Which home security systems train AI on your data?

Arlo’s only training language is about video you “donate”: it may “use and view the donated video recording” for research and development, and “This includes training our artificial intelligence algorithms or as we otherwise notify you”. Donation is per recording, with your permission each time. Opt-in is what a B is. The record calls the grade inferential, since the policy never writes that undonated footage is excluded. Arlo does run AI over live footage, alerts and an optional Person Recognition feature, as stated feature processing, not training.

SimpliSafe’s privacy policy points at another document, and that document answers: video trains its models only under a sharing programme that is “strictly voluntary”, entered by opting in inside the app. That’s a B. What the addendum also says plainly is what leaving doesn’t undo: “updated or retrained AI/CV models using the Shared Video Content, cannot and will not be deleted”. Illinois, Texas, Portland and premises outside the US are excluded from the programme. The addendum provides for review of shared video by SimpliSafe employees and representatives. It says nothing about de-identification.

Anker’s policy for Eufy gives training permissions, each gated on something you volunteer for. Forum material, “public materials only”. Product-testing data you “voluntarily provide, such as audio wake-up words”, “used for model training”. And, if you turn on its improvement programme or send feedback, “designs and uploaded images to improve our AI models”. Its assistant’s text requests get the reverse promise: “However, this data will not be used for training purposes.” For the cameras it states no training use and no commitment against one. Familiar-face recognition is described as on-device.

The copy of Ring’s notice read for its verdict never mentions training, AI or machine learning. What it does say is what Ring holds: “video or audio recordings, live video or audio streams, images, comments, and data our products collect from their surrounding environment”, for the purpose “To provide you with these services”. Nothing beyond improve-our-services boilerplate is stated for any of it. Its single use of “model” is the hardware kind. The US-served render links a “Generative AI Development Disclosure”, the same template as Amazon’s: generative AI services trained on datasets that “in some cases” hold personal information, with no word on doorbell video. Ring is Amazon’s, and the site’s reading of Amazon’s own consumer notice found the same silence.

Blink’s policy is silent in the same way, over “images, audio, and video stored in connection with Blink Services”. The nearest it comes is a purpose, to “provide and continually improve the products and services of Blink and its affiliates”, and where the document explains improvement, it means analysing performance, fixing errors and usability, not building models. The word biometric never appears, from a company selling person detection. Blink’s parent is Amazon, and the silence matches Ring’s and Amazon’s own.

ADT lists “faceprints, voiceprints, iris or retina scans, gait” as biometric information, alongside data from cameras and microphones. Its policy never mentions training, AI or machine learning. The nearest language is “Enhancing, improving, or modifying our websites, products, and services”, boilerplate this site declines to grade on. It states an implied-consent posture: “If you do not notify us, we will consider you to have consented to our continued collection, use, disclosure, and sharing of personal information.” ADT’s site turns away non-browser readers. The verdict rests on the Internet Archive’s copy.