What Happens to Your AI Chats

Every time you ask ChatGPT for help with a work email, share a personal dilemma with Claude, or paste code into Gemini, you're handing over data. Sometimes sensitive data.

But what actually happens to those conversations? Can someone at OpenAI read your chat about your failing marriage? Is your proprietary code being used to train the next model? And as a European, do you have any say in this?

I spent a few hours digging into privacy policies so you don't have to. The short version: it's complicated, and the defaults are not privacy-friendly.

The Default: Your Chats Train Future Models

Let's start with what happens if you sign up for a free account and use these services without changing any settings.

ChatGPT (OpenAI): Your conversations may be used to train future models. Human reviewers may read your chats for quality purposes. Data is retained indefinitely unless you delete it.

Claude (Anthropic): For free and Pro users, conversations may be used for training if you've opted in (or haven't opted out — the default has changed over time). They de-link conversations from your account before training use.

Gemini (Google): Conversations are used to improve products, which includes training. Human reviewers may see your chats. Google being Google, there's also the integration with your broader Google account activity.

Microsoft Copilot: For consumer versions, conversations may be used to improve the service. Enterprise versions have stronger guarantees.

The pattern: consumer tiers assume consent to training. Enterprise tiers don't.

What "Training" Actually Means

When companies say your data "may be used for training," they mean your conversations could become part of the dataset that teaches the next version of their AI how to respond.

This doesn't mean the AI memorises your chat and can repeat it verbatim. Models learn patterns, not specific memories. But:

Unusual or distinctive content has a higher chance of being "memorised"
Security researchers have extracted training data from models before
Even anonymised data can sometimes be re-identified

For most casual conversations, this isn't a big deal. But if you're sharing proprietary business information, personal health details, or anything confidential? That's different.

How to Opt Out of Training

Every major AI provider now offers ways to prevent your chats from being used for training. Here's how:

ChatGPT

Go to Settings → Data Controls → Improve the model for everyone and turn it off.

Alternatively, use ChatGPT's "Temporary Chat" feature — these conversations aren't saved to your history and aren't used for training.

Note: Even with training disabled, OpenAI retains your data for 30 days for "safety monitoring" before deleting it.

Claude

Go to Settings → Data Privacy Controls and disable the option to improve Claude.

Anthropic has been relatively transparent about this — they automatically strip your user ID from any data before training use, and they've published research on privacy-preserving analysis.

Gemini

Go to your Google Account → Data & Privacy → Gemini Apps Activity and turn it off.

Be aware this is separate from your general Google activity settings. Google's ecosystem means your AI usage intersects with your search history, email, and everything else they know about you.

Microsoft Copilot

For consumer accounts, check Settings → Privacy for any available controls. The options vary depending on which Copilot product you're using.

The Enterprise Difference

If you're using these tools through work, the rules are different:

ChatGPT Enterprise/Team: OpenAI explicitly states they don't train on your data. Conversations stay within your organisation.

Claude for Work/API: Business tiers don't use your data for training by default. This is why companies pay the premium.

Gemini for Workspace: Similar guarantees for business customers. Your organisation controls the data.

Microsoft 365 Copilot: Enterprise data stays within your Microsoft tenant and isn't used for general training.

The privacy gap between free and paid tiers is deliberate. If you're not paying, your data is part of the product.

Your GDPR Rights

If you're in Europe, you have specific legal rights that these companies must honour:

Right to Access (Article 15): You can request a copy of all personal data they hold about you, including your conversation history.

Right to Erasure (Article 17): You can request deletion of your data. This includes conversation history.

Right to Object (Article 21): You can object to processing of your data for training purposes. This is separate from their in-app toggles — it's a legal demand.

In practice, the in-app privacy controls are easier to use. But if you want formal confirmation that your data has been deleted, send a GDPR request.

For ChatGPT: privacy@openai.com
For Claude: privacy@anthropic.com
For Gemini: Use Google's GDPR request form in account settings

Use the magic words: "Under Article 17 of the GDPR, I request the erasure of all personal data you hold about me, including conversation history and any derived data."

What You Shouldn't Share

Regardless of privacy settings, I'd avoid putting certain things into any AI chat:

Passwords and credentials: Obvious, but people do it. "Can you help me fix this config file?" and then paste their API keys.

Personal identification documents: Don't paste your passport number or tax ID to "check if it's formatted correctly."

Truly confidential business information: Trade secrets, unreleased product details, M&A discussions. Your company's lawyers will not be happy.

Other people's private information: Don't paste your friend's email about their health problems to help you respond. That's their data, not yours to share.

Anything you'd be uncomfortable seeing leaked: Assume any chat could theoretically become public. Would you be embarrassed? Don't type it.

Local Alternatives

If privacy is paramount, you can run AI models locally. Nothing leaves your machine.

Ollama is the easiest way to run open-source models locally. Install it, pull a model like Llama 3 or Mistral, and you have a private AI assistant.

LM Studio offers a nice GUI for running local models on Mac, Windows, or Linux.

Jan.ai is another user-friendly option designed for local-first AI.

The trade-off: local models are less capable than GPT-4 or Claude. They require decent hardware (8GB+ RAM minimum, more for better models). But for straightforward tasks — drafting emails, summarising documents, coding assistance — they work fine.

I've been running Llama 3 locally for tasks I don't want touching cloud servers. It's not as smart as Claude, but it's perfectly adequate for most things, and I know exactly where my data goes: nowhere.

Practical Privacy Tiers

Here's how I think about it:

Casual use (recipes, general questions, creative writing): Cloud AI is fine. Enable the training opt-out if you can be bothered, but it's not critical.

Work tasks with non-sensitive data: Cloud AI with training disabled. Use your company's approved tier if available.

Sensitive personal matters (health, finances, relationships): Either use a privacy-friendly provider with training disabled, or go local. Don't use free tiers.

Confidential business data: Either enterprise tiers with explicit privacy guarantees, or local models. Never use consumer cloud AI.

Anything you truly can't risk: Local only. Or don't use AI at all.

The Browser and App Split

One more thing: the ChatGPT mobile app and web interface may have different default settings. Check both.

Also consider which account you're logged into. Using ChatGPT while logged into your personal account at work mixes your personal and professional data.

If you use AI at work, either use your company's official account/tool, or use an entirely separate personal account with different credentials.

My Setup

For what it's worth, here's what I do:

Claude Pro with training disabled for most work tasks
Local Ollama for anything I don't want leaving my machine
ChatGPT (with training disabled) for tasks where it's clearly better
I never paste credentials, personal documents, or other people's private info into any AI

Is this paranoid? Maybe. But data shared with cloud services exists outside your control, potentially forever. It takes two minutes to toggle the privacy settings and costs you nothing.

The Bottom Line

The default settings on AI chatbots are not designed for your privacy. They're designed to improve the product, which means training on your data.

If you care about privacy:

Turn off training data collection in settings
Don't share sensitive information with free tiers
Consider local models for confidential tasks
Remember your GDPR rights

AI assistants are incredibly useful. But treat them like what they are: services run by companies that want your data. Give them what's useful for you. Keep the rest to yourself.