AI and Your Privacy
Where does what you type actually go
① "Whatever you type into AI gets used for training" is only half right. Whether it enters a model's training data and whether a human can review it as a stored log are two different questions, and the answers vary by service and plan.
② The riskiest input isn't your own data — it's someone else's. The moment you paste in company files, a customer's information, or a coworker's name, the nature of the risk changes.
③ Most services put a training opt-out toggle in account settings. The problem is that few people know it's there.
① What happens the moment you type something
As covered in how an LLM works, most of the AI covered in this series runs on a server. Type a question and it leaves your device for a company's server, gets computed, and the answer comes back. That round trip isn't unusual — the same thing happens when you send an email or upload a file to the cloud.
What's different is what happens next. With an email, it's usually clear that only the recipient and you see it. What happens to what you type into an AI service afterward varies by provider, and the answer usually sits inside terms of service nobody reads. This piece doesn't summarize that fine print — it lays out what to actually ask.
| What can happen after you submit it | What to check |
|---|---|
| Used as training data to make the model better | Is there an opt-out toggle in account settings? |
| Kept as a log a human can review for safety or quality checks | How long is it retained, and can you request deletion? |
| Wired into other features of the service (memory, personalization) | Does information carry across conversations, or reset each time? |
② What "used for training" actually means
"What I type gets used to train AI" collapses two separate questions into one sentence.
- Question 1 — model training: Does your conversation become part of the dataset used to train the next version of the model?
- Question 2 — human review: Can a human directly read your conversation to check for safety issues or abuse?
These two switches toggle independently. A service that says "not used for training" often carries a separate clause allowing human review for policy violations. The reverse also happens — used for training, but never reviewed by a human. Confirming just one and concluding "this is safe" is where people go wrong.
| Account type | Typical default | What to watch for |
|---|---|---|
| Free / personal account | Training use is often on by default (opt-in) | Usually requires you to switch it off manually |
| Paid personal subscription | Varies by service — still worth checking | "Paid" is not a guarantee of "safe" |
| Business / team plan | No-training-use is often the contractual default | Confirm whether you're on the company license or a personal account |
A personal paid subscription and an enterprise license are not the same thing. A plan you pay for with your own credit card may carry different data-handling terms than a business license your employer negotiated separately. If you're using a personal paid account for work, it's worth asking IT or legal whether it carries the same protections as the company's actual enterprise contract.
③ How long conversations last, and who can see them
"The AI sees your conversation" is, strictly speaking, imprecise. The model itself remembers nothing once a conversation ends — as covered in tokens and the context window, the context window is a bowl refilled every turn. The real question isn't "does the model remember" but "does the company store this, and can a human open it?"
Many services reserve the right to have a human review some conversations to handle policy violations or safety issues. This is closer to an abuse-prevention procedure than covert surveillance, but the outcome for the user is the same: what you typed could be read by someone you've never met.
Deleting a conversation often doesn't make it vanish instantly. Legal retention requirements or backup policies can keep a copy inside the system for some period. It's more accurate to think of "delete" as "stop using this going forward" rather than "erase immediately."
④ Pasting in company data — where the stakes change
Typing your own personal information and typing someone else's information or company files are entirely different problems. With your own data, you decide how much risk to accept. The moment you paste in a customer's details or a coworker's name, you've handed that risk to someone else.
Cases that come up constantly in practice:
| Situation | What leaks |
|---|---|
| Pasting an entire file to debug code | API keys or passwords sometimes sitting right in the code |
| Having a contract draft reviewed | Undisclosed counterparty names, amounts, terms |
| Pasting a support ticket to draft a reply | Customer names, contact info, purchase history |
| Asking for a meeting summary | Coworkers' remarks, personnel-related discussion |
This is why companies roll out dedicated enterprise AI tools (products like Microsoft Copilot). This kind of enterprise AI demand is now covered as a core growth driver in recent big-tech earnings — commercial bookings hitting record highs is covered in the big-tech Q2 earnings guide. Enterprise tools more often carry explicit contractual terms around data isolation and non-training guarantees than personal accounts do.
Before pasting company material into AI, ask: would it be fine if this leaked exactly as-is? If not, use a company-approved enterprise tool rather than a personal account — and if no such tool exists, don't paste it at all. The same principle applies to sensitive documents an investor handles, like filing records or tax paperwork discussed in the overseas-stock tax guide — those, too, mix personally identifying and financial information that needs careful handling.
⑤ What leaks without you noticing — risk inside the prompt itself
Pasting in an entire file isn't the only way this goes wrong. Even a short question can carry information you never meant to share.
- Details that slip out while explaining context — a name and department end up in the prompt while you're describing a situation ("our team's John Smith said...")
- Metadata in an uploaded image — a photo file can carry hidden location or timestamp data
- Information reconstructable from several prompts — each item anonymous on its own, but identifiable once you combine them
These leaks don't come from bad intent — they come from habit. The casual, conversational way people write to AI carries over the things they'd never say out loud to a stranger, typed in without a second thought.
⑥ A practical checklist
Understanding the mechanics reduces the response to a handful of habits.
| Do this | Why |
|---|---|
| Check the training-use setting directly | Most services have a toggle in account settings. Don't leave the default unchecked — look at it once. |
| Use temporary/incognito mode for sensitive chats | Many services offer a mode that doesn't save the conversation — good for one-off sensitive questions. |
| Separate work and personal accounts | Keep company material inside company-approved tools only. |
| Skim before pasting | Before dropping in a long document, check once for names, contact details, or contract terms mixed in. |
| Route truly sensitive work on-device | There's a mode where nothing leaves the device at all — covered in on-device AI. Even then, some app features can still route to a server, so don't assume "on-device" automatically means "safe." |
The moment you type something into AI, it's not a private note — it's data sent to a company's server. "Used for training" and "reviewable by a human" are different questions, and the thing to watch most carefully isn't your own information — it's someone else's. Checking the training-use setting once removes most of the risk.
The mirror question to where your input goes is who owns what you get back — the copyright side is covered in who owns what AI writes.
※ Written as of August 2026, describing general principles. Data-handling policies vary by service and plan and change frequently — check the current privacy policy and terms of service of whatever you're using for exact details. This piece is not legal advice.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.