Open your file explorer. 

Navigate to the folder where Chrome stores its user data. On Windows, that is %LOCALAPPDATA%\Google\Chrome\User Data\. Look for a folder named OptGuideOnDeviceModel. Inside it, you may find a file called weights.bin. It is approximately 4 GB in size. It contains the weights for Gemini Nano, Google's on-device large language model. You almost certainly did not ask for it. Google put it there anyway.

This behaviour has been independently confirmed across Windows and macOS systems by multiple researchers and thousands of affected users [1] [2] [3] [4]. The download triggers silently through Chrome's background component updater when the browser determines your hardware meets a minimum threshold. No notification appears. No checkbox is ticked. And when users delete the file, Chrome quietly re-downloads it on the next launch.

For Indian users, this is not merely a consumer annoyance. It sits squarely in the crosshairs of India's new data protection framework, the constitutional right to informational privacy recognised by the Supreme Court in 2017, and existing civil liability provisions under the Information Technology Act, 2000.

Check your device right now On Windows: press Win + R, type %LOCALAPPDATA%\Google\Chrome\User Data\ and press Enter. Look for OptGuideOnDeviceModel\ and a file named weights.bin inside it. If it exists and is around 4 GB, Chrome placed it there without your consent.

What Is Actually on Your Disk and How Did It Get There?

The file is the weights for Gemini Nano, Google's smallest on-device large language model. Chrome uses it to power several AI-assisted browser features: "Help me write" in text boxes, on-device scam detection, tab-group suggestions, and smart paste. None of these features required your invocation. The model was pre-staged on your device so the features could run instantly if you ever used them.

Privacy researcher Alexander Hanff documented a forensic reconstruction of exactly how the download happens on macOS [5]. Using macOS's kernel-level filesystem event log, which Chrome cannot tamper with, he traced the following sequence on a profile that received zero human input:

  • Chrome creates the folder The OptGuideOnDeviceModel directory appears in the user's Chrome profile. No tab has been opened. No AI feature has been touched.
  • The unpacker fires Three background subprocesses spawn under Chrome's own bundle identifier. One writes weights.bin, a manifest, and a verified-contents file. Chrome batches this into the same idle window as a routine security certificate update, treating a 4 GB AI model download as equivalent to a 50 KB patch.
  • The model moves into place The unpacked file is installed at its final path. Smaller companion models for text safety and prompt routing also register. The entire process takes under 15 minutes. Human actions during this window: none.

Chrome's own internal configuration files confirm that the device was profiled, its GPU class, VRAM and unified memory were read, before the decision to push the model was made. The settings UI that would theoretically allow you to discover and disable this feature was enabled in lockstep with the download itself. That means the settings page through which you could refuse the model appears at the same time as the model is already being installed. This is design, not oversight.

The "AI Mode" Deception: Your Search Queries Still Go to Google's Servers

Here is the part that makes this worse. Recent versions of Chrome display an "AI Mode" button in the address bar, the most prominent piece of real estate in your browser. A reasonable user who knows their device now hosts a 4 GB Gemini Nano model will draw an obvious inference: AI Mode uses the local model, so my queries stay on my device.

That inference is completely wrong. The AI Mode address-bar button is a cloud-backed Search Generative Experience surface. Every query you type into it is transmitted to Google's servers for processing. The on-device Nano model is not involved in that flow at all.

The user pays the storage cost of the 4 GB file. The user's most visible AI experience still sends every query to Google's data centres regardless. The local model is a Google-side asset deployed on your device  it is not a user-side privacy benefit.

The actual features that use the on-device model, "Help me write" in textarea boxes, tab-group suggestions, are buried in right-click context menus that most users will never discover. The headline AI affordance you can see does not use the file sitting on your disk. The file that is sitting on your disk does not power the AI Mode you can see. They are entirely separate. Yet both arrived together, without your knowledge.

What Indian Law Says About This

The Constitutional Foundation: Right to Informational Privacy

In August 2017, a nine-judge bench of the Supreme Court of India delivered a unanimous verdict in Justice K.S. Puttaswamy (Retd.) v. Union of India, holding that the right to privacy is a fundamental right protected under Article 21 of the Constitution [6]. The judgment explicitly recognised informational privacy, the right of individuals to control access to information about themselves, as a core component of this right [7].

Justice D.Y. Chandrachud's lead opinion in Puttaswamy took direct note of the digital trails individuals leave behind, including the use of tracking technologies, the collection of browsing histories, and the profiling of users through algorithmic analysis. Chrome's silent hardware profiling of your device, its GPU class and VRAM, to determine whether you are an eligible recipient for the model push, is precisely the kind of non-consensual data processing that Puttaswamy places within the ambit of fundamental rights protection.

The Digital Personal Data Protection Act, 2023

India's Digital Personal Data Protection Act, 2023 (DPDP Act), enacted on August 2023, establishes India's comprehensive data protection framework. The DPDP Rules, 2025, notified on November 2025, operationalise the Act with full compliance expected by 13 May 2027 [8].

Under Section 6 of the DPDP Act, consent for processing personal data must be "free, specific, informed, unconditional and unambiguous with a clear affirmative action" [9]. The Act has extraterritorial reach: it applies to foreign entities that offer goods or services to data principals within India. Google Chrome is very much such an entity.

The Information Technology Act, 2000: Section 43

Section 43 of the Information Technology Act, 2000 creates civil liability for any person who, without permission of the owner of a computer or computer system, accesses or secures access to it, copies or downloads any data from it, or introduces any programme that damages or is likely to damage any computer system [10].

Chrome writing a 4 GB binary to the user's disk without permission is, in plain language, accessing and modifying a computer resource without the owner's consent. Section 43 does not require proof of fraudulent intent for civil liability to attach. Unauthorized access itself is sufficient. An Adjudicating Officer under the IT Act can direct compensation of up to Rs 5 crore. Where damages exceed that threshold, the remedy lies in civil court.

Section 66, IT Act, 2000 Section 66 elevates contraventions under Section 43 to criminal offences where they are committed with dishonest or fraudulent intent. Google's decision to deliberately obfuscate the file under the name OptGuideOnDeviceModel rather than a transparent name like GeminiNanoLLM, and to batch it with a routine security certificate update, could engage questions of intent that the Section 66 threshold requires.

Consumer Protection Act, 2019

The Consumer Protection Act, 2019 defines unfair trade practices to include practices that falsely represent goods or services, or mislead consumers about the nature of a product. Chrome's "AI Mode" label, displayed prominently in the address bar, creates a reasonable impression that local AI processing is occurring. When in fact every query is transmitted to Google's servers, the label is arguably misleading within the Act's definition. Indian consumers may lodge complaints with the relevant District or State Consumer Disputes Redressal Commission.

The Environmental Cost at India's Scale

India accounts for over 350 million Chrome users. At 4 GB of data per device push, the network-transfer cost of the Gemini Nano rollout to eligible Indian devices runs into hundreds of petabytes of data.

For Indian users on metered mobile broadband connections, which remain the primary form of internet access for hundreds of millions of people in Tier 2, Tier 3, and rural India, a 4 GB silent download is catastrophic. On a 1.5 GB/day prepaid plan, it erases nearly three days of mobile data without the user's knowledge or agreement. Google has made no public commitment to exclude metered connections from the model push, and the behaviour has been documented on mobile-broadband-connected Windows machines as well as those on fixed-line connections.

Scenario Devices Data consumed CO2 equivalent*
Conservative (10% of Indian Chrome users) 35 million 140 petabytes ~2,100 tonnes
Mid estimate (25%) 87 million 350 petabytes ~5,250 tonnes
High estimate (40%) 140 million 560 petabytes ~8,400 tonnes

 

CO2e estimates based on 0.06 kWh/GB network transfer intensity and 0.25 kg CO2e/kWh grid factor (delivery-only; excludes re-downloads and on-device inference).

What Google Should Do  and What Indian Users Can Do Now

What Google must do

  1. Ask first, always. Before Chrome downloads any AI model file, surface a clear, plain-language dialogue in the user's preferred language explaining what the file is, how large it is, and what features it will enable. Provide a genuine decline option.
  2. Pull, never push. Download the model only when a user invokes an AI feature for the first time. Feature use is the consent event. Pre-staging is not.
  3. Surface and honour deletion. List all downloaded AI model files in chrome://settings/ with their size, purpose, and a remove button. If a user deletes the file, treat that as a standing instruction, not a temporary state to correct.
  4. Publish clear notice in India. Chrome's Google Play Store listing, the Chrome download page, and any in-app onboarding should state plainly, in English and in major Indian languages, that Chrome downloads large AI model files on supported hardware. This is a basic DPDP Act notice obligation.
  5. Issue retrospective notice to Indian users. Every Indian Chrome user who received the model without notice should be informed, on next Chrome launch, of what was installed and given a one-click removal option. The DPDP Rules 2025 require retrospective notices for data processed before the Rules came into force.

What Indian users can do right now

  1. Check and delete the file. Navigate to %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel\ on Windows. Delete the folder. This frees the storage temporarily.
  2. Disable the Chrome flags. Open Chrome and go to chrome://flags/. Search for "optimization guide on device" and set it to Disabled. Also disable "Prompt API for Gemini Nano." Relaunch Chrome. This prevents re-download [1].
  3. Apply a permanent registry block (Windows). The most durable fix is an enterprise policy entry in the Windows Registry. Pureinfotech has a detailed, verified guide for this here. This prevents Chrome from re-downloading the model even after updates.
  4. Consider alternative browsers. Firefox, Brave, and other non-Chromium browsers do not implement Gemini Nano. If Chrome's conduct concerns you, switching is the most complete remedy available today.

The Larger Pattern: Your Device Is Not a Deployment Target

This is not an isolated incident. It is part of a broader industry pattern in which technology companies treat users' personal devices as infrastructure for their own product roadmaps. Chrome holds over 64% of global browser market share and reaches somewhere between 3.45 and 3.83 billion users worldwide. That scale turns a unilateral technical decision about a default setting into a policy that affects more people than most nations have citizens.

India's DPDP Act was designed precisely for this moment. It was built on the Puttaswamy principle that informational privacy is a fundamental right, not a privilege that technology companies can waive on your behalf through a default setting buried in a terms-of-service document you never read. The Act's requirement for free, specific, informed, unconditional and unambiguous consent is a direct response to the industry habit of treating consent as a formality to be designed around.

Google has not, to our knowledge, made any public statement specifically addressing Indian users' rights under the DPDP Act in connection with the Gemini Nano push. The compliance deadline of 13 May 2027 is approaching. Whether Google will treat that deadline as an opportunity to do right by Indian users, or as a deadline to manage through minimal technical compliance, will be instructive.

India's Supreme Court said in 2017 that informational privacy is a fundamental right. India's Parliament said in 2023 that consent must be free, specific, informed, and unambiguous. The question now is whether India's regulators will say the same thing to Google.

Is the 4 GB file on your device?

Follow the step-by-step guide to check, remove, and permanently block the Gemini Nano model from being re-downloaded to your Windows PC.

How to Remove It Permanently →

References & Sources

  1. Pureinfotech. "Stop Chrome from silently downloading Gemini Nano AI model on Windows 11." March 2026. [Verified working]
  2. Dhavale, V. "Chrome Installed a 4GB LLM on My Machine. Here's What I Found Out." April 2026. [Verified working]
  3. WinAero. "Google Chrome Secretly Downloads Huge Local AI Models." March 2026. [Verified working]
  4. AIBase. "Google Chrome Exposed for Forcing 4GB AI Model Installation." March 2026. [Verified working]
  5. Hanff, A. (That Privacy Guy). "Google Chrome silently installs a 4 GB AI model on your device without consent." May 2026. [Original forensic research]
  6. Wikipedia. Justice K.S. Puttaswamy v. Union of India (2017) — Right to Privacy verdict.
  7. Dvara Research. "The Right to Privacy Judgment: Initial Reflections on Implications for Digital Financial Services." 2017.
  8. matters.ai. "DPDP Act 2023: India's Digital Personal Data Protection Law Explained." 2026.
  9. Hogan Lovells. "India's Digital Personal Data Protection Act 2023 brought into force."
  10. ApniLaw. "Section 43 of the IT Act Explained: Hacking and Unauthorized Access to Computer Systems."
  11. LLMAdvocates. Section 43, Information Technology Act, 2000 — full text.
  12. MeitY. Digital Personal Data Protection Act, 2023 — official text (PDF).
  13. StatCounter. "Browser Market Share Worldwide."
  14. DemandSage. "How Many People Use Google Chrome (Updated 2026 Data)."
  15. DLA Piper Data Protection. "Data protection laws in India."