Google Chrome Installs 4 GB Gemini Nano AI Model Silently – No User Consent

✍️ OpenClawRadar📅 Published: May 6, 2026🔗 Source
Google Chrome Installs 4 GB Gemini Nano AI Model Silently – No User Consent
Ad

Google Chrome is silently installing its Gemini Nano AI model — weighing in at roughly 4 GB — on user devices without asking for permission. The model, part of Google's on-device AI push, is being downloaded in the background and stored locally, which can eat up significant disk space on unsuspecting machines.

How It Works

According to reports, Chrome is downloading the model as part of its built-in AI features (like smart compose or summarization), but the installation happens automatically. Users are not prompted, and there is no clear opt-in flow. The model is stored under Chrome's local data directory, typically ~/.config/google-chrome/GeminiNano/ on Linux or the equivalent on other OSes.

  • Size: ~4 GB for the full model download.
  • No Consent: The download begins without user interaction or notification.
  • Background Process: Happens via a Chrome updater service or built-in component updater.
Ad

What You Can Do

If you want to remove the model or prevent the download, you can:

  • Disable Chrome's AI features via chrome://settings/safetyCheck or chrome://flags/#optimization-guide-on-device-model.
  • Delete the model folder manually: rm -rf ~/.config/google-chrome/GeminiNano/ (Linux/Mac) or the corresponding Windows path.
  • Use an enterprise policy: Set OptimizationGuideOnDeviceModelEnabled to false.

This is a significant privacy and storage concern, especially for users on limited data plans or constrained disk space. Developers and power users should check their systems immediately.

📖 Read the full source: HN AI Agents

Ad

👀 See Also

Claude's policy filter blocks bioinformatics work with pathogen names
News

Claude's policy filter blocks bioinformatics work with pathogen names

A computational virology researcher reports Claude's usage policy filter flags legitimate bioinformatics scripts when pathogens are named, requiring workarounds like describing tasks without organism names or downgrading to Sonnet 4. The issue affects Claude Code, claude.ai, and both Opus 4.6 and Sonnet 4.6 models.

OpenClawRadar
Critique of MCP's Abstraction Boundary and Service Integration Approach
News

Critique of MCP's Abstraction Boundary and Service Integration Approach

A Reddit discussion critiques MCP for bundling API access, efficient tooling, and domain knowledge into one layer, arguing this creates limited interfaces compared to underlying APIs. The post uses Lattice as an example where their public API only covers HR admin workflows despite having a full GraphQL API.

OpenClawRadar
SPLICE Benchmark Reveals VLMs Struggle with Temporal Reasoning, Rely on Language Priors
News

SPLICE Benchmark Reveals VLMs Struggle with Temporal Reasoning, Rely on Language Priors

Research presented at EMNLP 2025 shows vision-language models score poorly on a video sequencing task where humans excel, with models like Gemini 2.0 Flash reaching 51% accuracy versus human performance of 85%. Models frequently rely on visual shortcuts and language descriptions rather than true visual understanding.

OpenClawRadar
AI Agents Prefer Structured Queries Over Natural Language in Cala MCP Server Test
News

AI Agents Prefer Structured Queries Over Natural Language in Cala MCP Server Test

Cala's team built an MCP server with three knowledge graph access methods: natural language queries, structured query language, and direct entity/relationship traversal. Agents abandoned natural language within minutes, choosing structured queries and graph traversal without prompting.

OpenClawRadar