The Hamburg Commissioner for Data Protection and Freedom of Information published today a final report concluding that Ray-Ban Meta AI Glasses create data protection problems for the people captured by the device's camera and microphones, not primarily for the wearers themselves. The 53-page report, dated September 10, 2026, combines a hardware teardown, an analysis of the companion app's network traffic and stored data, and a legal assessment under the General Data Protection Regulation (GDPR).

In Short

Germany's Hamburg privacy regulator took apart Meta's smart glasses and found that people standing near a wearer usually cannot tell they are being recorded, because the warning light is often too dim to notice. This matters for anyone who might appear in someone else's photo, video, or AI conversation without ever being asked. The regulator's conclusion is that wearers, not just Meta, are legally responsible for getting permission from bystanders, and that Meta's practice of using glasses conversations to train its AI models generally cannot be justified under EU law when it involves other people's faces or voices.

The investigation, carried out by the HmbBfDI's technical and legal teams, examined the first-generation Ray-Ban Meta Wayfarer model released in 2023. Investigators set up the glasses in a lab, intercepted network traffic from the companion Meta AI app, disassembled the frame to inspect its memory chip, and filed data access requests through Meta's Account Center to see exactly what the company logs. The resulting document lays out, in granular technical detail, how the glasses capture, store, and transmit data, and follows that with a lengthy legal analysis of who bears responsibility when a bystander's face, voice, or personal situation ends up inside Meta's systems.

What the technical teardown found

The Ray-Ban Meta AI Glasses combine a 12-megapixel wide-angle camera, five microphones, two speakers, 32 gigabytes of flash memory, Wi-Fi and Bluetooth radios, and a touch panel, all built into a frame styled after the classic 1950s Wayfarer. A Qualcomm processor designed for AI workloads sits inside, alongside a memory chip from the Chinese manufacturer BIWIN. Investigators could not read data directly from that chip; it lacked accessible contact points and could not be read using standard chip adapters, so the physical storage itself remained opaque to the inquiry.

Far more revealing was the companion Meta AI app, which uses the Android package name "com.facebook.stella." When investigators unpacked the app and monitored its network traffic on a Google Pixel 7 running Android 13, they found that most communication with Meta's servers was encrypted and could not be inspected. One endpoint, however, was decryptable: a diagnostic logging path that transmitted device fingerprinting data, including device model, operating system version, kernel version, and a persistent device identifier, every time the app synced.

Inside the app's local storage, the investigators found a SQLite database called StellaDatabase containing tables named "face," "face_group," "face_low_confidence_pair," and "face_to_face_group." These table names strongly suggest infrastructure for recognizing and grouping faces. Crucially, the tables were empty at the time of testing, leading investigators to conclude that no facial recognition was actively running, even though the database schema for such a feature already exists inside the software. A separate interaction log recorded every voice command cycle, timestamped in UTC, moving through states labeled LISTENING, PROCESSING, FINISHED_PROCESSING, and INACTIVE, with an AssistantResponse field capturing the text that Meta AI read back to the wearer through the glasses' speakers.

The investigators also tested how reliably the glasses distinguish real activation commands from incidental speech. The wake phrase, "Hey Meta" or "OK Meta," triggered recording almost exclusively when spoken by the person wearing the glasses, a behavior the report attributes to the directional placement of the five-microphone array. False activations were rare and limited to phonetically similar words, such as "metamorphosis," and the system was generally able to detect an accidental trigger and stop recording.

The visibility problem with the recording light

A central finding concerns the outward-facing LED meant to signal to bystanders that recording is underway. The report states plainly that this light is not a reliable transparency mechanism. Its visibility depends heavily on the distance, angle, and ambient lighting a bystander happens to be experiencing, and the investigators found that on the second-generation Gen 2 glasses, the LED does not illuminate at all when the wearer is simply talking to Meta AI about their surroundings, as opposed to explicitly taking a photo or video.

Even on the Gen 1 model examined in detail, the light glows only dimly during AI interactions, compared to a brighter, roughly one-second flash when a still photo is taken. During continuous video recording, the Gen 1 LED alternates between dim and bright, while the Gen 2 version alternates between fully off and bright, a higher-contrast pattern the report credits with somewhat better visibility. Regardless of generation, the investigators found the light essentially invisible outdoors or in direct sunlight, and functional only for people standing close to the wearer and looking directly at the frame.

The protective mechanism has a further weakness. Meta says the camera deactivates automatically if the LED's own light sensor is covered, but the investigators demonstrated that certain covering techniques, such as specific stickers or paint, let enough ambient light reach the sensor to avoid triggering the safeguard while blocking outward light entirely. They also found that during video recording, the system checks for a covered LED only at the exact moment recording starts; covering the light after recording has already begun lets it continue undetected for the rest of the clip.

Facial recognition, tested and found absent - for now

The report devotes particular attention to whether the glasses perform identifying facial recognition, meaning comparing a captured face against a database of known people rather than simply detecting that a face is present as an object in the frame. The investigators distinguish carefully between the two: object-level face detection, used to apply a masking circle over faces before some images are stored locally, is not the same as biometric identification.

Based on their technical testing, the HmbBfDI could not confirm that identifying facial recognition was active on the device, despite database structures that would support it. The report does note, however, that outside researchers reached a different conclusion earlier in 2026. According to the Electronic Frontier Foundation and a WIRED investigation from June 2026, facial recognition software was present and could be manually triggered to identify people the glasses had previously encountered, before Meta modified the software in late June 2026 to prevent that behavior from being reproduced. The discrepancy between the two findings likely reflects the difference between a fixed software build at the time of Hamburg's testing and a feature that briefly existed and was then patched.

The legal significance of facial recognition, were it enabled, would be substantial. Biometric data used to uniquely identify a person falls under Article 9 of the GDPR as a special category of data, subject to a general prohibition unless a narrow exception applies. The report states that if the facial recognition feature were switched on, both the glasses' wearer and Meta itself would face the strict requirements that provision imposes.

Why bystanders, not wearers, are the focus

The report's legal analysis centers on an argument that will be familiar to anyone who has followed how European regulators have approached bodycams and dashcams: the household exemption under Article 2(2)(c) of the GDPR, which excludes purely personal or family activity from data protection rules, does not reliably cover what these glasses do in practice. A grandparent filming a birthday party is one thing. A commuter wearing the glasses on public transport, capable of triggering an AI assistant to analyze strangers' faces, clothing, or conversations, is functioning more like an operator of a body-worn camera than a family photographer, according to the assessment.

The investigators lean on a body of case law from the Court of Justice of the European Union, including the Ryneš, Lindqvist, and Buivids rulings, to establish that the household exemption must be interpreted narrowly and does not extend to situations where personal data becomes accessible to an indefinite group of people or where third parties suffer a foreseeable disadvantage, such as covert recording. Publishing images to Facebook or Instagram removes any doubt: publication to an indefinite audience takes the activity outside the household exemption entirely, a principle the report grounds in the CJEU's 2019 Buivids judgment.

Where the glasses' camera and microphones pick up an identifiable third party, whether through a recognizable face, distinctive clothing, a visible name badge, or simply being specifically discussed with Meta AI by name, the report treats that as processing of personal data requiring a lawful basis under Article 6 of the GDPR. The wearer, as the person who chose to activate the device and aim it at a particular scene, is treated as the data controller for that processing, separate from and in addition to whatever role Meta itself plays.

The report works through three separate usage scenarios; taking photos and video, using the AI assistant, and posting to social media, and evaluates two possible legal bases for each: consent and legitimate interest.

Consent, under Article 6(1)(a), requires a clear signal delivered before processing begins, in line with a 2025 CJEU ruling on bodycams that the report cites directly, involving a Swedish transit authority. That ruling established that a data subject must receive the essential facts, such as who is filming and why, at the moment of collection, not afterward. Applied to smart glasses, the investigators concluded that a dim, easily overlooked LED does not meet this bar. Effective consent would require wearers to take active additional steps: verbally announcing that recording is underway, wearing a visible sign, or otherwise ensuring bystanders understand not just that a photo is being taken, but that the resulting data might be sent to Meta AI or, if AI training remains enabled, incorporated into Meta's model development pipeline.

Legitimate interest, under Article 6(1)(f), fares little better as a general-purpose justification. The report applies the three-part test that the CJEU has developed, requiring a genuine interest, a demonstration that the processing is necessary to achieve it, and a balancing exercise weighing that interest against the bystander's rights and reasonable expectations. Because the glasses can record covertly and without the transparency ordinary photography with a visible smartphone provides, the report concludes that this balancing test will favor the affected third party in the large majority of everyday scenarios, unless the wearer takes independent steps to raise awareness.

One narrow exception the report identifies involves visually impaired users, for whom navigation, object identification, and social participation may constitute a legitimate interest that, on a case-by-case basis, could outweigh a bystander's competing interest, particularly given that such users are often independently recognizable through a cane, harness dog, or armband. The report is careful to note this cannot be generalized as a blanket rule and still requires individual assessment.

The report's sharpest conclusions concern what happens when a wearer has not opted out of Meta's AI training program, which Meta refers to internally as generating "Flywheel data." Since April 2025, Meta has structured AI training on European users' interactions as opt-out by default, meaning content and voice interactions are used for training unless a user actively objects, a policy PPC Land covered when Meta first announced the change and Meta began implementing it on May 27, 2025. According to the report, Meta only offers third parties, people who appear in a glasses wearer's interactions but never consented to anything themselves, objection mechanisms that do not apply retroactively and are considerably more cumbersome to use than the wearer's own opt-out.

The report states that once AI training is enabled, the household exemption cannot be sustained at all, because a wearer's private use of the glasses is no longer contained within their own personal sphere; the data is transferred to a global technology company for a purpose, model development, that exists independently of the wearer's private activity. Citing a 2024 opinion from the European Data Protection Board on AI models, the investigators note that personal data absorbed into training data can become embedded in model parameters and, under some circumstances, extracted or regurgitated through later interactions with the model, meaning the intrusion is not a single, contained event but a potentially permanent and uncontrollable one.

Because of that, the report concludes that when AI training is active, the glasses' wearer and Meta become joint controllers under Article 26 of the GDPR with respect to third-party training data, a status the CJEU has defined functionally across several rulings, including the Wirtschaftsakademie, Jehovan todistajat, and Fashion ID decisions. Joint controllership does not relax the need for a lawful basis; it means both parties must be able to point to one. The report finds that neither consent nor legitimate interest realistically clears that bar for uninvolved bystanders. Consent fails because bystanders are typically unaware any AI processing, let alone training, is occurring. Legitimate interest fails the balancing test because affected third parties did not supply their own data, have no relationship with Meta, cannot meaningfully object in advance, and face processing that extends into an unpredictable, largely irreversible training pipeline.

The report distinguishes this situation from a May 2025 ruling by the Cologne Higher Regional Court, which found that Meta could rely on legitimate interest to train AI models using public Facebook and Instagram posts from registered adult users, a case PPC Land covered in detail at the time. The Hamburg report argues that ruling turned on facts specific to a platform relationship: registered users, information campaigns, opt-out mechanisms, and content those users had deliberately made public. Data picked up incidentally by smart glasses in physical space lacks all of these features. There is no platform relationship with the bystander, no realistic awareness of processing, and no meaningful opportunity to object before the data has already reached Meta.

Deletion and transparency gaps inside the app itself

Beyond the core legal findings, the technical investigation surfaced several practical shortcomings in how the Meta AI app handles user control. During testing in autumn 2025, investigators found that chat histories with Meta AI could not be deleted in bulk; a "Delete All Chats" feature was only added after a subsequent app update in spring 2026. Even after that update, the report notes that deleting chat history does not remove the same queries from a separate log called the "Voice Activity Log," located in a different menu entirely. A user seeking to fully erase their interaction history must clear both locations independently, a design the report treats as a transparency and control failure rather than a minor inconvenience.

The report also found that withdrawing consent to voice recording storage does not retroactively delete previously stored recordings; it merely stops new recordings from being saved going forward. And the "Data & Privacy" section of the app, where users might expect to find privacy controls, offers only a data export option; the actual settings to revoke voice interaction storage sit several menus deep under Device Settings.

Broader context for wearable AI regulation

The Hamburg findings arrive amid sustained regulatory attention to how Meta justifies AI training under European law, a debate PPC Land has tracked closely since Meta first proposed using EU personal data for AI training in April 2025. Separate German court proceedings later found that Meta's AI training program captures children's data despite platform protections aimed at minors, and privacy group noyb has continued pressing the Hamburg data protection authority on other matters involving biometric scanning. A survey PPC Land covered in August 2025 found that only seven percent of German Meta users actually wanted their personal data used to train AI models, a figure the Cologne court's own reasoning had to weigh against Meta's claimed legitimate interest.

Meta has continued expanding the smart glasses line commercially through the period covered by this report. PPC Land reported in November 2024 on Meta's expansion of the AI assistant to France, Italy, Ireland, and Spain, and more recently on Meta's donation of 15,000 units to visually impaired adults in Ireland through a national disability services program, illustrating the assistive-technology use case the Hamburg report treats as a genuine, if narrow, candidate for legitimate interest. Separately, Meta confirmed in October 2025 that interactions with Meta AI would begin feeding advertising personalization from December 2025 onward, a development that adds a commercial dimension to the training question the Hamburg report examines primarily through a data protection lens.

For advertising and marketing professionals, the report's relevance extends past the glasses themselves. Wearable cameras with always-on AI assistants represent an early instance of a broader pattern: consumer hardware capturing environmental data about people who never interacted with the product or its manufacturer, then feeding that data into commercial AI systems under legal bases designed for more conventional data relationships. The Hamburg analysis, by concluding that neither consent nor legitimate interest reliably justifies bystander data flowing into AI training, sets out a framework that regulators elsewhere in Europe, and potentially companies building competing wearable AI products, will need to reckon with as the product category grows.

Timeline

  • September 27, 2023 - Meta and Ray-Ban announce the first-generation Ray-Ban Meta smart glasses.
  • April 7, 2025 - Meta announces it will use public posts and Meta AI interactions from adult EU users to train its AI models, with an opt-out deadline of May 27, 2025.
  • May 13-14, 2025 - A German consumer association files for an emergency injunction against Meta's AI training plans; noyb separately sends Meta a cease-and-desist letter, as PPC Land reported.
  • May 23, 2025 - The Higher Regional Court of Cologne rejects the injunction request, finding Meta has a legitimate interest under Article 6(1)(f) GDPR for training on public profile data, with the Hamburg data protection authority participating in the hearing.
  • May 27, 2025 - Meta begins using public content from EU users for AI training.
  • August 2025 - A survey finds only 7 percent of German Meta users support having their data used for AI training; a separate German court ruling confirms Meta's training program captures children's data.
  • November 18, 2024 - Meta expands the glasses' AI assistant to France, Italy, Ireland, and Spain.
  • March 24, 2026 - The HmbBfDI submits its first data access request through Meta's AI Account Center.
  • April 16-24, 2026 - The HmbBfDI requests and downloads a second data export, including device and data logs.
  • June 2026 - The Electronic Frontier Foundation and WIRED report that facial recognition software was briefly present and exploitable on the glasses before Meta modified the software.
  • July 2026 - Meta donates 15,000 units of the glasses to visually impaired adults in Ireland.
  • September 10, 2026 - The Hamburg Commissioner for Data Protection and Freedom of Information publishes its final technical and legal review of the Ray-Ban Meta AI Glasses.

Summary

Who: The Hamburg Commissioner for Data Protection and Freedom of Information (HmbBfDI), Germany's regional privacy regulator, investigated Meta Platforms Ireland Limited's Ray-Ban Meta AI Glasses, with particular focus on the legal responsibilities of glasses wearers toward third parties captured by the device.

What: A combined technical and legal review found that the glasses' recording indicator light is frequently invisible or absent to bystanders, that no active facial recognition was detected despite supporting database structures, and that neither consent nor legitimate interest generally provides a lawful basis for transmitting bystanders' data to Meta AI, particularly when that data feeds AI model training.

When: The final report is dated September 10, 2026, following a technical investigation that included data access requests submitted in March and April 2026 and reference to external reporting from June 2026.

Where: The investigation was conducted in Hamburg, Germany, and its legal analysis applies GDPR standards relevant across the European Union and European Economic Area.

Why: The report matters because it establishes that glasses wearers, not only Meta, bear direct data protection liability for recording and AI-assisting with bystanders who never consented, at a moment when wearable AI devices are moving from novelty into mainstream consumer hardware, and while Meta's broader AI training practices already face active legal challenges across Germany and the wider EU.