AI hardware can't bypass privacy: the data boundary dispute over Meta glasses, Apple Watch, and connected cars

There is always a good reason in the tech industry for people to hand over more of their privacy.

In the internet era, the reason was convenience. To hail a ride, order takeout, or stay in touch with friends, people clicked agree beneath one lengthy agreement after another. Many of these authorizations were hardly the result of careful choice—it was simply that if you refused, whatever you were trying to do right then could not proceed. In the AI era, this exchange comes with an even bigger promise: devices should understand you, models should learn about the real world, and artificial general intelligence should keep moving forward.

So cameras and microphones have been built into glasses, cars, watches, and robots, drawing ever closer to people's bodies and daily lives. For some of these products, it is worth asking again whether consumers are the ones being served, or the ones helping manufacturers obtain data. Especially when the hardware price has already been paid yet data is still treated as a resource that can be continuously extracted, this bargain is far less simple than the ads suggest.

Frankly, a lot of today's hardware is not built to serve consumers at all—it is built for "data."

Several recent news stories have pushed this unease into the foreground. Research on car data flowing to third parties has drawn attention again, Apple Watch's 15-second conversation replay has sparked controversy, and Meta, while mired in a covert-recording controversy, has launched a camera-free model and added an option for some visual data to be excluded from AI training. The common question they raise is this: tech products enter our lives very quickly—can people's refusals come fast enough?

We need to state our position clearly: there must be brakes, and the cost must be paid. Authorizations becoming more cumbersome, data collection becoming more limited, and model progress perhaps not being as fast—the industry should seriously absorb all of this. Ordinary people have no obligation to use their own lives to fill in the conditions for a company's technology roadmap.

After buying a device, what else must be handed over?

"Trading privacy for convenience" once made the internet's data business seem perfectly natural. Platforms provide free services, users contribute attention and behavioral data, and companies recoup their investment through advertising. Even if this exchange is full of information asymmetry, consumers could at least see a surface-level explanation: I'm not paying directly, so I'm paying with something else—this kind of "transfer" was still defensible.

Hardware makes that explanation much harder to sustain. Cars, glasses, watches, and home robots all have clearly stated prices; some products even require ongoing subscriptions. Users have already paid for the device and the service, so why should they still be expected to accept data uses unrelated to the current service? Why should buying a product incidentally become a commitment to the manufacturer's long-term data needs?

On September 29, The Verge reported on a connected-car privacy study conducted by Northeastern University in collaboration with Consumer Reports, covering 21 cars and 30 companion apps. The study found connections between vehicles and advertising, tracking, and analytics domains, and in some companion apps found vehicle identification numbers, location, and other information being transmitted to ad networks.

It is fair to say that many consumers have no idea where the data behind their cars goes. The car is bought and brought home, yet the data relationships revolving around it far exceed the buyer and seller.

A more direct case is General Motors. The final settlement order issued by the U.S. Federal Trade Commission in January 2026 addressed allegations that GM and OnStar collected and sold precise location and driving behavior data without adequate notice or explicit consent. The order includes a five-year ban on disclosing such data to consumer reporting agencies and requires more consent, access, deletion, and opt-out mechanisms.

Cars need to sense the road, navigation needs to know location, and voice assistants need to process commands—these all have specific purposes. What truly needs to be justified item by item are the downstream uses: why should information obtained to complete this service be retained long-term, why should it be handed to other companies, and why should it be able to enter training pipelines?

AI makes this confusion even more likely. Data can help a product complete the task at hand, and it can also help manufacturers improve models and develop next-generation products. When both kinds of value exist at once, companies have an incentive to expand collection as much as possible. Rather than guessing whether a company's "intentions are impure," analyzing its data acquisition behavior is better served by looking at the actual product design: after users refuse additional uses, how much of the service they already paid for remains?

For robots, this question is also pressing. Robots that enter homes and devices that record human operations may come into contact with room layouts, household routines, work processes, and the people who live together. Manufacturers can explain the value of this information for training, but they still need to answer: who is qualified to authorize this slice of life? If a homeowner agrees to trial a robot, can the lives of visitors, domestic workers, and other family members all be handed over along with it?

The Apple Watch case mentioned above also reminds us that privacy controversies cannot focus only on whether "the manufacturer used it for training." According to Apple's published design, Live Rewind will retroactively display text snippets from the previous 15 seconds of conversation, and Siri Recap will generate conversation summaries; both features are planned to launch in beta within the year, with users choosing to enable them. Apple says it does not save raw recordings, and that saved text and summaries are protected by end-to-end encryption.

These protections should be acknowledged—they genuinely limit data exposure. But the other party in the conversation can still ask: do I want this exchange to leave behind a machine-generated record?

Personal hardware is producing effects that extend beyond the individual. In the past, if you didn't accept an app's rules, you could still try uninstalling it; but with the glasses, watches, and robots around us, refusal often means actively interrupting a conversation, asking the other person to stop using the device, or even leaving the scene. The convenience brought by the device goes to the user, while the trouble of maintaining boundaries falls on bystanders.

Meta's repeated flip-flopping exposes who is really in charge

Among these controversies, Meta is a fitting sample for observation. Its privacy boundaries keep changing with products, policies, and external pressure.

In April 2025, Meta notified US glasses users that it was removing the option to turn off storage of voice recordings; users can still manually delete recordings in settings. By June of this year, WIRED reported that code related to facial recognition, not yet publicly enabled, had appeared in the Meta AI app and was subsequently removed. Meta responded that the feature is still exploratory and no final decision has been made.

These online rumors cannot yet be exaggerated into "Meta has already opened to the public a feature that identifies strangers at will." But they are enough to make consumers ask: when buying a pair of glasses, where exactly is the privacy boundary they are accepting? This uncertainty itself consumes trust.

For this reason, two of Meta's recent adjustments deserve recognition. The Ray-Ban Meta Audio released in September removed the camera while retaining audio and AI interaction capabilities. Foreign media noted that the camera-free model still has audio privacy issues. There is no need to attribute a pure motive to the company; the fact that consumers now have a product option that collects fewer images is a concrete improvement, regardless of whether it was forced by various pressures.

The "training opt-out" option is the same. On September 25, Business Insider cited confirmation from a Meta spokesperson that glasses users can turn off "storing visual data in AI experiences," opting out of storing this portion of data for AI training. This option applies only to visual data and does not cover voice interactions.

Here it is also necessary to distinguish ordinary photography from AI visual interaction. Meta says photos and videos taken for the album are by default saved on the device or in the user's album; asking AI about things in front of you will invoke the camera to help answer. When a user says "what is this," a different data processing flow may be involved. The features look seamless, but the authorization boundaries need to be explained with particular clarity.

Allowing users to opt out of training at least acknowledges one thing: a user invoking AI to complete a current task and allowing the manufacturer to continue using the information for model improvement can be different choices. If these two things can be separated, the choice should be returned to the user. Uses that were previously bundled together also need to be re-examined.

However, the opt-out entry point still solves only part of the problem. What is more worrying is that everyone is accustomed to relying on companies' "voluntary" choices for this over the long term. If a company is willing to provide an additional opt-out switch, that is of course welcome; whether the company can take it away again in the next update also needs constraints. The reasonable expectations users form when purchasing hardware should not change arbitrarily with model training plans and growth targets.

Meta's repeated flip-flopping ultimately brings the discussion back to a simple question: who decides where the privacy boundary lies? Manufacturers can certainly design new features, but before extending new uses into consumers' lives, they need to seek consent again. Moreover, those who have not purchased the product or registered an account should also be included in the scope of product responsibility.

"Brakes" must be built into the product's cost accounting

"For better AI" is a reason that needs to be seriously examined. It can explain why companies want to obtain data, but it cannot by itself answer whether companies have the right to obtain data. Expanding the reason to AGI and the technology race does not eliminate the authorization, boundaries, and responsibility in between.

Once refusing to provide data is described as not understanding technology or not supporting innovation, consumers bear an additional moral pressure. Ordinary people buy devices hoping to solve their own problems; they have not thereby promised to join an indefinite data contribution plan. The future technological dividends are full of uncertainty, but the life handed over today is very concrete.

For refusal to be truly meaningful, products must first accommodate refusal. Consent for training and marketing uses that are not necessary to complete the service should be sought separately and cannot be obtained through bundled authorization. If necessary sensors are turned off and related functions are affected, that can be explained; if refusing extra training also restricts unrelated functions, manufacturers need to explain why. The two should not be vaguely packed into a prompt saying "the service cannot be provided without consent."

Protection for bystanders must also move earlier into product design. Which places should restrict collection, how recording status can be made easy to notice, whether devices can be easily paused, and how recorded individuals can complain and request handling all require clear arrangements. Indicator lights and notification sounds are useful, but they are only part of notification. Making restaurant servers, office colleagues, and household visitors research the meaning of each device's lights on their own and then actively negotiate is clearly not a reasonable division of labor.

Of course, these requirements may increase costs. Collecting less data may mean giving up useful samples; adding on-device processing may require more computing power and engineering investment; establishing authorization, access auditing, and deletion mechanisms requires ongoing maintenance; allowing users to refuse may affect the scale of training data. Companies should write these expenditures and trade-offs into product budgets and calculate them just like chips, battery life, and after-sales service.

Especially for data collection companies, costs cannot be calculated only as how much each collector is paid per hour, how much equipment depreciation amounts to, and how expensive data cleaning is. Whether the collection site permits it, whether other people in the frame are informed, how private information is handled, and whether the scope of use can later be proven should all enter the delivery standards. Data whose source and authorization cannot be explained, even if cheap to collect, cannot easily be treated as a high-quality asset.

Costs also include giving up part of the revenue, as well as bearing sufficiently clear consequences for overstepping boundaries. Rectification, compensation, auditing, and restrictions on the use of the relevant data should make companies think twice when making collection decisions. Otherwise, the party that obtains data first gains an advantage, while the party that earnestly obtains authorization bears higher investment, and the market will instead reward companies with looser boundaries.

Finally, we must be vigilant that these costs paid cannot simply turn into another "privacy protection fee" charged to users. Companies need to handle costs honestly in pricing and business models, but the fact that consumers cannot afford a more expensive version should not become a reason why their lives can be additionally exploited. When it comes to other people's privacy, buyers certainly cannot price and sell it on their behalf.

The tech industry is already very familiar with how to discuss computing power costs, customer acquisition costs, and mass production costs. After entering the physical world, it must also learn to calculate the cost of respecting others. It cannot, whenever privacy is mentioned, leave the hard problems to consumers, leave the losses to bystanders, and keep the benefits brought by collection on its own books.

We can acknowledge the value of AI hardware and welcome adjustments like Meta's, while insisting on actively demanding more. A tech industry that claims to understand the physical world should have the courage to understand others' refusals and clarify the boundaries of data. Brakes must exist, and costs must be paid, even if that means collecting less data, making less money, or delivering the so-called AGI future a little later.

This article comes from the WeChat public account "AI Wuzhao," author: Peng Kunfang, published by 36Kr with authorization.