Chrome Silently Installs 4GB Gemini Nano Model
Google Chrome places a 4 GB AI model file on users' devices without permission. This occurs at a scale of billions of devices. The climate impact from such distribution reaches thousands of tonnes of CO2-equivalent emissions. Reports date back to 4 May 2026 by Alexander Hanff.
Two weeks before that report, a similar issue appeared with Anthropic's Claude Desktop. It registered a Native Messaging bridge in seven Chromium-based browsers without consent. Those browsers included Brave, Edge, Arc, Vivaldi, Opera, and Chromium. The bridge reinstalled itself on each launch of Claude Desktop. No consent dialog showed up. No opt-out option existed.
File Details and Installation Process
Chrome creates a directory named OptGuideOnDeviceModel in the user profile. Inside sits a file called weights.bin, around 4 GB in size. This holds the weights for Gemini Nano, Google's on-device large language model. Chrome powers features like "Help me write," on-device scam detection, and other AI browser aids with it.
No consent prompt appears. No setting checkbox mentions downloading a 4 GB AI model. The download starts when Chrome's AI features activate. Those features turn on by default in recent versions. Chrome checks hardware requirements and treats qualifying devices as targets.
Users report a cycle on Windows: delete the file, Chrome redownloads it. Repeat happens. Permanent removal needs disabling AI via chrome://flags or enterprise policies, unavailable to most home users. Or uninstall Chrome. On macOS, the file sets to mode 600, owned by the user. Chrome tracks install state in Local State. Next eligibility check from the variations server triggers redownload.
Verification on Fresh macOS Profile
Existing reports came mostly from Windows users noticing disk space loss. To confirm on another platform, checks used macOS. Its kernel logs filesystem events in .fseventsd. This records every file create, modify, delete at OS level. Apps cannot alter it. Logs persist after file deletion.
A new Chrome user-data directory created on 23 April 2026 for a WebSentinel privacy sweep. The sweep used Chrome DevTools Protocol: load page, wait five minutes without input, capture events, close Chrome. No human keyboard or mouse input occurred. No AI mode UI touched. Omnibox never used.
By 29 April, the profile held 4 GB in OptGuideOnDeviceModel. A du -sh command during cleanup spotted it. .fseventsd page files showed exact times:
- 24 April 2026, 16:38:54 CEST (14:38:54 UTC): Chrome makes OptGuideOnDeviceModel directory (page file 0000000003f7f339).
- 16:47:22 CEST (14:47:22 UTC): Three unpacker subprocesses create temp directories in /private/var/folders/.../com.google.Chrome.chrome_chrome_Unpacker_BeginUnzipping.*/. One (5xzqPo) writes weights.bin, manifest.json, _metadata/verified_contents.json, on_device_model_execution_config.pb. Another updates Certificate Revocation List. Third updates browser preload-data (page file 00000000040c8855).
- 16:53:22 CEST (14:53:22 UTC): Unpacked weights.bin moves to OptGuideOnDeviceModel/2025.8.8.1141/weights.bin, plus adapter_cache.bin, encoder_cache.bin, _metadata/verified_contents.json, execution config. Four more model targets (40, 49, 51, 59) register in optimization_guide_model_store. None existed before (page file 00000000040d0f9c).
Total time: 14 minutes 28 seconds. No human action during. Audit dwelled on home pages or transitioned sites. Unpacker ran in background.
Temp directory prefix com.google.Chrome.chrome_chrome_* matches Chrome's bundle ID. Chrome process itself writes the file during unrelated tab activity.
Corroboration from:
Chrome's Local State JSON: optimization_guide.on_device block with model_validation_result: { attempt_count: 1, result: 2, component_version: "2025.8.8.1141" }. Matches fseventsd path. performance_class: 6, vram_mb: "36864". Chrome profiled hardware pre-push.
ChromeFeatureState: OnDeviceModelBackgroundDownload and ShowOnDeviceAiSettings enabled. Download triggers before settings UI.
GoogleUpdater logs: on-device-model control (appid {44fc7fe2-65ce-487c-93f4-edee46eeaaab}) from http://edgedl.me.gvt1.com/edgedl/diffgen-puffin/%7B44fc7fe2-65ce-487c-93f4-edee46eeaaab%7D/.... 7 MB compressed, arrived 20 April 2026. LaunchAgent runs hourly. HTTP URL, CRX-3 verifies integrity. Control points to weights; Chrome's OnDeviceModelComponentInstaller fetches from CDN.
Four evidences align: kernel events, profile state, flags, updater logs. Model arrived without consent on zero-input profile in 14 minutes 28 seconds Tuesday afternoon.
Scale, Patterns, and Legal Issues
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
OptGuideOnDeviceModel and weights.bin reports circulated forums over a year. In 2026, scale grew with Chrome's 64%+ global share, 3.45-3.83 billion users. Gemini rollout aggressive. Affects hundreds of millions across desktop OSes.
Matches Anthropic patterns:
- Forced bundling: Chrome writes separate ML model.
- Invisible default, no opt-in.
- Harder to remove: Multi-step, undocumented.
- Pre-staging unrequested capability.
- Generic naming hides purpose.
- Installs without feature use.
- Documentation gap for users.
- Auto re-install.
- Future consent no fix for past.
- Stable channel, code-signed.
Chrome 147 shows "AI Mode" pill in omnibox. Suggests on-device use with existing model. Wrong: Cloud-backed Search Generative Experience. Queries go to servers. Local model unused by it. Features using local (Help-Me-Write, etc.) buried in menus.
User pays storage/bandwidth. No transparency gain. Engages EDPB Guidelines 03/2022 deceptive patterns: misleading info, skipping choice, hindering removal.
Breaches:
- ePrivacy Directive Article 5(3): Stores info without consent. Not necessary.
- GDPR Article 5(1): Unlawful, unfair, opaque processing (hardware profile, logs, prompts).
- GDPR Article 25: No data minimisation; pre-stages.
Same under UK GDPR, PECR. CCPA notice issue. Possible computer-misuse crimes.
CSRD notifiable environmental harm.
Environmental and Bandwidth Costs
Unlike Anthropic's 350-byte push, Chrome's 4 GB across millions. Uses 0.06 kWh/GB energy (Pärssinen et al. 2018), 0.25 kg CO2e/kWh grid factor.
Per device: 0.24 kWh, 0.06 kg CO2e. Delivery only.
Bands:
| Devices | Bytes | Energy | CO2e | |, , , , -|, , , -|, , , , |, , , | | 100M | 400 PB | 24 GWh | 6,000 t | | 500M | 2 EB | 120 GWh | 30,000 t | | 1B | 4 EB | 240 GWh | 60,000 t |
Comps: 24 GWh = 7,000 UK homes; 6,000 t = 1,300 EU cars.
Excludes: Storage (0.16 kg CO2e/GB NAND; 640,000 t for 1B), inference, re-downloads, updates. Scope 3 Cat 11.
Bandwidth burdens ISPs, metered users (month's data in places like Africa, Asia, Latin America).
Recommended Fixes
- Ask via dialog before download.
- Pull on first feature use.
- List models in settings with remove button, persistent.
- Document in installers, stores.
- Respect deletion.
- Publish ESG footprints.
- Retrospective notice + revoke.
Both Anthropic and Google cases treat user machines as vendor deployment surfaces, not personal devices whose owner decides contents.