Nothing is trained: how an AI twin answers

The word people reach for is “trained”. As in: so you train it on me? No. Nothing is trained. Not a model, not a voice, not a personality. It is worth being exact about this, because the alternative sounds impressive and is mostly worse.

What training would mean

Fine-tuning takes a large model and nudges its weights with your data until it starts producing text that sounds like you. It is slow, and once your data is inside the weights it cannot be pulled back out. Delete your account and your sentences are still in there somewhere, smeared across a few million numbers. A voice clone does the same with audio. A “personality model” is a fancier name for the same bet: that you can be compressed into parameters and played back.

None of that happens here. There is one general AI, and it has never seen your answers until the moment somebody asks a question.

What a twin actually is

A twin is a page. Ten answers you wrote, any extras you added, and an About section if you filled one in. For a business, the items it lists too. Plain text, stored in a database, editable, deletable.

When a visitor sends a message, the AI is handed that page along with a short set of instructions. Answer only from what is on the page. Refer to the person by name. “I” means the AI. Do not guess. Do not rate. If the page doesn’t cover it, say so. Then the AI reads and answers, the way a careful assistant would answer from a briefing note.

Anything it says about you can show the lines it came from. A plain hello has no source, and the page says so when an answer cited nothing. When someone asks about you and there is no line to lean on, that is the signal to stop.

One thing can sit beside the page. If a visitor asked to be remembered, or the two of you are connected, the AI also gets one short note about that visitor: what they told it, what they asked, what was left open. The note is about them, never about you. It trains nothing, and they can read it and delete it. Here is how that memory works.

The path one question takes

Say a visitor types: “Has Deniz ever built anything for her friends?” Deniz is one of our made-up demo people.

First, the message is scrubbed. Any email address, phone number or social handle in it is removed before the text is stored or goes anywhere near the AI. The visitor’s IP address is not stored with it. What is kept is the message and the answer, for thirty days, and Deniz can read them; the visitor was told that before typing a word. The back-and-forth of this conversation is kept for two hours, so the next reply makes sense.

Second, the AI gets the page. Deniz’s ten answers, her extras, her About section, with the instructions on top.

Third, it reads. Her answer about a project she is proud of says a friend’s student club was tracking borrowed equipment in a group chat, and then: “I built them a tiny web app over a long weekend.” That covers it.

Fourth, it answers from there and only from there. “Yes. A friend’s student club was tracking borrowed equipment in a group chat, so Deniz built them a tiny web app over a long weekend.” Nothing added about how many people use it, what it cost or who asked. She didn’t write that.

Fifth, it shows the source. The visitor can open “Where does this come from?” and read the lines the answer leaned on, in Deniz’s own words. If the paraphrase drifted, the original is right there.

The moment it says “I don’t know”

Now the visitor asks: “What does she think about electric cars?”

The AI reads the page again. The club app is about borrowed equipment. Nothing about cars, nothing about opinions on them. So it stops.

“Deniz didn’t write about that, so I don’t know. Ask Deniz.” It doesn’t offer a phone number or an email, because a person’s twin never hands those out. The question goes on Deniz’s list of things her twin was asked and couldn’t answer, so she can write about it if she wants to.

This is the part that feels like a failure and isn’t. A fine-tuned model would have produced two confident paragraphs about electric cars, because confident paragraphs are what it is for. The twin produces one honest sentence, because it is only allowed to speak from the page, and the page is silent.

Why do it this way? Because it keeps you in charge. Change an answer and the twin changes with it, immediately. Delete your account and your answers, your page and the notes your twin kept go with it, because nothing was ever absorbed. And because a visitor can check. An answer that points at its source is an answer you can argue with, which is the only kind worth trusting.

Everything here is built on this and nothing else. If you want to hear what a page of your own answers sounds like, you can make your own twin now.