendors that lack SOC 2 reports or strong encryption. The risks grow when a summary becomes the accepted account of a meeting: the software chooses what to emphasize, which tasks to assign and what future readers are likely to treat as fact.

Recording has moved beyond the meeting app

Granola, an AI note taking service used by many of the authors’ professional circles, launched an Apple Watch app on July 28. A wearer can start recording with a tap, without adding a visible bot to a call or opening a laptop. The shift is practical for users, but it removes familiar cues that a recording is underway.

That same week, DEF CON barred Meta style smart glasses from its conference floor over recording concerns. Meta had acquired AI pendant company Limitless in 2025, adding personal recording technology to a major consumer hardware business. One company the authors have worked with now checks attendees for smart glasses before all hands meetings where unreleased plans are discussed. Those responses show how device design is already changing workplace expectations.

The authors describe testing a watch by placing it beneath a jacket sleeve, starting a recording with one tap and capturing an evening walk. The transcript was clear, and nobody nearby could tell the device was recording. The example exposes the gap between a policy written for meeting software and a device that can record ordinary conversations in public or at work.

Recording rules also differ across U.S. states. California generally requires permission from all parties to record a confidential conversation, while many states allow recording with consent from only one participant. Whatever the local rule, companies still face a practical problem: people have few shared expectations about disclosure, while wearable devices are increasingly designed to attract little notice.

The transcript can be steered or distorted

An AI transcript is not a neutral account. After speaking at the [un]prompted conference in April, the authors say they heard examples of people manipulating meeting summaries. One described using prompt injection during a live meeting. Another recommended joining early and, while alone with the note taking system, calmly stating the desired conclusions so they would appear central in the final summary. A third said he had found a way to make other companies’ note takers fail silently, leaving no meeting record. The authors point to academic research on these techniques and say they can work.

Errors can arise without deliberate interference. A system may mishear a word and then interpret later remarks in a way that fits the mistaken transcription. People whose accents are less well handled by a model bear more of this accuracy risk. That creates a workplace problem beyond transcription quality: if a summary shapes follow up decisions, an error can become an error in the company’s institutional memory.

Some tools also use information from earlier meetings and background about participants to generate current notes. Granola, for example, can bring prior context into a new summary. That may save time, yet it can also carry reasoning or sensitive details into a document shared with someone who wasn’t part of the earlier discussion. A summary can therefore disclose more than the conversation it appears to represent.

A case involving a Campbell’s security executive illustrates that the recording risk predates AI. The executive was terminated last November after an employee secretly recorded a lunch conversation the executive believed was private. Months later, the recording reached an employment lawyer, and the executive was dismissed within days. The authors’ point is about the mechanism: one person with one device can create a consequential record outside any meeting platform.

Build rules around the record, the device and the people

Boards and legal teams are asking CISOs to set policies for AI note takers. The authors recommend starting with disclosure. Everyone present should know when a recording tool is active. One simple option is to add “+ AI notetaker” to the user’s display name. The same standard should apply to employees who record colleagues: undisclosed recording should carry a clear workplace consequence.

Next, companies should name meetings that cannot be recorded. Possible categories include mergers and acquisitions discussions, legal meetings, HR conversations and executive sessions of the board. The specific list will vary, but the decision belongs in a written policy. Otherwise, each employee and each vendor effectively makes it by default.

Rules should cover devices brought in by people outside the company, too. If an auditor’s laptop has to meet workplace requirements, a recording device used by an outside organization should receive similar scrutiny. Companies should establish whether visitors are recording employees and whether the company can obtain a copy of the resulting record.

Retention needs an explicit decision. A complete transcript contains more detail than a short summary and may create a different kind of corporate record. Organizations should decide whether transcripts are deleted by default, subject to legal holds, or retained indefinitely. Vendor settings shouldn’t determine that choice by accident.

Vendor review should reflect the sensitivity of the material these services handle. Companies need to assess security controls, terms governing reuse of data and responsibility for claims. The authors cite Brewer v. Otter.ai, a class action alleging that Otter’s note taker recorded meetings without everyone’s consent and used recordings to train models, as well as Novacy’s sudden shutdown. These examples underline why procurement teams need to understand both data practices and what happens to records if a vendor stops operating.

Finally, every transcript that becomes an official record should have a named human owner. That person should review it for accuracy, regardless of which attendee or organization ran the tool. Employees may be tempted to shape what a system captures, much as a lawyer protects the record in a deposition. Checking a transcript is sensible; quietly optimizing an organization’s official account of events signals a deeper problem with its record keeping.

The operational question has changed. Companies once asked whether a meeting bot could join a call. Now a recording device can accompany a person through the workday, with no visible invitation or platform notification. Policies that set disclosure expectations, define restricted conversations and assign responsibility for records give organizations a way to manage that reality before wearable recording becomes routine.