A person wearing an AI device may know that the device is listening. The person sitting opposite the table cannot do this. This gap is becoming increasingly difficult to ignore as wearables move from tracking steps and heart rate to understanding conversations, recognizing their surroundings, and turning what they hear and see into information.
The new Apple Watch Series 12 provides a useful example. Its Live Rewind feature can display a text snippet of what was said in the previous 15 seconds, and Siri Recap can create overall conversation summaries. Both functions are designed to bypass the user’s conversations rather than simply respond to a command.
Almost at the same time, Meta is reportedly preparing a camera-less version of its smart glasses, codenamed Luna. As reported by Reuters, citing a report from The Information, the glasses are expected to remove the camera, but will retain microphones, speakers and artificial intelligence capabilities. The reported move comes amid growing attention to the privacy implications of camera-equipped wearables.
These two developments approach the same problem from different angles. The wearable device better understands what people are saying. The other reportedly removes one of the more obvious sensors associated with privacy concerns.
The question is what happens to everyone else who is within range of these devices.
The person wearing the device is only part of the conversation
Apple says the Live Rewind feature is voluntary and only activates when the wearer double-presses the Digital Crown. It captures the previous 15 seconds of speech, displays a snippet of text, and does not continuously record conversations in the background. Siri Recap is also user-controlled and can be scheduled for a specific time or location.
Apple has also built several privacy measures around these features. Live Rewind is processed on the device, and Siri Recap uses Apple’s private cloud computing for some of the processing. Apple says that saved Live Rewind summaries and snippets are fully encrypted via iCloud and that Siri Recap data is automatically deleted after seven days if the user doesn’t save it.
These controls give the user some information about when the function is active and what happens to the data received. The person sitting opposite the owner has different circumstances.
They may not have an Apple Watch. They may not be aware that the conversation is being processed. They may be comfortable being heard, but not when their words are transcribed into notes, stored, summarized, or potentially analyzed by an artificial intelligence system.
This creates what might be called witness confidentiality. The wearable device is owned by one person, but its sensors operate in an environment occupied by multiple people.
Dhananjay Yadav, co-founder and CEO of Indian wearable AI startup NeoSapien, said the industry needs to recognize this difference.
“The person who buys a wearable device is not the only one affected,” Yadav said. “Their agreement to the terms of the app cannot reasonably represent the interests of everyone they later meet.”
According to Yadav, consent should also be divided according to purpose. Someone might agree to have a meeting recorded but object to his emotional state being analyzed or his words being used to train an AI model.
“They are different permits,” he said.
The microphone can be as important as the camera
The privacy debate surrounding smart glasses has largely focused on the cameras.
This concern has already reached regulators. The Irish Data Protection Commission contacted Meta about its Ray-Ban smart glasses, raising concerns about whether people seized by the glasses would receive adequate notice. Following the interaction, Meta increased the size of the external privacy LED and added a flashing mode when recording.
The meta continues to face scrutiny over its record indicator. Earlier this month, The Verge reported that the company was introducing another privacy measure after users found ways to block the recording LED: The glasses now turn off the camera if the LED is covered.
However, the camera is just one way a wearable device can monitor its surroundings.
The microphone can pick up the conversation without creating a visual recording. AI can then turn that audio into searchable text, extract names, identify decisions, summarize discussions, or perhaps gain insights from how something was said.
This is where the shift to artificial intelligence changes the privacy equation.
The UK Information Commissioner’s Office has previously warned that immersive technologies such as smart glasses could intentionally or unintentionally track people near the device. It also highlighted the problem of increasingly covert devices making covert surveillance easier.
The privacy issue becomes more complicated when the device interprets rather than simply records. A photograph can become a description of a crime scene, an identified person, or an extracted piece of text. The audio can also serve as a transcript, summary, or summary of what was discussed. In this way, AI can create multiple layers of information in a single interaction.
Removing the sound does not necessarily solve the problem.
Gathering information is only the first stage. What happens to that information afterwards creates another set of privacy questions.
Yadav said companies need to distinguish between protecting information once it is captured and determining whether the information should have been collected in the first place.
“Protecting captured information and obtaining permission to collect it are separate responsibilities,” Yadav said.
This distinction is important because deleting the original record does not necessarily delete everything that was created from it. A transcript or summary may retain the other person’s words after the audio itself is deleted.
The UK ICO, in its 2026 guidance for consumer IoT products, said that consent should be real, specific and freely given, while privacy information should be provided in plain language and data collection should be limited to what is necessary.
For wearable AI, this principle becomes more complex because the person whose information is collected may not be the customer.
A wearable device may record a conversation that is important to its owner, but the conversation itself involves other people whose words become part of the final transcript or summary. This makes preservation another part of the witness confidentiality issue, especially when information obtained may persist beyond the initial interaction.
Recording light is useful, but it has limitations.
A visible indicator is an obvious solution. If a light comes on while the wearable device is recording, people nearby will at least get a signal that something is happening.
Apple took this approach with Live Rewind. The company says that when activated, the feature emits a beep, displays a microphone indicator, and displays a full-screen animation.
Yadav also sees indicators as an important starting point. He said Neosapien’s Neo1, a wearable pendant device that can transcribe and summarize speech, uses white light to indicate active listening and has a mode that allows capture to be paused.
But the indicator solves the problem of attention, and not necessarily the problem of choice.
A person may know that a device is recording but have no practical way to prevent their voice from being intercepted. In a restaurant, office or public transport, leaving may be impossible.
The ICO’s guidance on body worn recording makes a similar distinction. It recommends clear privacy messaging, visible indicators and storage controls, while warning that continuous recording can capture people going about their daily lives and can be excessive.
For consumer AI wearables, this means that the recording indicator may not be the entire privacy system.
Yadav said security measures should cover “what is captured, what is intended, where it is processed, how long it is stored and who can access it.”
He also argues that companies should provide a practical way for bystanders to raise concerns or request removal without requiring them to become customers.
Wearable devices may have to be designed for a person sitting at a desk.
The basic idea behind AI wearables is that the device can remember what the user forgets, understand what they see and hear, and reduce the need to interact with the phone.
But the environment around the user is part of the input.
This makes bystanders part of the product experience, whether the manufacturers intended it to be or not.
Yadav said companies should test products on people sitting opposite the user, and not just on customers. Their ability to understand what’s going on and say no should be part of the product design.
This could mean clearer recording signals, easier ways to pause capture, shorter retention periods, local processing, automatic deletion, and restrictions on reuse. It could also mean designing devices that minimize extraneous speech, rather than simply picking up whatever is in range of the microphone.
