I am going to make two arguments in this piece, and I want to be clear at the outset that they are independent of each other. Either one, on its own, is a serious problem for Amazon. Together they describe a company that made the worst available choice on both the privacy axis and the environmental axis, for a feature nobody asked their neighbours' permission to point at them.
The feature is Ring "Familiar Faces". Amazon began rolling it out to the United States and Canada in December 2025, with the usual reassurances: optional, off by default, the customer is in control [1]. What Amazon did not lead with - what you have to read the engineering and the legal correspondence to establish - is that the facial recognition runs in Amazon's cloud, not on the doorbell, even though the doorbell is carrying silicon perfectly capable of doing it locally. That single architectural decision is the spine of both arguments below. It is the reason the GDPR exposure is as wide as it is, and it is the reason the carbon argument exists at all.
So let me take the two arguments in order. First, what Familiar Faces actually does, and why it is a data protection problem of a kind the household-activity defence does not reach. Second, what it costs the atmosphere to run biometric matching in a datacentre that did not need to happen in a datacentre; and why that puts it in the conceptual territory the Environmental Crimes Directive was written to address.
What Familiar Faces actually does
When the feature is enabled on an eligible Ring device, the camera scans the face of every person who comes into view, converts that face into a mathematical template - a faceprint - and compares it against a catalogue of up to 50 people the device owner has tagged in the Ring app [2]. If there is a match, the owner gets a notification that says "Emma at Front Door" instead of "person detected". If there is no match, the system still generated and compared a faceprint to find that out.
Read that sequence again. To decide whether an approaching person is one of the 50 tagged faces, the system has to biometrically process the face of every single person who approaches - the tagged neighbour, the un-tagged neighbour, the delivery driver, the canvasser, the kid selling biscuits, the meter reader, and whoever happened to walk past on the public pavement within range. The 50-person catalogue is the owner's. The biometric processing is performed on everyone.
That processing does not happen on the doorbell. Ring's own support documentation states that the facial recognition information is encrypted and stored in the cloud, not on the Ring devices [3]. The class action filed against Ring in the United States this month pleads the same thing: Familiar Faces converts facial images into faceprints that are stored in Amazon's cloud rather than on the device [4]. And in Amazon's written answers to a US Senator, which I will come to, Amazon confirms the biometric data is processed in the cloud [5].
So the picture is: the camera captures the face, the image (or a derived representation) travels to AWS, AWS generates the faceprint, AWS does the match against the catalogue, AWS sends the answer back. The biometric template of a non-consenting passer-by is created and held on Amazon infrastructure.
The hardware could have done this on the device. Amazon chose not to.
Ring's higher-end devices run Ambarella CVflow system-on-chip silicon. The Ring Doorbell Pro 2 is documented as operating on the Ambarella CV25M [6]. The CV25 is not a generic video encoder with AI bolted on as a marketing line - it is a purpose-built computer-vision SoC, and Ambarella sells it on exactly this use case: on-device deep-neural-network processing for detection, classification and tracking, in an extremely low-power design optimised for wire-free, battery-powered cameras with long battery life [7]. Ambarella ships the tooling to let manufacturers port their own neural networks onto the chip using standard frameworks, and Amazon itself integrates its AWS edge-ML toolchain directly with the CV25 [8]. The CVflow family runs on-device inference in the range of roughly 2 to 8 trillion operations per second depending on the part [9].
Face detection is trivial for that silicon. A face-matching model against a catalogue of at most 50 templates - which is a tiny gallery, a handful of vector comparisons - is well within reach of a quantised embedding model running locally. The whole point of the CV25 is that you can do this on the camera, in under two watts, without sending anything anywhere [7].
Amazon sent it to AWS anyway.
That is the choice. The hardware on the device the customer already owns is capable of generating the faceprint and doing the match locally, ephemerally, without a single byte of biometric data leaving the home. Amazon instead architected the feature so that every face - including every non-consenting face - is processed in its datacentre and the faceprint is stored on Amazon infrastructure. This is not a hardware limitation that forced a cloud design. It is a cloud design chosen in spite of the hardware. Hold that thought; it is load-bearing for both halves of this article. Amazon only uses the CV25 platform to determine whether the shape approaching the door is human, and routes everything biometric to its datacentre - despite the device already running on-device computer vision for that detection step, and despite an on-device, ephemeral design being available that would never persist a stranger's faceprint at all.
That on-device, discard-on-no-match design is not a thought experiment. In a research project funded by RISE Research Institutes of Sweden, which I co-led, we demonstrated real-time redaction of the incidental capture of individuals by smart cameras - suppressing not only visual identifiers such as faces but also accidentally captured audio - as the frames were processed. The probe still has to be extracted to test it against the catalogue, so this does not make the comparison step disappear; but it does establish, with publicly-funded research behind it, that real-time on-device handling and immediate suppression of everything that is not a match was a solved problem years before Familiar Faces shipped. The hardware proves the match could have been local. The RISE work proves the discard could have been engineered. The persist-by-default cloud design was a choice, not a necessity.
Why the household exemption does not save anyone
The reflexive defence to any home-camera privacy complaint is the "purely personal or household activity" exemption in Article 2(2)(c) GDPR [10]. A homeowner pointing a camera at their own front door, for their own domestic purposes, sounds like the paradigm case. And it is worth being clear up front: for the homeowner, in narrow circumstances, that defence might have some life in it. For Amazon, it is dead on arrival, and it is important to see why, because this is the point that disposes of the more important defendant before any of the field-of-view argument even starts.
The household exemption is a defence about the nature and purpose of the processing, not about who happens to operate the camera. It exempts processing carried out in the course of a purely personal or household activity. Amazon's activity is not, and can never be, a personal or household activity - it is the commercial activity of an undertaking. In Lindqvist (C-101/01), the Court of Justice construed the carve-out narrowly and tied it to activities carried out in the course of private or family life, holding that the charitable and religious activities of a parish catechist who published colleagues' personal data on a website fell outside it [11]. The commercial-operator consequence is not subtle: a company processing personal data in the course of its business cannot be engaged in a "purely personal or household activity", by definition. The exemption is simply unavailable to it.
Amazon's activity here is commercial through and through. Running the biometric matching pipeline in AWS is a commercial activity. Storing the faceprints on Amazon infrastructure - even when the company frames it as doing so "for the customer" - is a commercial activity; the storage is a paid service feature of a subscription product (Familiar Faces requires a Ring Home Premium plan) [12], performed by a profit-making undertaking on its own infrastructure for its own business purposes. Building, training, testing and tuning the recognition model is a commercial activity. None of that is the operator pottering about in their private life. It is a multinational processing biometric data at scale in the operation of a paid product. The household exemption could not attach to it under any reading of Article 2(2)(c) as Lindqvist interprets it.
This matters because it means the analysis does not depend on where the camera is pointed. Even the most charitable hypothetical - a camera that sees only the owner's own hallway, owner enrols only their own family, no stranger ever appears - does not put Amazon inside the exemption, because Amazon's processing is commercial regardless of the field of view. The exemption is the homeowner's to argue, in limited cases, and never Amazon's. So the joint-controllership analysis below is not the fallback; for Amazon it is the whole of it.
The field-of-view point is the homeowner's problem, and it is worth stating because it removes the exemption from the customer too in the ordinary case. In Rynes (C-212/13), the Court held that a home CCTV system whose field of view extends, even partially, to a public space or to a neighbour's property is not covered by the household exemption, because the processing reaches data subjects outside the operator's private sphere [13]. Familiar Faces is Rynes with a biometric upgrade. A doorbell camera, by its entire purpose, watches the approach to the door - the path, the step, the pavement. It does not merely record those people, it runs facial recognition on them and turns them into biometric templates. So in the typical deployment the customer loses the exemption too, the moment the lens reaches the public space. But keep the two points separate: the customer can lose the exemption on the facts; Amazon never had it to lose, as a matter of law.
And there is a category of subject for whom the exemption could never have applied to anyone in the first place: the people Amazon's own documentation calls "unnamed" faces. These are the faces the system scanned, generated a faceprint for, failed to match against the 50-person catalogue, and then retained. More on the retention in a moment, but note the point of principle now - these are by definition not the household, not the tagged friends and family, not anyone in a relationship with the device owner. They are strangers whose biometrics were processed and stored. There is no reading of Article 2(2)(c) that covers the systematic biometric profiling of unidentified strangers.
Once the household exemption is gone, you are in full GDPR scope, and the data in question is biometric data processed for the purpose of uniquely identifying a natural person - special category data under Article 9(1), which is prohibited unless one of the Article 9(2) conditions is met [10]. For the non-consenting passer-by, none of them is met. There is no explicit consent (they were never asked). There is no other Article 9(2) condition that fits a doorbell scanning strangers. The processing is, by default, prohibited.
Joint controllership: Amazon is not a bystander
Amazon's entire public posture pushes the legal responsibility onto the homeowner. When a customer turns the feature on, Amazon shows an in-app message reminding the customer that they should comply with applicable laws that may require obtaining consent [3][5]. That is the move: the customer is the controller, the customer must get consent, Amazon merely provides the tool.
This does not work under EU law, and Fashion ID (C-40/17), Wirtschaftsakademie (C-210/16) and Jehovan todistajat (C-25/17) are the reason it does not work. The CJEU has held, repeatedly, that where a party determines the purposes and means of processing jointly with another, it is a joint controller, and that joint controllership does not require equal responsibility or equal access to the data - it requires participation in determining the why and the how [14][15][16].
Apply that here. Who determined that the means of processing would be cloud-based rather than on-device? Amazon. Who determined that the faceprint would be generated and stored on Amazon infrastructure? Amazon. Who determined the retention period for unnamed faces? Amazon. Who built, trained, tested and tuned the recognition model, and decided the catalogue cap of 50, the matching threshold, and the alert behaviour? Amazon. The homeowner determines one thing: who goes in the 50-person catalogue. Amazon determines essentially everything else about the purposes and means.
That is not a processor relationship. A processor acts on the controller's instructions and does not determine purposes and means. Amazon determined the means so completely that the homeowner could not change them if they wanted to - the homeowner cannot move the processing onto the device, cannot shorten the unnamed-face retention, cannot inspect the model. Amazon is a joint controller, and as a joint controller it carries its own Article 9, Article 5, Article 6, Article 13/14 and Article 25 obligations directly, to every data subject the system processes - including the strangers [10]. The in-app reminder to the customer does not discharge a single one of Amazon's own obligations. It is, if anything, evidence that Amazon knew the consent problem existed and chose to address it by contractually pushing the liability on to the customer.
There is a further dimension to that offloading that sits outside data protection law entirely. By telling the customer, through an in-app notice, that the customer must obtain consent and the customer must comply with applicable biometric law, Amazon misrepresents where the legal responsibility actually sits. Amazon is a joint controller and bears those obligations itself; presenting them as the customer's alone gives the average consumer a false picture of their own liability. That is capable of being a misleading action and a misleading omission under Articles 6 and 7 of the Unfair Commercial Practices Directive (Directive 2005/29/EC), because it is liable to cause the consumer to take a transactional decision - enabling the feature, keeping the subscription - that they would not have taken had they understood that the legal exposure was shared with, and in large part borne by, Amazon [17]. The same conduct engages the Unfair Contract Terms Directive (Directive 93/13/EEC): a term that purports to shift a trader's own statutory liability onto the consumer is close to the paradigm of a term causing a significant imbalance in the parties' rights and obligations, contrary to the requirement of good faith, and is therefore liable to be found unfair and non-binding [18]. One regime governs the practice of telling the consumer the false thing; the other governs the term that does the shifting. Both point the same way, and both are concerned with the consumer relationship specifically - they sit alongside, not inside, Amazon's separate and undischarged obligations to the people the cameras actually scan.
Amazon's answers to Senator Markey, read the way a lawyer reads them
On 30 October 2025, US Senator Ed Markey wrote to Amazon with a list of pointed questions about Familiar Faces. Amazon replied on 21 November 2025, over the signature of Brian Huseman, Vice President for Public Policy [5]. That reply is a public document, and it is far better evidence than any press paraphrase, because it is Amazon's own careful wording about its own conduct. The careful wording is the story.
Markey asked (question 1e) whether Amazon uses biometric data to "train machine learning models or improve facial recognition algorithms", and whether subjects are informed and able to opt out. Amazon's answer, verbatim:
"Outside of limited data sets where participants have expressly consented to the use of their Ring data, for example beta trials, Ring does not use its customers' biometric data for algorithm or machine learning model training purposes." [5]
I have seen this reported as "Amazon says it does not use the data to train AI models". That is not what the sentence says. The sentence is hedged three separate ways, and each hedge is a door left open.
First, the consent carve-out is an admission, not a denial. The sentence opens by exempting "limited data sets where participants have expressly consented ... for example beta trials". So Amazon does use Ring biometric data for training, in the cases where it considers consent present. "For example" means beta trials are an illustration, not the boundary of the carve-out. The headline reads as "we do not train on faces". The actual claim is "we do not train on faces, except in an undefined set of cases we have decided are consented".
Second, "its customers' biometric data" is doing enormous work. Markey's question expressly named "passersby, delivery workers, and guests". Amazon's answer narrows to its customers' data. The faceprints generated at the door are overwhelmingly not customers - they are visitors and strangers. The sentence says nothing about the biometric data of the non-customers. The scope of the promise is drawn precisely around the population with the strongest legal claim and the least ability to enforce a customer agreement. That is not an accident of drafting. That is the drafting.
Third, "training purposes" is not "improvement". The question asked about training or improvement. The answer addresses only "training ... purposes". Evaluation, benchmarking, threshold tuning, quality assurance and bias testing are not "training", and Amazon admits to doing exactly those things elsewhere in the same letter. In its answer on bias (question 2a) it describes "systematic data collection and testing across diverse demographic groups, including gender, ethnicity, age, and appearance variations", while declining to publicly disclose the test results [5]. So Amazon collects and processes facial data to improve and validate the system. It just does not file that activity under the word "training".
Two more facts from the same letter complete the picture, and they matter more for the EU analysis than the training question does.
On retention, Amazon states that reference data for a named familiar face is stored until the customer deletes it, and that the feature automatically removes reference data for unnamed faces after 30 days [5]. Read that as a controller's admission: faceprints of unnamed, non-consenting people are generated and stored for up to 30 days. There is no de-minimis window in Article 9. A biometric template of a stranger held for 30 days is prohibited special-category processing held for 30 days.
On consent for the people being scanned, Amazon's answer to how it will obtain informed consent from passers-by, delivery workers and guests is, in substance, that it tells the customer to comply with applicable laws [5]. Amazon offers no mechanism of its own to obtain consent from the data subjects it processes despite being a joint controller (or even an independent controller where the data is used for product "improvement").
The data protection conclusion
Pulling the first argument together. Familiar Faces performs special-category biometric processing on non-consenting data subjects. The household exemption is unavailable to Amazon as a matter of law, on Lindqvist, because Amazon's processing is commercial - it cannot be a "purely personal or household activity" no matter where the camera points. (It also fails for the customer on Rynes in the ordinary case, the moment the lens reaches the pavement, and it could never have covered the "unnamed" strangers for anyone.) Amazon is a joint controller, on Fashion ID and its line of authority, because Amazon determines the purposes and the means - and the means most of all, because Amazon chose the cloud architecture that makes all of this happen on its own infrastructure. As a joint controller it owes Article 9, 5, 6, 13/14 and 25 duties directly to every person scanned, and its own correspondence shows those duties unmet: no Article 9 condition for the strangers, 30-day retention of strangers' faceprints, no controller-side consent mechanism, and an Article 25 data-protection-by-design posture that is the inverse of what Article 25 requires, given that the privacy-preserving on-device design was available in the hardware and rejected.
That is the case before we add a single gram of CO2.
The second argument: the carbon cost of a choice that did not need to be made
Now the part nobody else is costing, because the data protection story and the environmental story are being told by different people and almost never in the same room.
The feature runs biometric inference in AWS. Every face processed is a face captured on the device, shipped to a datacentre, run through a detection-and-embedding pipeline on server-class hardware with the full datacentre overhead (cooling, power conversion, networking) layered on top, matched, and answered. The on-device alternative - the one the Ambarella silicon was built for - would have run the same detection and match locally for a tiny fraction of the energy and with no transmission cost at all.
I am going to put numbers on this, using the same discipline I apply in the WebSentinel environmental methodology and in my Chrome and Anthropic pieces: a defensible mid-band, every assumption stated, and the parameter that dominates the result flagged. The accounting follows the structure of the ISO-accredited Software Carbon Intensity standard (ISO/IEC 21031:2024), which expresses software emissions as operational energy multiplied by region-specific carbon intensity plus embodied hardware emissions, on a gross basis with no offsets [19].
The assumptions
Device population. Amazon does not break out Ring's EU installed base. Ring is heavily US-skewed but the EU is a growing market. Triangulating from a plausible global installed base of mid-tens of millions of devices and an EU-27 share in the high single digits, I use a central estimate of 3 million EU-27 devices, with a defensible band of 1.5 to 6 million. This is the single biggest lever in the model; the results scale linearly, so halve or double them with the device count as you prefer.
Per-face energy. Edge case: roughly 0.05 joules per face for a lightweight on-device detect-and-embed pipeline on a CVflow-class NPU, no transmission, no datacentre overhead. Cloud case: roughly 5 joules per face for a heavier server-side detect-and-embed pipeline including host and server overhead, multiplied by a datacentre power-usage-effectiveness factor of about 1.2, plus a modest attributed upload term. These sit at the cheap, task-specific end of the published inference-energy range; benchmarking of deployed models confirms that single-purpose vision inference is orders of magnitude cheaper than the generative workloads that dominate AI-energy headlines [20]. The scandal here is not that each match is enormous, it is that Amazon multiplied a cheap operation by a needless 100x and then by millions of devices.
Grid intensity. Edge uses the EU-27 domestic grid average, about 0.25 kg CO2e per kWh - a deliberately conservative figure, as the European Environment Agency's early estimate puts EU generation intensity at roughly 220 to 240 gCO2e/kWh for 2023 to 2024 and still falling [21]. Cloud uses an AWS European-region blend of about 0.35 kg CO2e per kWh - higher than the domestic average, because Amazon's main EU regions sit on dirtier grids than the hydro-and-nuclear-heavy parts of the continent, and because the methodology here counts gross emissions and ignores market-based renewable matching, exactly as the Software Carbon Intensity standard requires [19].
Daily volume. The plausible band runs from a quiet household (a handful of faces a day) to a busy one (a doorstep on a through-route, deliveries, the school run). I model four scenarios: 1, 10, 30 and 50 faces processed per device per day.
These numbers cost the recognition step only. They deliberately exclude the embodied carbon of the camera fleet, the always-on video recording and cloud storage that happens regardless of Familiar Faces, and the model training. The marginal carbon of the feature is the inference, and the inference is what I am costing.
The result: cloud (what Amazon actually built), 3 million EU devices
| Faces per device per day | Fleet faces per year | Energy per year | CO2e per year |
|---|---|---|---|
| 1 | 1.1 billion | 1,521 kWh | 0.53 tonnes |
| 10 | 11.0 billion | 15,208 kWh | 5.3 tonnes |
| 30 | 32.9 billion | 45,625 kWh | 16.0 tonnes |
| 50 | 54.8 billion | 76,042 kWh | 26.6 tonnes |
The result: edge (what the hardware could have done), 3 million EU devices
| Faces per device per day | Fleet faces per year | Energy per year | CO2e per year |
|---|---|---|---|
| 1 | 1.1 billion | 15 kWh | 3.8 kg |
| 10 | 11.0 billion | 152 kWh | 38 kg |
| 30 | 32.9 billion | 456 kWh | 114 kg |
| 50 | 54.8 billion | 760 kWh | 190 kg |
What the gap means
At the busy end, the feature as built emits on the order of 27 tonnes of CO2e a year across a 3-million-device EU fleet. The same feature, run on the silicon already sitting in the doorbell, would emit on the order of 190 kg - about a hundredth. The difference, roughly 26 tonnes a year at the high scenario, is pure waste: emissions that buy the user nothing, that the hardware did not require, and that exist solely because Amazon decided the processing would happen in its datacentre instead of on the device the customer already paid for and powers from their own wall socket or even battery.
27 tonnes a year is not, in isolation, a climate emergency - it is roughly the annual footprint of a handful of EU residents. If the EU fleet is closer to 6 million devices, or daily volumes run higher, or you add the re-processing of unnamed faces over their 30-day retention, the number climbs, but it stays in the tens-to-low-hundreds of tonnes for the recognition step alone. The argument is not "this single feature will cook the planet". The argument is narrower and sharper, and it is the argument the Environmental Crimes Directive actually invites.
Where the Environmental Crimes Directive comes in
The EU Environmental Crimes Directive (Directive (EU) 2024/1203) reached its transposition deadline last month and member-state criminal law implementing it is now coming into force [22]. It broadens the categories of conduct that constitute a criminal environmental offence and reaches conduct that causes, or is likely to cause, substantial damage at scale. I made the underlying argument in my CIPA and ECD piece and in the Chrome analysis: where processing is unlawful, and that unlawful processing carries a measurable environmental cost at scale, you are in the territory the directive was written for [23][24].
Familiar Faces fits that template more cleanly than the Chrome case did, for one reason. With Chrome the unlawfulness was the ePrivacy and GDPR breach and the carbon was a consequence of it. Here the unlawfulness and the environmental waste are the same decision. The cloud architecture is what creates the joint-controllership exposure, the stranger-faceprint processing and the Article 25 failure - and it is, simultaneously, what creates the 100x carbon penalty. Amazon did not breach data protection law over here and waste energy over there. It did both with a single architectural choice, and that choice was avoidable on the hardware Amazon itself shipped.
So the chain is: the biometric processing of non-consenting strangers is, on the analysis above, unlawful special-category processing under Article 9 with no household exemption and an Article 25 failure baked into the design. That same unlawful processing is being performed in a datacentre, at a roughly hundred-fold carbon premium over the lawful, privacy-preserving, on-device alternative that the device's own silicon supported. The environmental harm is not incidental to the unlawful processing - it is produced by the very same design decision that makes the processing unlawful.
I will be as careful here as I was in the CIPA piece, because this is an emerging and as-yet-untested theory and overclaiming would discredit it. There is no settled case law applying the Environmental Crimes Directive to cloud compute. What there is: a directive now in force, an underlying processing activity that is unlawful on grounds that are not novel at all, a quantifiable environmental cost calculated from a method that survives review, and - uniquely in this case - a documented, available, vastly lower-carbon alternative that the manufacturer rejected. That last element is what an environmental prosecutor looks for. It is the difference between "this activity has an unavoidable footprint" and "this defendant chose the high-carbon path when the low-carbon path was sitting in their own bill of materials".
What Amazon should have done
None of this is a hard list. It is the same shape of list I gave Anthropic and Google.
- Process on the device. The CVflow silicon in the eligible cameras can do detection and a 50-template match locally. Generate the faceprint on the device, match it on the device, and never transmit or store a biometric template. This eliminates the stranger-faceprint problem and the carbon penalty in one move.
- Never store a stranger's faceprint. If the match is against a 50-person catalogue, the only template that ever needs to persist is the catalogue itself. An unmatched face should be discarded in memory, not retained for 30 days. There is no legitimate purpose served by keeping a stranger's biometric template at all, let alone for a month.
- Obtain consent as a controller, or do not process. Amazon cannot discharge its own Article 9 obligation by reminding the homeowner to comply with the law. If Amazon cannot obtain a valid Article 9 condition for the strangers it processes - and at a doorbell it cannot - then the lawful answer is to not biometrically process the strangers. On-device, discard-on-no-match design achieves exactly that.
- Stop the verbal engineering. "We do not use customers' biometric data for training purposes" is a sentence built to be technically true while leaving improvement, evaluation and non-customer data entirely uncovered. Say plainly what is collected, from whom, for what, and for how long - including the strangers.
- Account for it. Publish the aggregate compute energy and carbon of the cloud recognition pipeline, and publish the on-device alternative's footprint alongside it, so the avoidable delta is visible. Treat the delta as what it is: waste the company chose.
- Do not ship it in the EU as designed. As built, with cloud processing of non-consenting strangers' biometrics and 30-day stranger-faceprint retention, the feature should not be enabled for EU data subjects at all. Amazon has, tellingly, already declined to launch it in Illinois, Texas and Portland, which is the clearest possible signal that the company itself doubts it survives scrutiny where biometric consent law has teeth [25].
Closing
Strip away the convenience framing and Familiar Faces is a feature that biometrically scans everyone who approaches a door, performs that scan in Amazon's cloud rather than on a doorbell that could have done it locally, stores the faceprints of strangers for up to 30 days, and arrives wrapped in carefully drafted assurances that fall apart on a close read. The data protection analysis does not need the carbon argument and the carbon argument does not need the data protection analysis. But they meet at a single point, and that point is the architecture: Amazon took a feature that the device's own silicon could have run privately and cheaply (and was created explicitly for that purpose), and chose to run it publicly and wastefully instead - worse for the people scanned and worse for the atmosphere, from one decision.
The household-activity defence does not apply, because the household exemption ends where the neighbour's face begins. The processor defence does not apply, because Amazon determined the means down to the last detail. And the "it is only a small amount of energy" defence does not reach it either, because the question an environmental prosecutor asks is not "how big is the footprint" but "was the footprint avoidable", and here the answer is in Amazon's own bill of materials.
When will the regulators and public prosecutors treat biometric mass-processing of non-consenting people, performed unlawfully and at an avoidable environmental cost, as the breach it plainly is - or are the largest technology companies simply exempt from the criminal and civil statutes the rest of us are bound by?
References
[1] TechRepublic. "Ring's Facial-Recognition Feature Sparks Privacy Debate." https://www.techrepublic.com/article/news-ring-familiar-faces-ai-privacy/
[2] Fox News / CyberGuy. "Amazon adds controversial AI facial recognition to Ring", 25 December 2025. https://www.foxnews.com/tech/amazon-adds-controversial-ai-facial-recognition-ring
[3] Ring Support. "Familiar Faces." https://ring.com/support/articles/z3yhg/familiar-faces
[4] The Register. "Ring faces class action over facial-recognition feature", 3 June 2026. https://www.theregister.com/personal-tech/2026/06/03/ring-faces-class-action-over-facial-recognition-feature/
[5] Amazon (Brian Huseman, VP Public Policy). Letter to Senator Edward J. Markey responding to questions on Ring Familiar Faces, 21 November 2025. https://www.markey.senate.gov/download/amazon-markey-response-ring-frt-november-2025
[6] Security Info Watch. "Ambarella's New SoC Makes AI Processing a Reality" (identifying the Ring Doorbell Pro 2, model 5AT2S2, as operating on the Ambarella CV25M SoC). https://www.securityinfowatch.com/video-surveillance/article/21254042/ambarellas-new-soc-makes-ai-processing-a-reality
[7] Ambarella. "Ambarella Introduces CV25 SoC with CVflow Computer Vision", 7 January 2019. https://investor.ambarella.com/news-releases/news-release-details/ambarella-introduces-cv25-soc-cvflowtm-computer-vision-enable
[8] AWS Machine Learning Blog. "ML inferencing at the edge with Amazon SageMaker Edge and Ambarella CV25", 1 March 2022. https://aws.amazon.com/blogs/machine-learning/ml-inferencing-at-the-edge-with-amazon-sagemaker-edge-and-ambarella-cv25/
[9] Ambarella processor selection guidance (CVflow CV5/CV25/CV22 deliver approximately 2 to 8 TOPS of on-device inference). https://www.ambarella.com/products/consumer/
[10] European Parliament and Council. Regulation (EU) 2016/679 (GDPR), Articles 2(2)(c), 5, 6, 9, 13, 14, 25. https://eur-lex.europa.eu/eli/reg/2016/679/oj
[11] Court of Justice of the European Union. Case C-101/01 Bodil Lindqvist, judgment of 6 November 2003, ECLI:EU:C:2003:596. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62001CJ0101
[12] Reader's Digest. "Ring Doorbells Can Now Identify Faces" (Familiar Faces requires a Ring Home Premium subscription). https://www.rd.com/article/ring-doorbells-facial-identification/
[13] Court of Justice of the European Union. Case C-212/13 Frantisek Rynes v Urad pro ochranu osobnich udaju, judgment of 11 December 2014, ECLI:EU:C:2014:2428. https://curia.europa.eu/juris/liste.jsf?num=C-212/13
[14] Court of Justice of the European Union. Case C-40/17 Fashion ID GmbH & Co. KG v Verbraucherzentrale NRW eV, judgment of 29 July 2019, ECLI:EU:C:2019:629. https://curia.europa.eu/juris/liste.jsf?num=C-40/17
[15] Court of Justice of the European Union. Case C-210/16 Unabhangiges Landeszentrum fur Datenschutz Schleswig-Holstein v Wirtschaftsakademie Schleswig-Holstein GmbH, judgment of 5 June 2018, ECLI:EU:C:2018:388. https://curia.europa.eu/juris/liste.jsf?num=C-210/16
[16] Court of Justice of the European Union. Case C-25/17 Tietosuojavaltuutettu v Jehovan todistajat, judgment of 10 July 2018, ECLI:EU:C:2018:551. https://curia.europa.eu/juris/liste.jsf?num=C-25/17
[17] European Parliament and Council. Directive 2005/29/EC concerning unfair business-to-consumer commercial practices (Unfair Commercial Practices Directive), Articles 6 and 7. https://eur-lex.europa.eu/eli/dir/2005/29/oj
[18] Council Directive 93/13/EEC on unfair terms in consumer contracts (Unfair Contract Terms Directive). https://eur-lex.europa.eu/eli/dir/1993/13/oj
[19] Green Software Foundation / ISO. Software Carbon Intensity (SCI) specification, ISO/IEC 21031:2024, published 22 March 2024. https://greensoftware.foundation/standards/sci/ and https://www.iso.org/standard/86612.html
[20] Luccioni, S., Jernite, Y., Strubell, E. "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", 2024 (task-specific inference is orders of magnitude cheaper than generative inference). https://arxiv.org/abs/2311.16863
[21] European Environment Agency. "Greenhouse gas emission intensity of electricity generation" (EU early estimate approximately 242 gCO2e/kWh in 2023, down a further approximately 9% in 2024). https://www.eea.europa.eu/en/analysis/indicators/greenhouse-gas-emission-intensity-of-1
[22] European Parliament and Council. Directive (EU) 2024/1203 on the protection of the environment through criminal law (Environmental Crimes Directive). https://eur-lex.europa.eu/eli/dir/2024/1203/oj
[23] Hanff, A. "CIPA and the Environmental Crimes Directive: why forensic web evidence just became the most contested thing in privacy litigation", That Privacy Guy!, 22 May 2026. https://www.thatprivacyguy.com/blog/cipa-environmental-crimes-directive-forensic-evidence
[24] Hanff, A. "Google Chrome silently installs a 4 GB AI model on your device without consent. At a billion-device scale the climate costs are insane.", That Privacy Guy!, 4 May 2026. https://www.thatprivacyguy.com/blog/chrome-silent-nano-install
[25] Electronic Frontier Foundation. "The Legal Case Against Ring's Face Recognition Feature", 4 November 2025 (Familiar Faces not offered in Illinois, Texas and Portland). https://www.eff.org/deeplinks/2025/11/legal-case-against-rings-face-recognition-feature
