Google Gemini routes personal Drive files and emails to human annotators
Google’s Gemini AI is reportedly sending data from connected personal services like Gmail, Drive, and Photos to low-paid human reviewers, raising alarm over how deeply user content is mined to train generative models. Commenters highlight confusing consent flows and repeated prompts to enable “connected apps,” arguing that most email and cloud files are inherently sensitive despite Google’s warnings not to link confidential data. Many see this as a breaking point for trust in large cloud providers and describe migrating to self‑hosted or privacy-focused alternatives, while noting that Google’s enterprise Workspace AI currently has different, stricter data-handling promises.
Privacy concerns and user reactions
- Many see routing Drive, Gmail, and Photos data to human annotators as a major breach of trust, especially for products viewed as “rock solid” over decades.
- Some report unsettling experiences where Gemini’s answers seem tailored to private emails, without explicit disclosure that it read them.
- Several commenters express anger and use this as a trigger to fully de-Google their lives, or at least to disable Gemini and its integrations.
How Gemini and Connected Apps are described
- Consumer Gemini with “Connected Apps” can autonomously search across Drive, Gmail, and Photos to pull file contents or snippets into its context.
- A subset of that data, plus chat history, can be sent to low-paid human reviewers for quality control and model improvement.
- Google’s own help pages explicitly warn not to connect apps if they contain information you wouldn’t want reviewers to see.
- Some note it’s easy to trigger Gemini or accept permissions accidentally due to repeated prompts.
Consumer vs Workspace Gemini
- Multiple comments stress that the controversy is about consumer Gemini.
- Workspace Gemini for business/education is said (by Google’s docs, per commenters) to be covered by stricter contracts and not used for training or human review.
- Others argue the distinction is confusing and risks eroding enterprise trust anyway.
Self-hosting and alternative services
- Many advocate local LLMs and self-hosted replacements (Nextcloud, Immich, self-hosted mail receiving) as a way to escape data mining.
- Counterpoints: self-hosting email sending and public-facing Nextcloud instances are described as fragile, complex, and security-sensitive.
- Recommended patterns include:
- Self-host receive-only mail plus third-party relays for sending.
- 3-2-1 backup strategies, offsite encrypted backups (e.g., restic/rclone/Kopia to S3-like storage).
UX, dark patterns, and regulation
- Repeated consent prompts and unclear scopes are seen as dark patterns designed to harvest more data.
- Some anticipate class actions or GDPR scrutiny, but outcomes are unknown.
- Broader sentiment: big tech will use any stored data “for any purpose they feel like,” pushing people toward open-source and local alternatives.