Privacy & Data Sharing
How the group thinks about AI data exposure, and the patterns that minimize it.
On this page
- The recurring warnings
- The pragmatic position
- Specific concerns the group has surfaced
- Connecting AI to your full Gmail or Drive
- Vibe-coded apps with banking info
- Password storage
- Workspace admin visibility
- Public Claude artifacts
- Ethnicity-based marketing lists
- Patterns that minimize exposure
- Open questions
- Related
Where the group oscillates between paranoia and pragmatism, and lands somewhere in the middle.
The recurring warnings
"They use your chat data to train their models."
"If you have Gemini for Workspace, users cannot easily delete their chats, and the admin can see all the chats."
"I've found Gemini for nonprofit doesn't integrate into email the same way it does for a regular account."
The pragmatic position
"We here have already sold our liberty to technology."
The honest framing: most shluchim already use Gmail, WhatsApp, ChatGPT, ChabadOne, and social media. Adding Gemini integration does not change the threat model materially; it just makes it visible.
Specific concerns the group has surfaced
Connecting AI to your full Gmail or Drive
One member asked whether there was any reason for concern. No consensus emerged; positions ranged from "go ahead" to "compartmentalize."
Vibe-coded apps with banking info
"Keep programs local, don't connect to banking, don't give file access. Use professional developer oversight before deploying."
Password storage
"Use a dedicated password manager (Bitwarden, 1Password), not a spreadsheet. Zero-knowledge architecture, AES-256 standard."
Workspace admin visibility
On managed Workspace accounts, the admin can see Gemini chat history and users cannot easily delete it. This affects shluchim on org-wide accounts.
Public Claude artifacts
The iframe-into-ChabadOne workaround publishes the Claude artifact to a public URL. Fine for marketing pages, not for anything with PII.
Ethnicity-based marketing lists
These carry a real legal limit (Data Axle and similar), not just an AI guardrail. The restriction applies regardless of which model you ask.
Patterns that minimize exposure
- In-browser, no server vibe-coded apps, where nothing persists. Movers-referral-tool, my-charity-box, and countomer all do this right.
- MCP into vetted backends instead of browser automation against payment systems.
- Forward sanitized data; do not connect your full inbox.
- Anonymize or aggregate before feeding to AI when you can.
- Sefer-style buffer layers: a curated corpus the AI can read, and vetted summaries it can re-present.
Open questions
- Free-tier training opt-out: does it work? It varies by vendor and changes silently.
- Workspace Nonprofit Gemini: a different data-use policy than consumer? Unclear.
- Encrypted WhatsApp backups block AI mining of personal chat data; is that good or bad?