Is a digital twin safe?

Is a digital twin safe? It is the right thing to ask before you put your words on a page that talks to people you may never meet. Nothing online is free of risk, and we won’t pretend otherwise. What we can do is tell you exactly what a twin knows, who can read what, how deleting works and what can still go wrong. Then you can decide.

What it knows

Only what you typed. Your ten answers, any of the forty optional ones you add, and the first part of each section on your About page. It doesn’t look at the web, your social accounts, your contacts or your phone. The links on your About page never reach the AI. If something isn’t on your page, your twin doesn’t know it, and it says so.

The first thing to be careful about: your answers are public. Anyone with your link can ask about them, and under an answer they can read your exact words. So write what you would say to a stranger, and leave out phone numbers and email addresses. A person’s twin never gives those out.

Nothing is trained on your words, by us or by anyone we work with. Answers are written by a third-party AI model provider under contract with us, whose servers are in the United States. For each question it gets your twin’s page and the visitor’s messages, writes the answer and hands it back. It doesn’t keep your page. This post follows one question all the way through.

Who can read what

You can read every chat your twin has, for 30 days.

Visitors are told that before they write a word. A visitor who isn’t signed in shows up as “Visitor”: no name, no account, no address. A signed-in visitor shows up by name, unless they switch that chat to anonymous first.

We don’t read chats as part of our day. We open a conversation when it is reported, when a bug can’t be found any other way, or when the law requires it.

Email addresses, phone numbers and social handles are removed from a visitor’s message before it is stored and before it reaches the AI, and their IP address isn’t stored with the chat. There are no ads and no trackers, and we don’t sell data.

A twin can also keep one short note about a visitor, but only if they asked or you are connected. That person can read it word for word and delete it.

So here is one plain warning for visitors: a twin isn’t a private diary. Don’t tell it anything you wouldn’t tell its owner.

Deleting

Change or remove an answer and your twin changes with it. You can pause your twin, and you can export everything to a file.

Delete your account and your answers, your page, your photos, your connections and the notes your twin kept are gone, and because nothing was trained, nothing of yours is left inside the AI. There is no undo. Backups age out within thirty days. We keep only your old username on a list, so nobody else can take it.

How it stays honest about being an AI

A digital twin that lets people think it is human isn’t safe, however good it sounds. Before the first message, every twin page says this, here for a made-up person called Deniz: “This is an AI built from what Deniz wrote. It is not Deniz, and it can be wrong.” That line can’t be switched off. In the chat, “I” means the AI, and it talks about you by name. If someone asks whether it is really you, it says it is an AI.

It can be wrong. It puts your words into a new sentence, and a new sentence can drift. That is why any visitor can open “Where does this come from?” and read the lines it used. If it used none, the page says so.

It can’t commit you to anything, either. It makes no promises about money, dates or plans, and it doesn’t rate or rank anyone.

What can still go wrong

Someone might make a twin pretending to be you. That is against the rules. Use the Report link at the bottom of any twin, the report page, or write to hello@itwin.co. You don’t need an account, and a twin that pretends to be someone else comes down.

Someone might try to trick your twin into saying something you never wrote. It is told to ignore those attempts and stay with your page, but no AI is perfect. That is one more reason to keep your page to things you would say in public.

You need to be 18 to make a twin, for now. Anyone can talk to one.

So, is a digital twin safe?

It depends on what it is built on. One that reads only what you wrote, says what it is, lets you read everything and really deletes is about as safe as any public page you control. The risk that remains is mostly the one you already know: what you put on a public page is public.

The privacy and security pages have the full detail, and the questions page has the short version. If you are happy with all that, start here, and write only what you would stand behind.