ChatGPT App Permissions and Privacy
What each app actually sees, how to review scopes before you connect, when to narrow access, and the two-step revocation most people miss. Practical, no-fluff — the privacy posture is one of the seven criteria we score in our ranking methodology, and this page is the longer-form explanation.
The short version
ChatGPT apps are not magically more (or less) risky than any other OAuth-based third-party integration. The risk profile is entirely determined by what you authorize and who is downstream of that authorization (OpenAI plus the source service). If you would not authorize a random Chrome extension to write to your Drive, you should not authorize a ChatGPT app to write to your Drive without thinking about it first.
The good news: the OAuth scopes are explicit and reviewable. The connection screen lists exactly what the app is asking for. That's your moment to read, narrow, and decide — not after the fact when you're trying to figure out why an app modified something you didn't expect.
The five permission tiers (and what to do about each)
Most ChatGPT apps fall into one of these tiers. The tier determines how careful you should be at the connection screen.
Read-only on a narrow scope
App can read a specific set of files, pages, or records you explicitly select. Cannot modify anything.
Notion connector scoped to a single database. A Drive connector scoped to a single folder. A read-only Stripe scope.
Lowest-risk option. Use this tier first when you're testing what an app can do. Most queries don't actually need write access.
Read-only on a broad scope
App can read everything in the connected account but cannot modify anything.
Full Google Drive read scope. Slack workspace read scope. HubSpot CRM read scope.
Acceptable for accounts that don't hold sensitive data. For work / regulated accounts, prefer narrower scoping if the app supports it.
Read + write on a narrow scope
App can read and modify a specific subset of resources.
Notion access limited to one project workspace, with page-create permission.
Use when the workflow genuinely requires write access (e.g., creating new pages). The narrow scope contains blast radius.
Read + write on a broad scope
App can read and modify anything in the connected account.
Full Google Drive read/write. Full HubSpot CRM write. Full Stripe API access including refunds and payment links.
Highest risk. Grant only when you trust both OpenAI's connector data policy and the third-party service, and only for accounts where the worst-case write action is recoverable.
Administrative or financial
App can perform actions with real-world or money consequences — issuing refunds, sending agreements, ordering goods, controlling automations.
Stripe refund scopes. DocuSign envelope-send scopes. Zapier scopes that fire write Zaps. Instacart checkout.
Confirm every action before approving. Consider keeping these in dashboard-only flows rather than chat-initiated unless you have a strong reason.
Before you connect: the 60-second checklist
Run this checklist every time you authorize a new app. It doesn't make you safer than thinking carefully; it just turns careful thinking into a habit.
- Which account am I connecting? Personal or work. If work, does your IT policy allow this?
- What's the smallest scope that works? If the app supports folder / page / workspace-level scoping, use it. Don't grant the whole account when a folder is enough.
- Read-only or write? Start read-only. Upgrade to write only when a specific workflow genuinely requires it.
- What's the worst-case action? If the worst case is "model summarized something I'd rather keep private," that's lower-stakes than "model deleted my CRM."
- Will I remember to revoke? For one-off projects, write the revocation step into your calendar at the same time you grant access.
What ChatGPT does (and doesn't) do with connector data
When you ask a question that touches a connected app, the relevant content is fetched from the source service and used to answer your question. The standard expectations apply:
- Per-request fetching: Content is fetched when needed, not bulk-mirrored. There is no permanent ChatGPT-side copy of your Drive.
- Training opt-out depends on your plan: Enterprise and Team plans typically exclude content from training by default. Consumer plans depend on account settings. Check the current OpenAI data policy for specifics.
- Source-service terms still apply: Notion, Google, Microsoft, Stripe — each has its own data policy that governs the data on their side. The ChatGPT connection doesn't replace those policies; it stacks on top.
- Logs and audit trails: Most source services log connector access. If your org has auditing requirements, the audit trail lives on the source side (Google Workspace logs, Microsoft 365 audit logs, etc.).
How to revoke (the two-step that most people miss)
Disconnecting an app inside ChatGPT removes the app from your chat surface, but the OAuth grant in the source service often persists until you revoke it there too. For thorough cleanup, do both:
- In ChatGPT: Settings → Apps and connectors → find the app → Remove / Disconnect.
- In the source service: Account / Security settings → Connected apps (or Third-party access, or OAuth apps) → find the ChatGPT entry → Revoke.
- Google: myaccount.google.com → Security → Third-party connections
- Notion: Settings → Connections
- Microsoft: account.microsoft.com → Security → Connected apps
- Stripe: Dashboard → Settings → Authorized applications
- Slack: Workspace settings → Manage apps
- Salesforce: Setup → OAuth Connected Apps
If you do step 1 but skip step 2, the OAuth grant technically still exists. Reconnecting the app from ChatGPT may not even require re-authorization — it picks up the lingering grant. That's why the source-service revocation matters.
How we score privacy in our rankings
"Privacy clarity" is one of the seven criteria in our editorial methodology. We don't just check whether a privacy policy exists — we look for:
- Whether the app supports narrow scoping (folder / page / workspace level rather than entire account).
- Whether the OAuth scopes shown at connection time are clear and human-readable, or buried in legalese.
- Whether the source service publishes clear data-handling terms for connector access (separate from general product privacy policy).
- Whether write actions are clearly distinguished from read actions, and whether the app surfaces confirmation prompts before consequential write operations.
- Whether revocation in both ChatGPT and the source service is straightforward.
Apps that score high on these dimensions earn a higher privacyClarity score, which in turn lifts their overall editorial score. We do not rank apps higher just because they're more permissive or more popular — privacy clarity is a real tiebreaker.
Frequently asked questions
Frequently asked questions
- Are ChatGPT apps safe?
- Safety depends on three things: the scope you authorize, the third-party service's data policy, and OpenAI's connector data policy. A ChatGPT app is no riskier than directly granting an OAuth app access to the same account on the same scopes — but it is also no less risky. Review the scopes before authorizing.
- What data does a ChatGPT app see?
- Whatever the OAuth scope you grant says it can see. A read-only Drive scope means the app can read your files. A write scope adds the ability to create or modify them. There is no special hidden access; the scopes are explicit and reviewable at the connection screen.
- Is my chat content used to train models?
- OpenAI's data policy governs prompt content for the ChatGPT side. Connector-fetched content (your Google Drive file, your Stripe report) is governed by both OpenAI's connector data policy and the source service's terms. Enterprise plans typically opt out of training; consumer plans depend on your account settings.
- How do I revoke an app I no longer use?
- Two-step revocation: in ChatGPT (Settings → Apps and connectors → remove), and in the source service (Account → Connected apps → revoke). The second step is the one most people skip — until you remove the OAuth grant in the source service, the connection technically still exists.
- Can I limit which files or pages an app sees?
- Yes, in most cases. Apps that use folder/page-level scoping (Notion, parts of Google Drive in some flows) let you grant access to a specific subset rather than the whole account. Use that when available — it's the single most effective privacy control.
- What about regulated data (HIPAA, financial, customer PII)?
- For regulated content, defer to your organization's data classification and security policy. Many ChatGPT enterprise and business plans have specific data-handling commitments; consumer plans typically do not. Don't connect regulated data sources on a personal account.