Here’s an uncomfortable question: when you paste a client’s contract into ChatGPT for a quick summary, where does that data go? Who can see it? Could it show up in someone else’s response?
I read the privacy policies and terms of service for ChatGPT, Claude, and Gemini so you don’t have to. All 47,000 words of them. Here’s what I found — including some things that genuinely surprised me.
Why This Matters Now
Three things have changed in 2026 that make AI privacy more relevant than ever:
1. Enterprise adoption is real. Companies are integrating AI into workflows that handle sensitive data — legal documents, medical records, financial reports, HR files. The “just don’t paste anything sensitive” advice stopped being practical a year ago.
2. Data breach incidents are happening. In May 2026, Anthropic confirmed a data exposure affecting Claude Free tier conversations from a misconfigured logging pipeline. OpenAI had a similar incident with ChatGPT in early 2025. These are rare, but they demonstrate that nothing is zero-risk.
3. Regulation is catching up. The EU AI Act’s data governance provisions kicked in fully in 2026. California’s updated privacy bill specifically addresses chatbot data. Companies can no longer just say “trust us.”
ChatGPT (OpenAI)
What they collect: Everything you type into ChatGPT. Plus technical data (IP address, device info, usage patterns).
Training on your data: This is the big one. OpenAI’s default policy is that ChatGPT Free and Plus conversations may be used to improve their models, unless you opt out.
How to opt out: Settings → Data Controls → “Improve the model for everyone” → toggle OFF. This is a one-click toggle in the ChatGPT interface. Once you toggle it off, conversations from that point forward are excluded from training. But — and this matters — conversations from before you toggled it off remain in the training dataset forever.
Enterprise/Team plans: ChatGPT Team and Enterprise plans are different. By default, OpenAI does NOT train on Team/Enterprise data. This is a contractual guarantee, not just a setting.
Data retention: OpenAI retains conversation data for 30 days even after deletion (for abuse monitoring). After that, it’s permanently deleted. API data is retained for 30 days by default, configurable to 0 days for Enterprise.
Third-party sharing: OpenAI shares data with “service providers” for hosting and analytics. They do not sell your data to third parties. The policy explicitly states this.
What surprised me: The “opt-out doesn’t apply to past conversations” detail. If you’ve been using ChatGPT Free for a year without opting out, all those conversations are permanently in the training corpus. Opting out now protects future conversations, not past ones.
Claude (Anthropic)
What they collect: Conversations, uploaded files, and technical metadata. Anthropic collects less metadata than OpenAI — they don’t track detailed usage patterns for personalization.
Training on your data: Anthropic’s policy is the strongest of the three. By default, Claude does NOT train on user conversations from any paid plan (Pro, Max, Team, Enterprise). For the Free tier, they may use conversations for safety research and model improvement, but they explicitly say they filter out personal information first.
Opting out: For Free tier users, you can request data deletion through Anthropic’s privacy portal. For paid users, you’re already opted out by default.
Data retention: Anthropic retains conversation data for 90 days for Pro users (for abuse monitoring), then deletes it. API data retention is configurable (default 30 days, can be set to 0 for Enterprise).
Third-party sharing: Anthropic states they don’t sell data and only share with service providers (AWS for hosting, etc.). Their constitutional AI approach includes a privacy principle baked into the model.
What surprised me: The Free tier still has training implications, even though Anthropic’s policy is more protective than OpenAI’s. If you’re using Claude Free for anything sensitive, upgrade to Pro. The $20/month is worth it for the data protection alone.
Gemini (Google)
What they collect: This is where it gets complicated. Google collects your Gemini conversations AND can connect them to your broader Google account activity. If you have Web & App Activity turned on, your Gemini conversations are linked to your Google account and used across Google services.
Training on your data: Google’s policy is the most opaque. They state that “human reviewers may read, annotate, and process your Gemini Apps conversations” to improve quality. They don’t explicitly say they use Gemini conversations for model training, but their broader privacy policy allows them to use Google service data to improve Google services.
The Google Workspace version of Gemini has stronger protections — by default, Workspace Gemini conversations are not used for training and not reviewed by humans.
How to opt out: Gemini App Activity → turn OFF. This prevents conversations from being saved to your Google account and reviewed by humans. But it also disables some features (Gemini won’t remember context from previous conversations).
Data retention: If Gemini App Activity is ON, conversations are retained until you delete them. If it’s OFF, conversations are retained for 72 hours for service delivery, then deleted. But — Google’s privacy policy allows them to retain “anonymized” data indefinitely.
Third-party sharing: Google’s privacy policy allows data sharing across Google services. This means your Gemini conversation data could theoretically be used to improve Google Search, Google Ads, or other Google products. The policy separates Gemini data from ad personalization (they claim they don’t use Gemini conversations for ads), but the Google ecosystem is so interconnected that the boundaries are blurry.
What surprised me: The Google ecosystem integration is a double-edged sword. It’s convenient — Gemini can reference your Gmail, Calendar, and Drive. But it also means your data lives in Google’s broader data ecosystem, which is designed for data monetization even if Gemini specifically isn’t used for ads.
Side-by-Side Comparison
| Privacy Factor | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Trains on Free tier data | Yes (opt-out available) | Limited (filtered, safety only) | Unclear / complex |
| Trains on paid tier data | No (with opt-out) | No (default) | No (Workspace only) |
| Opt-out available | Yes (toggle) | Paid = auto opt-out | Yes (toggle, loses features) |
| Past data protected by opt-out | No | N/A (limited training) | N/A (72h retention if off) |
| Minimum data retention | 30 days | 90 days (Pro) | 72 hours (activity off) |
| Cross-service data sharing | No | No | Yes (Google ecosystem) |
| Enterprise protections | Yes (contractual) | Yes (contractual) | Yes (Workspace) |
| Privacy policy readability | Medium | Good | Poor (fragmented) |
What About API Usage?
If you’re building apps that use these models through APIs, the privacy picture is much clearer:
OpenAI API: No training on API data by default. Data retained 30 days for abuse monitoring, configurable to 0 for qualifying organizations. This has been OpenAI’s policy since March 2023.
Anthropic API: No training on API data. Data retained 30 days default, 0 days for Enterprise. Anthropic’s API privacy stance has been consistent since launch.
Google AI API (Gemini API): No training on API data. But Google’s API terms allow them to use API data for “service improvement” unless you explicitly opt out through a separate agreement. This is the weakest API privacy of the three.
Bottom line for API: Anthropic and OpenAI are roughly equivalent (no training, configurable retention). Google is slightly behind because of the “service improvement” clause.
5 Practical Steps to Protect Your Data
1. Pay for a plan. Free tiers are not private. The $20/month for ChatGPT Plus or Claude Pro is the cheapest data protection you’ll ever buy.
2. Opt out of training everywhere. ChatGPT: Settings → Data Controls → toggle OFF. Gemini: Gemini App Activity → toggle OFF. Claude Pro: already opted out, but verify.
3. Use the API for anything truly sensitive. API usage comes with stronger privacy guarantees across all three platforms. If you’re processing contracts, medical records, or PII, the API is the only responsible route.
4. Don’t paste raw sensitive data. Even with privacy protections, best practice is to sanitize data before sending it to any AI. Redact names, account numbers, and other identifiers. The AI doesn’t need to know it’s “John Smith at Acme Corp” — “a client at a mid-size company” is usually sufficient.
5. Check the Enterprise option. If you’re using AI for business, the Enterprise plans across all three platforms include contractual data protection guarantees that consumer plans don’t. This is the only way to get true legal protection for your data.
The Honest Answer
If you ask me “is my data safe with AI tools,” my honest answer is: it depends on what you mean by “safe.”
Your data is almost certainly safe from malicious actors. All three companies have competent security teams, and breaches have been rare and limited. Your grocery list conversation is not interesting enough to steal.
But your data may be used to train the next generation of AI models, reviewed by human annotators for quality control, or — in Google’s case — connected to your broader digital profile in ways you didn’t anticipate.
The practical reality: use paid plans, opt out of training, avoid pasting raw sensitive data, and use the API for anything truly confidential. It’s not about paranoia. It’s about knowing what you’re agreeing to.