Google figured asking permission would hinder their ability to push this model
In the quiet hours of late April 2026, millions of computers running Google Chrome received a 4GB artificial intelligence model they never requested — a file called weights.bin, installed without notification, consent, or obvious means of removal. The episode is not merely a technical curiosity but a reflection of a recurring tension in the digital age: the gap between what powerful institutions can do and what they ask permission to do. Privacy advocates and regulators are now weighing whether this silent deployment represents a violation of foundational legal principles, or simply the latest step in a long normalization of consent-free software expansion.
- Google Chrome silently installed a 4GB AI model called Gemini Nano on millions of eligible devices in late April and early May 2026, with no notification, no permission dialog, and no clear path to removal.
- Security researcher Alexander Hanff uncovered the covert rollout and publicly accused Google of violating EU GDPR principles requiring lawful, fair, and transparent data processing — a charge with potentially significant regulatory consequences.
- The model's purpose is to run AI features locally on users' machines, offloading computational costs from Google's servers onto users' storage, battery, and processing power without their explicit agreement.
- Google revealed it had quietly built opt-out tools into Chrome settings as early as February — meaning the infrastructure for consent existed before the rollout, but was never activated as a condition of installation.
- Removing the model requires navigating deeply buried system folders on Mac or Windows, with some users also needing to manually disable flags in Chrome's developer settings — a process designed for the technically determined, not the average user.
- Regulators and privacy advocates are now watching to see whether this incident is treated as a legal violation requiring accountability, or quietly absorbed as yet another precedent for unilateral AI deployment at scale.
Somewhere on your hard drive, there may be a 4GB file you never asked for. It is called weights.bin, and it is part of Gemini Nano — an artificial intelligence model that Google Chrome began quietly installing on eligible desktop computers in late April and early May of 2026. No notification appeared. No permission was requested. You would only know it was there if you went looking.
Swedish privacy researcher Alexander Hanff discovered the silent rollout and raised the alarm. Chrome had deployed the model to devices meeting certain hardware thresholds — sufficient processing power, RAM, and storage — without informing users or seeking consent. The exact number of affected machines is unknown, but the decision itself is clear: Google installed software on people's computers without their knowledge, then made it difficult to find.
Gemini Nano is designed to run locally rather than routing requests to Google's servers. It can detect fraudulent calls, help compose messages, and summarize audio — a piece of artificial intelligence living on your machine, consuming your resources, present without your permission. Google's response acknowledged that users could disable and remove the model through Chrome settings, a capability the company said had existed since February — months before the rollout became public. The infrastructure for consent was built. It simply was not used.
For those wishing to remove it, the process is buried deep in system folders on both Mac and Windows, with some users also required to manually disable settings in Chrome's developer flags before deleting the relevant directory. The simplest solution, as many observers noted, is to uninstall Chrome entirely.
Hanff argues the practice likely violates EU GDPR requirements for lawful, fair, and transparent data processing, and that Google should have disclosed the environmental cost of distributing billions of copies of a 4GB model under EU sustainability reporting rules. The underlying logic, he suggests, is economic: by running AI on users' own hardware, Google avoids enormous server costs, shifting the computational burden — and the storage, battery drain, and processing overhead — onto the people using the browser. "Google has given us every reason not to trust them," Hanff told CNET. "So, I suspect they figured asking permission would hinder their ability to push this model."
The question regulators must now answer is whether this episode will be treated as a violation demanding accountability, or quietly absorbed as the new normal for how artificial intelligence arrives in our lives — uninvited, already installed, waiting to be discovered.
Somewhere on your hard drive right now, there may be a 4GB file you never asked for, installed by software you use every day. The file is called weights.bin. It's part of Gemini Nano, an artificial intelligence model that Google Chrome began quietly placing on eligible desktop computers in late April and early May of 2026. No notification appeared. No permission dialog. No setting that obviously let you stop it. You would only know it was there if you went looking.
Security researcher Alexander Hanff, a Swedish computer scientist and privacy advocate, discovered the silent rollout and raised the alarm. Chrome, he found, had deployed this 4GB model to devices that met certain hardware thresholds—sufficient processing power, RAM, storage, and bandwidth—without informing users or asking consent. The exact number of affected machines remains unknown. What is clear is that Google made a choice to install software on people's computers without their knowledge, and then made it difficult for those people to discover what had been done.
Gemini Nano is designed to run locally on your device rather than sending requests to Google's servers. It can detect fraudulent phone calls, help compose text messages, summarize audio recordings, and analyze screenshots from Pixel phones. It is not the same as the AI Mode feature in Chrome's address bar, which still routes queries to Google's cloud servers. This is different—a piece of artificial intelligence living on your machine, consuming storage space and system resources, present without your permission.
Google's response was measured but revealing. A company spokesperson told CNET that the model will automatically uninstall itself if a device runs low on resources. More significantly, the company claimed that in February—months before the rollout became public—it had already begun allowing users to disable and remove the model through Chrome settings. Once disabled, the spokesperson said, the model would stop downloading and updating. The implication was clear: the infrastructure for consent existed, but it was not deployed during the initial installation.
For those who want to remove Gemini Nano, the process varies by operating system and is not intuitive. On a Mac, users must navigate through Finder to Library, then Application Support, then Google, then Chrome, then Default, looking for a folder called OptGuideOnDeviceModel containing a file named weights.bin. On Windows, the path is similarly buried: AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel. Once found, users can toggle off "On-device AI" in Chrome settings, though some Windows users must also navigate to chrome://flags and disable the optimization guide, then manually delete the folder. The easiest solution, of course, is simply to uninstall Chrome entirely.
Hanff's concern goes beyond inconvenience. He argues that the practice likely violates the European Union's General Data Protection Regulation, which requires that any processing of user data be lawful, fair, and transparent. Installing software on someone's device without their knowledge fails that test. He also suggests that Google should have disclosed the environmental impact of distributing billions of copies of a 4GB model under the EU's Corporate Sustainability Reporting Directive. The company, he noted, has a twenty-year history of privacy violations at scale. "Google has given us every reason not to trust them," Hanff told CNET. "So, I suspect they figured asking permission—what the law requires—would hinder their ability to push this model."
The deeper logic is economic. Running artificial intelligence on users' own hardware means Google avoids the enormous computational costs of processing those requests on its own servers. By shifting the work to millions of devices, the company reduces its infrastructure burden and accelerates the deployment of AI features. The trade-off is borne by users: storage consumed, battery drained, processing power diverted, all without their explicit agreement. What began as a technical capability—on-device AI—became a business decision made unilaterally, and only later did the company provide the tools to opt out. The question now is whether regulators will view this as a warning or a precedent.
Citas Notables
Running inference on users' own hardware allows them to push AI features without the compute costs— Alexander Hanff, security researcher
Google has given us every reason not to trust them with a history spanning two decades of global privacy violations at massive scale— Alexander Hanff, security researcher
La Conversación del Hearth Otra perspectiva de la historia
Why would Google install something this large without asking? That seems like it would cause immediate backlash.
Because asking permission would likely mean fewer people accept it. Moving AI computation off their servers onto yours saves them billions in electricity and hardware costs. They probably calculated that the backlash from discovery would be smaller than the cost savings from deployment.
But people found out anyway. Doesn't that defeat the purpose?
Not entirely. By the time it became public, the model was already on millions of devices. And Google could point to the February settings update as evidence they'd already given users control. The narrative shifts from "we did this without permission" to "we've now made it easy to disable."
Is this actually illegal in Europe?
That's the real question. GDPR requires lawfulness, fairness, and transparency. Installing software on someone's device without notification fails on transparency at minimum. Whether it's actually illegal depends on how regulators interpret it, but Hanff's argument is solid enough that Google should have seen the risk.
What's the environmental angle?
Distributing a 4GB file to hundreds of millions of devices has a carbon cost. The EU's sustainability directive requires companies to disclose material environmental impacts. Google arguably should have reported this as a significant distribution event, but didn't.
Can people actually get rid of it?
Technically yes, but it's deliberately obscure. The settings exist now, but they're buried. Most users will never find the file or know to look for it. That's not accident—it's design.
What happens next?
Watch for EU regulators. If they fine Google or force changes, this becomes a template for how other tech companies will be held accountable for silent deployments. If nothing happens, it signals that companies can push features first and ask permission later.