What AI Companies Know About You

You typed something into ChatGPT last Tuesday. Maybe it was a question about your health. Maybe it was a draft resignation letter, a message to an estranged family member, a business idea you hadn’t told anyone about yet. You got your answer, closed the tab, and moved on.

That conversation didn’t disappear.

It went to OpenAI’s servers. It may have been flagged for human review. It was likely used — unless you explicitly opted out — to train the next version of the model. And even if you deleted it, it sat in OpenAI’s systems for another 30 days under what the company calls a “safety review” retention period.

Most people have no idea this is happening. And OpenAI is just one of six major AI companies doing some version of it.

The short answer: Every major AI company logs your prompts, plus metadata like your device, IP address, and session times. Most use free-tier conversations to train future models unless you opt out. OpenAI, Google, and Anthropic default consumer chats into training; enterprise and Workspace accounts are protected. Meta AI is embedded in WhatsApp and Instagram with no real opt-out outside the EU. Apple processes most requests on-device. The per-company settings that limit this are below, and our full opt-out guide lists the exact toggle for each.

Jump to a company:

OpenAI / ChatGPT

ChatGPT has somewhere north of 200 million weekly active users. That’s a lot of people typing very personal things into a chat box.

By default, every conversation you have with ChatGPT is used to train future versions of the model. OpenAI calls this “improving the model for everyone.” You can turn it off under Settings > Data Controls > Improve the model for everyone. The toggle defaults to on, and most users never find it.

If you delete your conversation history, the data doesn’t vanish immediately. OpenAI retains it for 30 days after deletion for what it describes as safety and abuse review. That’s 30 days of retention on content you already told the system to remove. (A 2025 court order briefly forced longer retention; it was lifted in late 2025, so the 30-day window applies again as of mid-2026.)

The mobile apps add another layer. ChatGPT’s iOS and Android apps collect device identifiers, usage data, crash diagnostics, and interaction patterns. You are not just tracked on what you say. You are also tracked on how you use the app, when, from what kind of device, and how long each session runs.

OpenAI’s privacy policy gives the company latitude to share your data with third-party service providers and for “legal obligations.” That last phrase is particularly open-ended. A legal obligation in one jurisdiction may look nothing like what you’d expect in another.

The tier structure here matters. Enterprise and API customers get meaningfully stronger protections — conversations are not used to train models by default, and data processing agreements are more specific. Free users get the weakest protections. The math here is not subtle: if you’re not paying, the product is your data.

Google Gemini

Google has been reviewing AI conversations with humans longer than most people realize. Gemini is no different. Google’s own FAQ confirms that Gemini conversations are reviewed by human reviewers for safety training — not just automated systems, actual people reading what you wrote.

If you access Gemini through a personal Gmail account, your conversations are stored for 18 months by default (users can select 3 months or 36 months). If you used Gemini today and forgot to turn anything off, Google may hold that conversation until late 2027.

You can find and pause your Gemini activity at myactivity.google.com — look for Gemini Apps Activity. Pausing future collection is straightforward. Clearing past activity is more work, and like most “deletion” controls in Google’s ecosystem, the timeline on when data actually leaves their systems is not specified precisely.

Google Workspace accounts — corporate Gmail, school accounts — have stronger protections. In those environments, conversations are not used to train Google’s AI models. If your employer uses Workspace, you’re covered. If you’re using a personal account, you’re not.

Meta AI

Meta’s AI isn’t a separate product you choose to use. It’s embedded directly into WhatsApp, Instagram, Facebook, and Messenger — applications that hundreds of millions of Indians use every day. The Meta AI button appears in your WhatsApp search bar whether you want it or not.

Conversations with Meta AI are subject to Meta’s standard privacy policy, which means they may be used to improve Meta’s products. That’s a broad framing that gives Meta significant room to do what it wants with what you share.

Here’s the part that should give you pause: Meta’s data practices are already extensive before the AI enters the picture. The Meta Pixel tracks you on millions of third-party websites. Meta builds advertising profiles on people who have never created an account, using contact data uploaded by their friends. Adding an AI layer that reads your actual typed questions compounds everything that’s already happening.

EU users get somewhat stronger protections. GDPR enforcement has pushed Meta into more explicit consent flows in Europe, and German and Irish regulators have challenged specific practices. Indian users have no equivalent enforcement mechanism in place. The DPDPA’s cross-border transfer rules are still being operationalised, and there has been no enforcement against Meta under DPDPA as of mid-2026.

If you want to stop Meta AI from seeing your data, you effectively cannot — not if you continue using WhatsApp or Instagram. The only meaningful opt-out for most users is to stop using those platforms. EU users can object under Article 21 of GDPR; everyone else is working within Meta’s terms.

Microsoft Copilot

Copilot is not one product. It’s a brand name applied to several different things with different privacy rules, and the difference matters enormously.

Microsoft 365 Copilot — the enterprise version deployed through corporate accounts — comes with a documented privacy commitment: your organizational data is not used to train Microsoft’s foundation AI models. Microsoft publishes this clearly. If you’re using Copilot through a company-managed Microsoft 365 account, your documents, emails, and Teams conversations stay within your organization’s data boundary.

Consumer Copilot — the free version available in Windows 11, Edge, and at copilot.microsoft.com — has different terms. Conversation data from consumer Copilot may be reviewed by humans for quality and safety purposes. The privacy dashboard is at privacy.microsoft.com, and you can clear activity there, but the underlying data retention policies are less specific than the enterprise documentation.

The integration depth of Copilot is what makes this worth paying attention to. It’s built into Windows 11 itself, into Edge, into Outlook, Word, and Excel. This is an AI assistant that potentially touches your work documents, your calendar appointments, your email drafts. The scale of access is qualitatively different from a standalone chat interface.

Apple Intelligence

Apple has made privacy architecture a core part of its AI pitch — and unlike most corporate privacy claims, this one has a technical structure you can actually inspect.

Most Apple Intelligence requests are processed directly on your device using the A-series chip. Your query, your result, your data — none of it touches Apple’s servers. This is the design, not a setting.

For more complex tasks that require server-side processing, Apple routes requests through what it calls Private Cloud Compute. Apple’s claim has three parts. These servers don’t retain your data. The code running on them is auditable by independent security researchers. And Apple cannot inspect the contents of your requests, even in transit. Apple has made the Private Cloud Compute code available for external verification, an unusual step for a major tech company.

Apple also does not use Apple Intelligence data to build advertising profiles. This is a genuine differentiator. Google’s business model runs on advertising. Meta’s business model runs on advertising. Apple’s hardware revenue model means it has a different structural incentive — the product is the iPhone, not the user profile.

There’s nothing to configure in Apple Intelligence for privacy. On-device processing is the design decision. You either trust the architecture or you don’t.

Anthropic (Claude)

Anthropic operates Claude, which is available directly through claude.ai and through an API used by many other products and platforms.

For enterprise API customers, Anthropic’s default is not to train on customer data, similar to OpenAI’s enterprise policy. Developers building on Claude through the API have contractual protections around their users’ data.

Consumer access through claude.ai sits in a different category, and it changed in 2025. Anthropic had previously promised not to train on consumer conversations. In August 2025 it reversed that. Free, Pro, and Max users now have their new chats used to train Claude unless they opt out. The retention period extends to five years if they leave the setting on. Turning it off keeps you on the previous 30-day retention. You control this under Settings > Privacy > “Improve Claude for everyone.” Anthropic is also explicit that human reviewers may read flagged conversations. That review happens for safety and quality evaluation, even if you opt out.

One detail worth noting: Claude processes longer, more complex conversations than most AI assistants. Users frequently paste in entire documents, contracts, emails, or code repositories. The volume of sensitive material flowing through a typical Claude session can be substantial. The review and training policies apply to that material.

What every one of these companies shares

Regardless of the specific training policies, every major AI company collects the same baseline metadata. That means timestamps of when you used the service, device information, IP address, session duration, and interaction patterns.

This metadata is not the transcript of your conversation. It is, however, enough to build a detailed picture of your behavior — when you use AI, from where, for how long, how often. Metadata analysis has been used in law enforcement contexts for decades precisely because it’s so revealing even without content.

None of these companies give you a single dashboard that shows everything they hold on you in one place. Data is spread across conversation logs, training datasets, billing systems, safety review queues, and product analytics. Subject access requests, where they’re honored, typically return a partial and formatted export — not everything.

EU and UK users have the right to submit a Subject Access Request under GDPR, compelling companies to disclose what personal data they hold. Indian users have a similar right in principle under the Digital Personal Data Protection Act. In practice, the DPDPA’s enforcement infrastructure is still being built, and no major AI company has faced a meaningful Indian regulatory action over data handling as of 2026.

The India problem specifically

Several major AI providers operate Indian data centers. Microsoft Azure has regions in Pune and Chennai, and Google Cloud operates a Mumbai region. But the AI services most consumers use — ChatGPT, Gemini consumer, Claude.ai — route through infrastructure outside India. Every prompt you type into those services leaves India the moment you hit send. Your data travels to servers in the United States, the European Union, or wherever each company’s AI infrastructure sits.

The DPDPA was supposed to address this. Its cross-border data transfer provisions require that transfers happen only to countries with comparable data protection — a list the Indian government has not yet finalized. The rules exist; the enforcement does not. Until that changes, every Indian user sending data to an American AI company is doing so without any regulatory backstop protecting them.

The practical result: Indian users on free consumer AI tiers have among the weakest data protections of any major user group globally. EU users have GDPR enforcement with real teeth — billion-dollar fines have been levied. US users have the California Consumer Privacy Act in some states and sector-specific rules. Indian users have a law with strong language and no enforcement history yet.

This is not a reason to avoid AI tools. It is a reason to understand exactly what you’re accepting.

How to actually limit what they collect

Each platform has a different process. Here’s what works:

ChatGPT: Open Settings, go to Data Controls, find “Improve the model for everyone,” and turn it off. Also consider turning off chat history entirely if you want the most protection — this stops conversations from being stored beyond the session. Delete chats you don’t need, and remember the 30-day retention window applies after deletion.

Gemini: Go to myactivity.google.com, find Gemini Apps Activity, and turn it off. You can also delete past activity there. If you have a Google Workspace account through work or school, check with your administrator — your organization may already have stronger controls in place.

Meta AI: There is no meaningful opt-out for Indian users. If you use WhatsApp or Instagram, Meta AI is part of the package under Meta’s existing privacy terms. EU users can file an objection under GDPR Article 21, which has some force. Everyone else has the nuclear option: stop using the platforms that embed Meta AI.

Microsoft Copilot: Use the Microsoft Privacy Dashboard at privacy.microsoft.com to review and clear activity data. If your employer provides Microsoft 365, use that account rather than a personal consumer account — the enterprise protections are substantially stronger and documented in Microsoft’s commercial terms.

Apple Intelligence: Nothing to configure. The on-device processing model is the privacy protection. If you’re using Apple devices and Apple Intelligence, that’s already the most private AI option available from a major company.

For the exact toggle at each company, including Adobe, LinkedIn, X, and Spotify, see our step-by-step guide to opting out of AI training.

What you can do right now

Want to see what data brokers expose about you? See what a scan finds → shows an example report: the kinds of sites that list your name, phone, and address. Once KnytX Shield launches it will find and remove these automatically. Join the waitlist to get in early.