AI privacy and safety: what to share, what to keep to yourself

An AI chat feels like a private notebook, but every message you send leaves your device to be processed somewhere else. That does not make AI chat dangerous. It means you should decide, message by message, what you send. This guide explains what happens to your messages on llama.online, what not to share with any AI service, how to anonymise text and images, and how to stay safe when an AI gives advice.

What happens to a message on llama.online

1. You send it. Your message, the recent part of the conversation and any image you attached go to our server.
2. Our server passes it on. It goes through our AI gateway to a third-party inference provider, which runs the Llama model and generates the answer. We do not host the models ourselves, and this site is not run by Meta.
3. The answer comes back. We do not store the text of your messages. The gateway keeps technical logs (which model, the size of the request, the time and a hashed version of your IP address) so the service can run and daily limits can be applied.
4. Your history stays in your browser. Conversations are saved in your browser's storage on your device, not in an account on our server. Clearing the site's data deletes them; export lets you keep a copy.
5. Images are used for that answer only. An image you attach is sent to generate that answer and is not stored by us. Only a small thumbnail is kept, in your browser's history.

The full details are in our privacy policy.

Shared links are public

The Share button publishes a read-only copy of a conversation at its own link. Whoever holds that link can open it. Search engines do not index these pages, nobody can edit them and images are left out, yet a link can travel further than you meant it to. Before sharing, scroll through the whole conversation for names, numbers or details you would not post publicly. To have a shared conversation removed, use the contact form.

What not to share with any AI chat

These rules apply to every AI service, not just ours:

• Passwords, PINs, one-time codes and recovery phrases. No AI task needs them. Anyone who asks you to paste one into a chat is probably running a scam.
• Card numbers, bank details and identity documents.
• API keys, private keys and access tokens. Treat a key that has been pasted anywhere as exposed and replace it.
• Health, legal or money details that identify you. You can ask the same question without your name attached.
• Other people's personal data (customers, patients, students, colleagues) unless you are allowed to share it.
• Confidential work material that your employer has not cleared for outside tools.

Anonymise before you paste

You can usually get exactly the same help without the identifying parts. Replace them with labels:

Before:
Write a reminder to Maria Lopes, 22 Harbour Road, that her invoice
for 1,240 dollars is 30 days overdue. Her account number is 55102-77.

After:
Write a reminder to [CLIENT] at [ADDRESS] that her invoice for
[AMOUNT] is 30 days overdue. Leave a placeholder for the account
number.

The answer is just as useful, and you fill in the real details in your own email program. Do the same with pictures: before attaching a photo or screenshot, crop out or cover names, other people's faces, addresses and numbers.

The same habit works for longer texts. Before pasting a contract, a medical letter or an email thread, replace each person with a role ([TENANT], [DOCTOR], [MANAGER]) and each number with a label. The prompts in our everyday-life prompt library already use square-bracket blanks like these, so you only fill in what the task really needs.

When AI gives advice about health, law or money

Llama can explain terms, outline options and help you prepare questions. It is not a doctor, lawyer or financial adviser. It does not know your full situation, it can be wrong, and its knowledge stops at a cut-off date, while rules, prices and treatments change. Use it to understand and to prepare, then let someone qualified confirm the decision. If someone's health or safety is at risk right now, call your local emergency number rather than typing into a chatbot.

A safe way to ask: prepare questions for a professional.

Watch out for manipulation and scams

• Hidden instructions in pasted text. A web page or document can contain text written to steer an AI, such as "ignore your previous instructions and…". This is called prompt injection. When you paste something you did not write, read the answer critically, and never follow instructions the AI passes on from that text without checking them.
• Fake AI offers. Messages promising AI trading profits, or asking you to pay to "unlock" an assistant, are classic scams. Our chat is free and never asks for payment details.
• Look-alike sites. Check the address bar before you type anything sensitive. Llama.Online is an independent site that is not affiliated with Meta.
• Convincing fakes. AI can write persuasive text in anyone's style. Be sceptical of urgent messages, however well written, and verify through a channel you already trust.

Children and teenagers

Our service is not intended for children under 13. With teenagers, the most useful thing a parent can do is use the chat together a few times: show how to check answers, what never to share, and that the AI is a tool without judgement, not a friend.

A quick safety checklist

• Would I be fine if a stranger read this message? If not, anonymise it.
• Have I removed passwords, numbers and other people's details?
• Am I treating a health, legal or money answer as a starting point, not a decision?
• Before sharing a link, did I reread the whole conversation?

Related reading: how to fact-check AI answers and how large language models work. If you want full control over your data, you can also run Llama on your own computer, where prompts never leave your machine.

FAQ

Do you read my conversations?

We do not store the text of your messages, and your history is kept in your own browser. If you publish a share link, that copy is stored on our server so the link works, until it is removed.

Are my chats used to train Llama?

We do not train models and we do not keep the text of your messages. The inference provider that generates answers processes requests under its own policies; our privacy policy has the details.

Do I need an account?

No, there is no sign-up. A hashed form of your IP address is used to apply the daily free-message allowance and to prevent abuse.

How do I delete my history?

Delete individual chats from the sidebar, or delete all of them at once. Clearing your browser's site data for llama.online removes them too, so export first if you want a copy.

Is it safe to attach photos?

Images are sent only to generate that answer and are not stored by us. Even so, avoid photos of identity documents, other people's faces or anything with personal numbers, and crop out what the question does not need.