What does AI actually know about you? transcript

Every day, people type things into AI tools they wouldn't broadcast publicly at work. Medical questions. Money worries. Drafts of emails they're nervous to send. So it's fair to ask: what does an AI actually know about me? Does it remember what I said last week? Is someone reading my chats? Will I get ads based on what I just asked? I'm Zoe, and I work on the education team here at Anthropic, the company that makes Claude.

Let's walk through what happens when you share information with AI. We hope this can be helpful to you regardless of which AI you work with, whether that's Claude or others. Just keep in mind: policies vary between providers, so always confirm the specifics for whichever tool you choose. Here's a way to think about it. You should be able to see and control what your AI provider retains about you. What matters is where your data goes, why, and how long it sticks around.

Some of what you share lasts only as long as the conversation. Some of it persists to improve your experience with the product. A little of it could improve the experience for others. At every point along the way, there should be a setting that puts you in control. So let's take one detail you might share and ask, what happens to it when you close the chat? Say you're shopping for a birthday gift, and you tell an AI, "My partner is really into hiking and we live near Denver, any ideas?" For the rest of that conversation, the AI is working with those details: your partner, their love of hiking, and your proximity to Denver.

As for where these details go and what happens to them, your data can be used in different ways. The first use is the conversation itself. On its own, the model only holds context for the duration of a conversation. Open a fresh chat tomorrow and the conversation starts blank– on its own, the model has no idea what your partner likes. Your control here is simple: it only knows what you choose to put in front of it. The second use is the product's memory.

Many AI tools save details from one chat to the next. This isn't the model itself remembering you. These are details saved into your account that the model has access to when you start a new chat. Depending on the product, this may be on by default or something you switch on. So next week, the tool might incorporate your partner's love of hiking into a conversation about brainstorming fun weekend plans. In most products, you can edit it, clear it, or turn off memory entirely.

The third use is the provider's own systems. Providers' privacy policies often allow for data use in cases like keeping the service running, safety and abuse review, fixing bugs, research, and commercial purposes like serving ads, and where applicable, improving the model. At this point, close the chat and your information is kept with the provider. How that access is controlled depends on the provider, so the reliable move is to check the privacy policy and data retention settings.

The fourth use is training future models. Some providers may use conversations to improve the next version of a model. Personal data is generally removed from chats before they're used for training and your conversations aren't stored in the model to be read back to other users. But close your chat, your detail isn't really your exact words anymore– it's a pattern. Where providers train future models on conversations, many let you opt out.

So what does this mean in practice? The reassuring part is that you have control. A provider's privacy policy describes how your data is used, and you can check the product settings for additional controls. At Anthropic, we work to make those choices easy to understand. When your organization brings in Claude, model training is off by default, so there's nothing you need to do. A few habits make this easier: One: Spend a few minutes reviewing your settings.

Memory, chat history, connected apps, and the training toggle typically live in your account or privacy settings. That's where you'll find out what's being saved and how it's being used. Two: Share based on your own comfort, not a blanket rule. Before you share something, think about its uses– across the chat, the product's saved memory, or a future training run. Three: Leave out what you aren't comfortable with. If you're using an AI to reword an email, the AI may not need the recipient's real name to do a good job.

Swap in placeholders and you'll get the same value. Four: Match the product to the sensitivity. For everyday use cases, the consumer version is usually fine. For confidential or regulated data at work, look for business or enterprise plans– these often come with different data-handling terms which you can confirm in the product's documentation. So, what does AI know about you? In the moment and at baseline, it's whatever you've put in the conversation.

Beyond that, how durable it becomes depends on the product's defaults and the choices you make in settings. A few minutes upfront, plus thoughtful engagement, ensures that you stay in control of your information.