Local doesn't mean private: What websites can do with the AI in Chrome and Edge
Chrome and Edge now both have an AI language model built in, and websites can talk to it directly. We compared when the multi-gigabyte downloads kick in, what a website can do without your consent, what it sees of your input and where you can switch the feature off.

What this is about
Chrome and Edge now come with a small AI language model built in, and websites can talk to it through JavaScript. The Prompt API is an interface a page uses to send text to the model and get a response back. Google opened it up to every website on May 5, 2026, with Chrome 148. The model behind it is Gemini Nano. On November 19, 2026, Microsoft made Aion Instruct, a 2.3-billion-parameter model, Edge's built-in model, replacing Phi-4-mini. Microsoft also launched an Origin Trial at the same time. That's a time-limited test that websites sign up for, after which they can use the API with real visitors.
Both companies stress that everything runs locally. Microsoft says prompts “never leave” the device, and Google says “no data is sent to Google or third parties”. That's true, but it's only part of the story. Our argument: local doesn't automatically mean private here, and the control lies with the website, not with you. We went through the documentation, the specification, position papers and independent research. We didn't run any measurements of our own.
The comparison at a glance
| Chrome (Gemini Nano) | Edge (Aion Instruct) | Firefox | Safari | |
|---|---|---|---|---|
| Websites can access it | yes, all sites since Chrome 148 | only registered sites in the Origin Trial | no | no |
| Model size on disk | approx. 4.27 GB (PCWorld) | not specified (open bf16 version approx. 4.5–5 GB) | – | – |
| Free disk space required | 22 GB | 20 GB | – | – |
| Model gets deleted when | less than 10 GB free | less than 5 GB free | – | – |
| GPU | more than 4 GB VRAM, or a CPU with 16 GB RAM and 4 cores | at least 3 GB VRAM | – | – |
| Click required | only for the first download | not covered in Microsoft's docs | – | – |
| Toggle | Settings > System > On-device AI | Settings > Copilot and AI > On-device AI | Settings > AI Controls | – |
Both browsers only download the model over unmetered connections, so it won't eat into your mobile data plan. Microsoft also lists Windows 10/11, macOS 13.3 or later, and 64-bit Linux with Vulkan drivers from Edge 159 onward. Chrome supports Windows, macOS 13 or later, Linux and Chromebook Plus devices. Android and iOS aren't supported in Chrome.
When the gigabytes land on your drive
Chrome doesn't download Gemini Nano only for websites. According to The Register, the model has been on some devices since Chrome 126. Chrome also uses it for its own features, such as scam detection. PCWorld found the 4.27 GB file weights.bin in the OptGuideOnDeviceModel folder. If you delete it by hand, Chrome downloads it again at the next opportunity. According to Tech Times, the model arrived on eligible devices without users being asked.
Edge downloads Aion, according to Microsoft, “only when a website or feature needs it for the first time, and only on devices that can run it well”. After that, all websites share the same copy. In the Canary, Dev and Beta test channels, the “On-device AI” option is off by default. Microsoft only points this out for those channels, so we conclude that the option is on by default in the regular version of Edge.
Does a website need your permission?
Neither browser has a permission prompt like the ones for your camera or location. The specification requires user activation, meaning a real click or keypress, in only one case: when creating a session would trigger a download. Without that click, the browser stops with an error. Clicking any button anywhere on the page counts. The spec doesn't require a dialog asking whether you really want to download 4 GB. It only allows browsers to add one.
If the model is already on your computer, even that click isn't needed. William Fleshman (Army Cyber Institute) and Joseph Lucas (NVIDIA) confirm this in an arXiv paper published on November 18, 2026: once the model is available, no further gestures are required. The researchers were also able to get around the click requirement for the download with simulated mouse clicks. Google marked both reports as “Won't Fix”, meaning it doesn't plan to change anything. The authors write that the browser does “not treat” access to the local model “as a security boundary”.
There is some protection. Embedded third-party content, such as ad iframes, can't use the model by default. The page that embeds it has to explicitly allow it with allow="language-model". In Chrome, the API currently isn't available in Web Workers, the background processes a page can run.
Running on your hardware: the cryptojacking question
In cryptojacking, websites secretly use their visitors' computers to mine cryptocurrency. Apple's WebKit team doesn't use that comparison directly, but it describes exactly this pattern. It says the API has “no consent mechanism for inference” and that any page could “silently run inference, consuming the user's battery, CPU and GPU”.
“The user pays dearly in battery life and device responsiveness, while the website benefits from free compute.”
Google engineer Mike Wasserman replied that browsers already protect users from sites looking for “free compute (e.g. cryptojacking)”. He called the API an attack surface that's easy to audit. We couldn't find any public documentation of the specific limits Chrome sets, such as throttling in background tabs or caps on the number of requests. The specification only says in general terms that a background tab shouldn't block the foreground tab with endless loops. In 2025, the W3C's Technical Architecture Group (TAG) also criticized that there were “no guardrails for background usage”.
The arXiv paper shows how much load this creates. On a laptop with 8 GB of VRAM, Chrome loaded about 3 GB of model data onto the GPU, and a single request took about 1.4 seconds. Without a GPU, the CPU needed about 3.7 seconds per request.
What the website sees of your input
Everything. The page writes the prompt itself and gets the response back. If the text comes from you, say from a form you want summarized, the website already has it anyway. WebKit puts it this way: the page can “log, exfiltrate or share with third parties” prompts and responses, “regardless of where inference runs”. The Register noticed that in April, Google removed the phrase “without sending your data to Google servers” from the text in Chrome's settings. Google says nothing has changed about how the data is processed. The Register suspects the phrase no longer fit once websites, including Google's own, started using the model. What happens to that data is then governed by each website's privacy policy, not Chrome's.
Edge has one extra feature. Since version 158, generated text carries an invisible watermark that marks it as AI-generated. It's left out of responses in a fixed format such as JSON.
Fingerprinting: the model gives away your hardware
In fingerprinting, websites recognize your browser from lots of small technical traits, with no cookies needed. The specification itself admits that the exact download status of the model can be one of those traits, and it suggests that browsers hide it. There's no documentation on whether Chrome or Edge actually do. In an analysis, the bot-protection company DataDome writes that a single availability check is enough to sort devices into performance tiers, because the hardware requirements are so strict. DataDome lists the time until the first word of output as another measurable trait. We couldn't access the English version of the study, so we can't fully verify the details.
Our take: the arXiv paper's measurements (1.4 vs. 3.7 seconds) show how clearly response time alone separates GPU devices from CPU-only ones. The W3C TAG and WebKit also name model availability as a fingerprinting risk.
Why Mozilla and WebKit say no
Mozilla officially lists its position on the Prompt API as “negative”, and WebKit lists it as “oppose”. Mozilla's main argument is interoperability, the idea that websites should work the same in every browser. Developers would tailor their prompts to the quirks of one model, and browsers with other models would then work worse or get locked out entirely. Mozilla's Jake Archibald warned of “severe negative consequences” for the web's neutrality. Mozilla also objects to Google attaching terms of use for developers to a web interface. There were doubts even inside Microsoft. On April 30, Edge developer Alex Russell suggested on the Chromium mailing list that the API be switched off again for Chrome 148 and sent back to the trial phase. Google shipped it anyway.
Firefox doesn't offer websites a Prompt API. The browser uses local AI only for its own features, such as translations, link previews and tab groups, and it downloads the models only when you first use them. Extensions can access them through an experimental interface, but only if you grant permission. Since February 2026, the “Block AI enhancements” toggle under “AI Controls” in the settings also blocks future AI features. According to Mozilla, it deletes any downloaded models as well. Safari has no such API. WebKit points out that on Apple devices, the operating system manages AI models, not the browser.
How to turn it off
- Chrome: Go to Settings > System and turn off “On-device AI”. According to PCWorld,
weights.bindisappeared right away. Chrome's scam warnings stop working after that.chrome://on-device-internalsshows how big the model currently is. - Edge: Go to Settings > Copilot and AI and turn off “On-device AI”. According to Microsoft, this removes the model and stops future downloads.
edge://on-device-internalsshows whether your device qualifies (performance class “Medium” or higher). - Work PCs: Admins can block the feature with the policies
GenAILocalFoundationalModelSettings(both browsers) andBuiltInAIAPIsEnabled(Chrome).
The bigger picture
The benefit is real. A summary generated in your own browser doesn't end up with OpenAI or Google, and websites don't have to ship their own model. The cost is a multi-gigabyte download that's switched on by default, plus GPU time that any page can use without asking. You also get a new fingerprinting signal. Running a model locally only tells you who doesn't get your data: the browser maker. The website still gets it.
Our take: for now, Edge is a bit more cautious than Chrome. Only registered sites can access the model, and it downloads only when it's needed. With the Origin Trial, though, Microsoft is clearly preparing for a broad rollout. The arXiv researchers suggest a visible indicator whenever a page uses the model, like the ones for camera and microphone, and that would make sense. Until it exists, the toggle in the settings is the only control you have.
Sources
- Microsoft Edge Blog: Aion Instruct und Prompt API in Edge blogs.windows.com
- Microsoft Learn: Prompt API in Edge learn.microsoft.com
- Chrome for Developers: Prompt API developer.chrome.com
- W3C-Spezifikation: Writing Assistance APIs (Aktivierung, Datenschutz) webmachinelearning.github.io
- W3C-Spezifikation: Prompt API webmachinelearning.github.io
- The Register: Google ändert Chrome-Datenschutztext theregister.com
- PCWorld: Chrome lädt still ein 4-GB-Modell pcworld.com
- arXiv: Living off the Local Model arxiv.org
- DataDome: Hardware-Fingerprinting über Chromes KI-APIs (FR) datadome.co
- GitHub: Mozilla standards-positions #1213 github.com
- GitHub: WebKit standards-positions #495 github.com
- W3C TAG Design Review #1093 tag-github-bot.w3.org
- blink-dev: Intent to Ship Prompt API groups.google.com
- blink-dev Archiv: Alex Russell und Antworten mail-archive.com
- Tech Times: Google liefert Prompt API trotz Einwänden aus techtimes.com
- CyberInsider: Mozilla warnt vor Chromes Prompt API cyberinsider.com
- Mozilla Blog: KI-Steuerung in Firefox blog.mozilla.org
- Firefox Source Docs: WebExtensions AI API firefox-source-docs.mozilla.org
- Digital Applied: On-Device-KI in Edge und Chrome digitalapplied.com
- dev.to: Your Browser Now Has an AI Model Built In dev.to
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.