When you type something into ChatGPT or Gemini, that message doesn't just disappear after the AI replies — it travels to company servers, gets stored, may be reviewed by human contractors, can be used to train future AI models, and in ChatGPT's case, has already been handed over to a federal court in batches of tens of millions.
Most people assume AI chat is private the way a phone call feels private. It isn't. And the specifics of what happens to your data differ significantly between platforms, between free and paid plans, and between where you live in the world. I went through the actual privacy policies, the 2026 court records, and the third-party audits to put together the clearest picture I could find of where your words actually go.
This isn't an abstract privacy lecture. It's information that matters if you've ever typed a health symptom, a work frustration, a personal relationship problem, or an embarrassing question into one of these chatbots — and most of us have.
What Happens the Moment You Hit Send
The basic mechanics are the same across platforms. Your message leaves your device encrypted, arrives at the company's servers, gets processed by the AI model, and a response comes back. Simple. But then the storage and usage decisions kick in, and that's where the differences start.
What happens next depends on four questions: Is the conversation saved? Can employees read it? Is it used to train the model? And how long is it kept? The answers vary by platform, by plan tier, and by which country you're in. The table below shows where each platform stands on those four questions for their standard consumer accounts.
Quick Comparison: ChatGPT vs Gemini vs Claude (Consumer Plans)
| Question | ChatGPT (Free/Plus) | Google Gemini (Consumer) | Claude (Anthropic) |
|---|---|---|---|
| Are conversations saved? | Yes, until you delete them. After deletion, kept 30 more days. | Yes. Default retention: 18 months. Can be set to 3 or 36 months. | Yes, retained to provide the service and for safety monitoring. |
| Can human employees read your chats? | Yes — OpenAI staff and third-party contractors can access for safety/training review. | Yes — Google's trained human reviewers access conversations for quality/safety review. Retained up to 3 years if reviewed. | Yes — Anthropic staff may review conversations for safety and improvement. |
| Used to train AI by default? | Yes by default. You can opt out in Settings → Data Controls → turn off "Improve the model." | Yes by default. Opt out: Google Account → Data & Privacy → Gemini Apps Activity → Turn off. | No by default. Anthropic states it does not train on user conversations without explicit consent. |
| Your protection if you upgrade to paid? | ChatGPT Team/Enterprise: no training on your data by default. | Workspace Business/Enterprise: no training on your data by default. | API/Claude.ai Pro: same default no-training policy. |
| Location matters? | EEA/UK users excluded from some court-ordered data preservation rules. | EEA/Switzerland/UK users get paid-tier data protections even on free plans. | GDPR applies to EEA users. |
ChatGPT: What OpenAI Actually Does With Your Data in 2026
I'll start here because it's the most complicated story, especially after 2025's legal drama. OpenAI's privacy policy, updated May 2026, lays out the basics clearly: your conversations are stored until you delete them, and then kept for 30 more days. They can be reviewed by OpenAI employees and third-party contractors. They're used to train the model unless you turn that off. None of this is hidden — it's just written in legal language most people skip.
The part that caught a lot of people off guard was what happened in 2025 and into 2026. In May 2025, a federal judge ordered OpenAI to preserve all ChatGPT conversation logs indefinitely as part of the New York Times copyright lawsuit. Even if you clicked "Delete," your conversations were archived in a legal hold. That preservation order was modified in October 2025, restoring standard deletion rights for most users going forward — but in January 2026, a judge then ordered that 20 million preserved logs be handed over to the plaintiffs. Twenty million conversations. That's not a privacy policy edge case. That's a real-world outcome of using a cloud AI service during active litigation.
There's also a newer risk most people haven't noticed: Agent Mode. When ChatGPT's agent operates for you — browsing the web, accessing your files — it retains data including screenshots of your browser for 90 days. That's significantly longer than standard chat retention, and the scope of what gets captured is much wider. If you use ChatGPT agents to do anything involving sensitive sites or documents, you're now dealing with a 90-day screenshot archive, not just a chat log.
What you can do right now: Go to Settings → Data Controls and toggle off "Improve the model for everyone." For sensitive conversations, use Temporary Chat mode, which isn't saved to history and isn't used for training (though OpenAI still keeps it for 30 days for safety checks). If you're on a business plan, model training is already off by default.
Google Gemini: The Ecosystem Problem
Gemini's data situation is different from ChatGPT's in a specific way that matters a lot. When you evaluate ChatGPT, you're asking what OpenAI does with your AI conversations. When you evaluate Gemini, you're asking what Google does with your AI conversations and everything else it already knows about you — your Gmail, your Search history, your YouTube viewing, your Maps location history, your Chrome browsing. Gemini doesn't just store what you type. It can connect it to that entire existing profile.
Google's Gemini Apps Privacy Hub confirms the retention timeline: consumer conversations are stored for 18 months by default (you can change it to 3 or 36 months). Conversations reviewed by human reviewers are kept for up to 3 years, disconnected from your Google Account. And training is on by default for free-tier consumer users — you need to actively turn it off. On the Gemini mobile app, the picture gets broader still: Gemini can access call logs, installed apps, device usage patterns, and whatever's on your screen when you use the overlay feature.
The exception worth knowing: if you're in the European Economic Area, Switzerland, or the UK, Google applies the stronger paid-tier data protections to you even on the free plan. If you're in the US, the default free-tier settings apply, which means training is on, human review is possible, and your conversations can be connected to your broader Google account data.
There's also a 2026 lawsuit worth tracking: Thele v. Google LLC, filed in the Northern District of California, alleges that Google used existing "Smart Features" settings as a backdoor to enable Gemini's deeper Gmail access without obtaining clear additional user consent. As of March 2026, that case was in discovery.
What you can do: Go to myaccount.google.com → Data and Privacy → Gemini Apps Activity → Turn off. This stops future conversations from being used for human review or model training. For Gmail access specifically, go to Gmail Settings → General → Gemini for Gmail and disable it there separately if you want to limit cross-service data access.
The Things That Don't Change No Matter What You Do
Here's the part that most privacy guides gloss over: some data risks aren't solvable by adjusting settings.
First, once data has been used for model training, it doesn't "un-train." When you type proprietary code, a specific personal situation, or sensitive details into an AI model and that data gets incorporated into training, you can delete the conversation — but you can't extract the learned patterns from the model. As one engineer who has worked on AI systems explained it: the model doesn't store your text, it stores the relationships between words. You can strip identifiers but you can't un-learn what was learned.
Second, courts can demand your data regardless of your privacy settings. The January 2026 handover of 20 million ChatGPT logs was a demonstration of this. If you discuss a legal situation, a business dispute, or anything that might become relevant to litigation, your AI conversations are potentially discoverable — and "I turned off model training" doesn't change that.
Third, data breaches happen. There was a significant ChatGPT data leak in 2023. In July 2025, search engines were indexing thousands of ChatGPT conversation links, briefly exposing sensitive queries that users had accidentally made public. The more data a company stores, the larger the target it presents.
The Practical Bottom Line
- For everyday questions and tasks: The default settings on free ChatGPT and Gemini are fine for non-sensitive use. Just know that training is on and employees may read flagged conversations.
- For anything sensitive: Turn off model training in settings, use Temporary Chat on ChatGPT, use Gemini without a Google account or with activity turned off. Don't type health details, personal relationship information, financial specifics, or anything related to legal matters you're involved in — on any platform.
- For work or business use: Paid enterprise tiers from both OpenAI and Google offer no-training commitments, admin controls, and better data governance. Free consumer plans are not appropriate for confidential business information.
- For maximum privacy: Consider local AI models (Ollama, LM Studio, Jan) that run entirely on your own device. Nothing leaves. But you'll trade off capability for privacy, and the tradeoff is real.
Frequently Asked Questions
Does ChatGPT use my conversations to train its AI?
Yes, by default on free and paid personal plans. You can turn it off in Settings → Data Controls → "Improve the model for everyone." Business and Enterprise plans have training off by default. Turning off training doesn't prevent OpenAI employees from accessing conversations for safety review, and it doesn't affect data that's already been used.
Does Google Gemini share my chats with advertisers?
Google states it does not use Gemini conversations for ad targeting directly, but Gemini data is integrated with your Google account profile, which does inform advertising across Google services. Conversations are retained for 18 months by default and may be connected to your Gmail, Search, and browsing history.
Is Claude more private than ChatGPT?
On training specifically, yes — Anthropic's default policy states it does not train on user conversations without explicit consent, which is a meaningful difference from ChatGPT's opt-out default. However, conversations are still stored and staff may still review them for safety purposes, so Claude is not fully private either.
Can I delete my ChatGPT or Gemini conversations and have them truly gone?
Not immediately. ChatGPT keeps deleted conversations for 30 additional days, and any conversation reviewed by a human at either platform may be retained up to 3 years. Legal cases can also require preservation regardless of your deletion requests, as happened with tens of millions of ChatGPT logs in the 2025-2026 New York Times lawsuit.
What's the safest way to use AI if I need to discuss something sensitive?
Use Temporary Chat on ChatGPT or disable Gemini Apps Activity before the conversation, and don't include identifying details unless necessary. For anything involving legal, medical, or financial specifics, consider a local AI model that processes data entirely on your device, or consult a professional rather than an AI.