Google quietly revises Chrome wording around Prompt API and Gemini Nano privacy

Chrome’s built-in AI is the story here — specifically the Gemini Nano model that powers the Prompt API and other local features inside Chrome. The stakes are simple: if a browser downloads a multi-gigabyte model onto your machine and lets websites call it, people want very clear rules about what stays local, what gets downloaded, and when that happens. (winbuzzer.com) The gap is that Chrome’s behavior and Chrome’s wording have not always been equally clear. This week, reports showed Google had softened some of its privacy wording even while keeping the basic “runs on device” setup in place. (developer.chrome.com) ### What is the Prompt API? (adsm.dev) The Prompt API is Chrome’s browser feature that lets a website send natural-language requests to Gemini Nano running inside the browser, not in a remote cloud service. Google’s docs pitch it for things like summarizing pages, classifying content, extracting details, and building writing tools. It is part of Chrome’s broader “built-in AI” stack. ### Where does Gemini Nano actually run? For these built-in APIs, the model runs on the user’s device. Chrome’s developer docs still say “No data is sent to Google or any third party when using the model,” and they point developers to local inspection tools like `chrome://on-device-internals` to see what is installed. (winbuzzer.com) That is the core technical claim Google is still making. ### So what changed? Turns out the dispute is mostly about wording, not a newly discovered server-side data flow. Reports published on May 11 said Google removed or toned down some of Chrome’s clearest end-user privacy language around on-device AI. The practical complaint is that the old wording did real work — it told users why an automatic local download was supposed to be privacy-preserving. (developer.chrome.com) When that wording gets weaker, the product feels less explicit even if the architecture stays the same. ### When does Chrome download the model? Chrome’s own model-management docs say Gemini Nano is downloaded on demand when a built-in AI API needs it, typically on the first call to a `*.create` function. (developer.chrome.com) The browser may keep downloading in the background even if the triggering tab closes, and it checks for model updates when Chrome starts. Basically, the model is treated like browser infrastructure, not like a one-off file the user manually manages. ### How big is this thing? That is part of why people noticed. Independent testing and developer writeups describe a GPU build around 4 GB and a CPU build around 2.7 GB, while Google’s docs say the exact size varies and users need at least 22 GB of free space on the profile volume. (winbuzzer.com) That is not a tiny helper file — it is a meaningful storage and bandwidth event. ### Why is this flaring up now? Because Chrome 148 made the Prompt API available by default on desktop, instead of keeping it tucked behind earlier preview-style access. Once a capability moves into the default browser experience, quiet background downloads and fuzzy consent language stop feeling niche. (developer.chrome.com) They become product-policy questions. ### Is the real issue privacy or consent? Both — but they are different problems. Privacy is the narrow technical question of whether prompt data leaves the device during local model use. Consent is the broader product question of whether users clearly agreed to a large download and understood what it enables. Google has defended the on-device design, and Chrome leadership has said users can disable the model so it will not re-download. (adsm.dev) But that still leaves a wording problem: local processing is not the same thing as informed consent. ### What should developers take from this? (adsm.dev) If you build on Chrome’s AI APIs, do not lean on vague “on-device” shorthand. Spell out what triggers downloads, what storage is required, what data stays local, and which features actually use the local model versus Google’s cloud AI. The awkward part is that Chrome now has both kinds of AI paths in the same product. ### Bottom line? Google’s apparent wording change did not, on the evidence available, flip Chrome’s Prompt API from local to cloud. But it did remove some of the plain-English certainty around that promise. (digitaltrends.com) In a browser that now ships more AI by default, that kind of edit matters almost as much as the code. (winbuzzer.com)