AI and Data Privacy: A Complete Guide to What Gets Collected and How to Protect Yourself

The Direct Answer

What happens to your data depends heavily on which AI tool you’re using, which pricing tier you’re on, and which specific settings you’ve touched — there is no single honest answer that covers “AI chatbots” as a category. The short version: consumer free and low-cost tiers from OpenAI, Anthropic, and Google can all use your conversations to train future models unless you actively opt out, while business, enterprise, and API tiers from all three generally exclude your data from training by default. This guide covers the current, provider-specific policies, the regulations that actually apply, and the practical steps and honest limits of protecting yourself.

What OpenAI Actually Does With ChatGPT Conversations

On Free, Plus, and Pro consumer accounts, OpenAI may use your conversations to improve its models unless you turn that off — the setting lives under Settings → Data Controls → “Improve the model for everyone.” Once switched off, new conversations are excluded from future training runs. On Team, Enterprise, and API usage, business data is not used for training by default, without requiring any opt-out action. Separately from training, OpenAI generally retains conversations for a period — commonly cited around 30 days — to check for abuse before deletion, meaning “not used for training” and “not stored” are two different things even when both privacy protections are active.

There’s a real, documented limit worth knowing here: in 2025, a federal court ordered OpenAI to preserve ChatGPT output logs — including conversations users had deleted — as part of a New York Times copyright lawsuit, before later narrowing that obligation. OpenAI restricted access to the preserved data to a small legal and security team and said none of it would be handed to outside parties absent further court process, but the episode is a concrete example of how a legal proceeding can override a user’s own deletion settings.

What Anthropic Actually Does With Claude Conversations

Anthropic’s policy, updated in 2025, moved consumer accounts (Free, Pro, and Max) to a model where you actively choose whether new or resumed conversations can be used to train future Claude models — this only applies going forward, not to your prior chat history. If you opt in, Anthropic states it may retain that data in de-identified form for up to five years to support training and safety work; if you opt out, retention drops to 30 days and the data isn’t used for training. Commercial usage — Claude for Work, the API, and similar business products — is governed separately and isn’t affected by the consumer training setting either way. One caveat worth flagging plainly: conversations flagged for safety review can still be used to help train Claude’s safety systems regardless of your training preference, an exception that’s easy to miss in a quick read of the settings page.

What Google Actually Does With Gemini Conversations

On standard consumer Gemini plans, activity is reviewed and can be used to improve Google’s AI products unless you turn off “Gemini Apps Activity” (found under myaccount.google.com → Data & Privacy). With that setting off, future conversations aren’t sent for human review or used for training. Gemini also offers a Temporary Chat mode, which isn’t saved to your account, isn’t used to improve the model, and is automatically deleted after a short window. Default retention for regular Gemini activity has been reported around 18 months, extending to as long as three years for any specific conversation a human reviewer has looked at. For Google Workspace, Google Cloud, and paid Gemini API usage, Google states it does not use that data to train its models without the customer’s explicit permission — the same consumer-versus-business split seen at OpenAI and Anthropic.

Why Business and API Tiers Are Treated Differently

The consistent pattern across all three providers is that consumer conversations are the default source of training data (opt-out at OpenAI and Google, opt-in going forward at Anthropic), while paying business customers on enterprise, Workspace, “for Work,” or API tiers get a contractual promise that their data won’t train the underlying models without separate permission. This isn’t an accident — it reflects genuine commercial incentive: consumer conversations at scale are a valuable, low-cost training data source, while business customers routinely handle confidential information and would not adopt tools that quietly fed that data into a shared model. If you’re regularly discussing anything sensitive with an AI tool — client information, internal business data, anything covered by a confidentiality obligation — using a business or API tier with an explicit no-training data agreement is a meaningfully different privacy posture than using the free consumer app, even from the same company.

The Regulatory Landscape

The EU’s GDPR remains the strongest baseline privacy law that applies to AI providers operating in or serving the EU, giving users rights to access, correct, and delete their personal data, including data processed by AI systems. In the US, there’s still no comprehensive federal privacy law, but the patchwork of state laws has grown substantially — by 2026, roughly fifteen states have active comprehensive data privacy laws, and several have layered AI-specific rules on top. California’s AI Transparency Act adds disclosure requirements around AI-generated content, and Colorado’s AI Act (with a primary effective date pushed to mid-2026) requires “reasonable care” from developers and deployers of high-risk AI systems to prevent algorithmic discrimination. None of this amounts to a single clear national rulebook, which is exactly why relying on regulation alone to protect sensitive information is not a safe assumption — the specific provider’s policy, not a general sense that “there must be a law,” is what actually governs your data in most everyday chatbot use.

Practical Steps You Can Actually Take

A few concrete actions apply across most providers: turn off the model-training toggle in your account’s privacy or data-control settings if you don’t want conversations used for training; use temporary or incognito-style chat modes for anything you don’t want saved to your account at all; and periodically review and delete old conversation history rather than assuming it ages out on its own. Beyond settings, the more reliable protection is behavioral: don’t paste anything into a consumer AI chatbot that you wouldn’t be comfortable with a company employee or a future data breach exposing — that includes Social Security numbers, full financial account details, medical records, passwords, and confidential work material, unless you’re specifically using a business tier with a contractual no-training guarantee and your employer has cleared that use.

The Honest Limits of Privacy Settings

None of these settings amount to a guarantee that your data was never seen or processed by anyone. Turning off training doesn’t stop the message from being transmitted and processed on the provider’s servers to generate a response, and every major provider retains some window of data for abuse and safety monitoring regardless of your training preference. Deletion typically isn’t instantaneous or absolute — most providers keep a short retention window even after you delete something, and as the OpenAI litigation example shows, a court order can override a deletion setting entirely. And opting out of training only affects future conversations at every provider covered here — it does not retroactively remove your past conversations from data that may have already been used in a completed training run. Treat these settings as real, meaningful controls over what happens going forward, not as an undo button for the past or an absolute privacy guarantee.

When AI Assistants Connect to Your Other Accounts

A separate and growing privacy consideration comes from AI assistants that don’t just answer questions but connect directly to your other accounts — Gemini reading your Gmail and Google Drive, Copilot working with your Microsoft 365 documents, or Claude and ChatGPT connecting to third-party apps through integrations. This is a meaningfully different exposure than typing a question into a chat box: once an assistant is connected, it can potentially see the contents of documents, emails, or files you never explicitly typed into the conversation, and the privacy implications of that access are governed by the same consumer-versus-business policy split covered above, but with a larger practical surface area. Before granting an AI assistant this kind of standing access to another account, it’s worth checking specifically what that integration can read, whether that access is scoped to a single session or persists until manually revoked, and whether the underlying account (personal Gmail versus a company Workspace account, for instance) falls under a consumer or business data policy — because the answer changes what, if anything, that connected data contributes to model training.

A Few Common Real-World Scenarios

A handful of situations come up often enough to address directly. Uploading a resume or cover letter to a free consumer chatbot for editing help is common and low-risk in most cases, since the content is typically already meant to be shared publicly with employers — the bigger risk is pasting in unpublished, proprietary work documents under the same casual assumption. Asking a general medical or legal question is different from pasting in an actual diagnosis, prescription, or case file with identifying details attached; the former is fine on a consumer account, the latter starts to look more like the confidential information a business tier’s no-training guarantee actually exists for. And using an AI coding assistant on proprietary company code is worth a second look specifically at whether your employer has approved that tool and tier for the purpose — several companies have had internal source code end up processed by consumer-tier AI accounts before official policies caught up with employee habits, which is exactly the scenario a company’s IT or legal team would want flagged before it happens rather than after.

Bottom Line

The single most useful fact in this guide is the consumer-versus-business split: free and consumer AI accounts can train on your conversations by default (OpenAI, Google) or by opt-in going forward (Anthropic), while business and API tiers generally don’t train on your data at all — so which tier you’re using matters as much as which provider you’re using. Combine the right tier with the training opt-out, temporary chat modes for sensitive one-off conversations, and a simple personal rule about what never gets typed into a consumer chatbot, and you’ll have addressed the parts of this that are actually within your control.