Privacy & Data Sharing
Where to turn off model training in ChatGPT, Claude and Gemini (checked October 2026), what never to paste into a chat, and how much access to give an AI agent.
On this page
A shliach's AI chats hold the things a Chabad House is trusted with: donor amounts, family situations, a student's struggle, a board dispute. The group swings between "we already sold our privacy to Gmail" and "never connect anything." The practical middle: turn off training, keep certain things out of chats entirely, and give agents only the access they need.
Turn off model training
On consumer plans, ChatGPT, Claude and Gemini can use your chats to improve their models unless you switch it off. A group admin posted these settings after an in-person session in August 2026; they were re-checked on the vendors' help pages on 2026-10-21.
ChatGPT. Open your account menu, then Settings, then Data controls, and switch off the model-training toggle (labeled "Improve the model for everyone" in the group's August 2026 check). On the phone app: sidebar, your profile, Data controls. OpenAI says new chats with it off are kept out of training but still appear in your history. On a personal plan the same setting covers Codex.
Claude. Open your name menu, then Settings, then Privacy, and switch off "Help Improve our AI models" (Anthropic's article was updated August 3, 2026; the label used to read "Help improve Claude"). It covers Free, Pro and Max, including Claude Code on those plans. Anthropic notes that chats flagged by its safety systems may still be used for safety work. Incognito chats are never used for training, even with the setting on; they stay out of your history and memory but are kept for 30 days.
Gemini. Go to gemini.google.com, then Settings & help, then Activity, and turn off Keep Activity. Two catches:
- With Keep Activity off, Gemini does not keep your chat history either. You cannot have one without the other.
- Google still holds chats for up to 72 hours to run the service.
If you leave it on, Gemini deletes activity older than 18 months by default; you can change that to 3 or 36 months. On a work or school Google account, these settings belong to the Workspace admin, not you.
Business and nonprofit plans. The August 2026 admin post said business, team and nonprofit plans and the APIs default to no training. Anthropic's privacy center confirms it for Claude's commercial products (Team, including the nonprofit Team plan, Enterprise and the API): inputs and outputs are not used for training by default, unless you send feedback or opt in. For ChatGPT Business, check OpenAI's business data page for your plan.
Turning off training does not make a chat private. The company still stores it, staff may review flagged chats, and a hacked account exposes everything in it. A member put it this way in July 2026: storage security and whether they use your data are two separate questions.
What never goes into a chat
- Passwords, bank and card numbers, Social Security numbers, passport details. If you pasted a password, treat it as leaked and change it.
- API keys and login tokens, including in screenshots and pasted error logs.
- A full donor list with amounts, when names are not needed for the job. Replace names with ID numbers; a group admin pointed out in July 2026 that this is the standard advice for student data.
- Pastoral and counseling details with names attached.
- Anything someone told you in confidence, unless you would be comfortable telling them you shared it with an AI company.
A member made the consent point in September 2026: donors and contacts never agreed to have their details train an AI model, so keep their data on plans that do not train.
For passwords, use a password manager such as 1Password or Bitwarden, not a "secure spreadsheet" (March 2026).
Connecting Gmail, Drive and other apps
Connecting your inbox is useful and it is a real risk. Members' split from March 2026 onward:
- For: "we all sold our liberty already"; the AI finds RSVPs and donations faster than a CRM report.
- Against: a hacked AI account could read and send your email; the AI can email the wrong person; text inside an email can carry hidden instructions to the AI (prompt injection).
What the group settled on:
- Drafts only. Let the AI write drafts; you read and send. A member's ChatGPT sent a Communicator email within a minute without waiting for approval (October 2026). Say "do not send" every time.
- Assume it will use what is connected. In September 2026 an AI asked to check a sign-up sheet searched the shliach's email without being asked. Once connected, everything in that account is in play.
- Separate where you can. Forward specific emails to a separate address the AI reads, instead of connecting your whole inbox.
How to connect apps safely is on MCP Connectors.
Agents: how much access
Personal agents (GabAI, Hermes, Muse, Instinct and others) can act on your behalf, which raises the stakes. A shliach's eight questions for any agent (October 2026):
- Does it ask before sending, publishing, buying or changing records?
- Does it keep a log of what it did?
- Can you limit each connection?
- Can you fully delete your data, not just disconnect?
- Is training on your data opt-in or opt-out?
- Can family, donor, staff and client information stay separate?
- Can you export everything and leave?
- Who is responsible when it makes a mistake?
Practical rules members use:
- Give an agent only what you would give a new assistant on day one. Another member's version: only share what you would be fine uploading to a private Facebook album (October 2026).
- Give it its own email address and its own CRM login with limited rights, so you can see what it entered and shut it off. Free Gmail accounts created only for an agent have been suspended by Google (two members' reports, May and June 2026); a role account inside your own Google Workspace held up better.
- Keep credentials out of the agent's memory. Tools like OneCLI hold keys and insert them into requests so the agent never sees them.
- Read-only or draft-only for anything financial, with your confirmation before anything is written.
- Do the "verify you're human" and two-factor steps yourself.
The October 2026 debate about Instinct and Meta's Muse showed the trade-off. Press coverage from August to October 2026 raised privacy concerns about both, and an AI professional quoted in the group advised against connecting Instinct at all; several shluchos use these agents daily and say you can limit permissions and turn off training. The group did not settle it. More on Agent Safety and Personal AI Agents.
Other things that came up
- Recording meetings. Lawyers' advice relayed in 2026: an AI notetaker on a call with your lawyer may break its confidentiality. Recording rules depend on the state. See Meeting Tools.
- Workspace admins. On a managed Google Workspace account, the admin controls Gemini settings and may be able to see activity.
- Public Claude artifacts. Publishing an artifact for the ChabadOne iframe trick makes it public. Fine for a flyer page, never for personal data.
- Vibe-coded apps. The group's 2026 rules: no logins, no stored personal data, no payments and no access to your email without a developer's review. Ask the AI to check that no secret keys are exposed in what it built. See Vibe Coding.
- Looking people up. A January 2026 thread: a shliach used Gemini to dig up a stranger's contact details after the person declined to give them. The group called it creepy. Just because AI can find something does not mean you should look.
- WhatsApp agents. Agents that link to WhatsApp through unofficial libraries can get the number blocked. Use a separate number. See WhatsApp Automation.