The Verge reports that Google is set to save media from your Search-service interactions, the photos you run through Lens, the audio and video from Search Live, and the audio you speak into Translate, and to use it to train its AI. The mechanism, confirmed in Google's own support documentation, is a setting called Save Media that lives under Search Services History. When it is on, media from those interactions is retained to your history, and Google says that media "may be used to develop and improve Google's AI models and technologies." When the data is used for training, Google states, it is disconnected from your Google Account.
There is an off switch, and the precise shape of it matters. According to Google's documentation, when Search Services History or the Save Media subsetting is turned off, media from future interactions will not be used to train Google's generative AI models, unless you submit feedback. So this is a setting-gated data flow rather than a blanket grab, and the honest version of the story is that the contested part, what the default state is and whether existing users find themselves newly included, is exactly what makes it news. The alarmed framing is The Verge's headline; the mechanism is Google's, and the responsible reading is that this is a control you should check, not a default you can assume.
What makes it worth covering is the timing, because the same week produced the opposite philosophy in full. Apple's Private Cloud Compute, the cloud tier behind Apple Intelligence, staked its entire pitch on data that is never stored, never used for training, and verifiable by cryptographic attestation you can inspect yourself. Google's approach here is the mainstream one and the older one: your camera scans and spoken translations are genuinely useful training data, retained behind a toggle, with account-disconnection as the privacy mitigation rather than non-retention. Two answers to the same question landed in one news cycle, do not keep it versus keep it but de-identify it.
For users the action is small and concrete: Search Services History and the Save Media subsetting are where you turn this off. For everyone watching the shape of "private AI," the signal is larger. The multimodal surfaces that Google has spent two years pushing, the camera as a search box, live spoken translation, Search Live, are exactly the surfaces that generate the richest training data, and they are now intake by default of the product rather than by separate consent. The phrase doing the load-bearing work is "disconnected from your Google Account," which is a weaker guarantee than never-retained and a stronger one than nothing. The week's two privacy stories bracket the range of what private AI is going to mean, and the difference between them is whether the guarantee is enforced by math or by a policy you have to take on trust.