Host: So, we're talking about a document today that opens with a pretty bold claim: 'We don't build models.' And while most tech companies would just stop there and call it a day on privacy, this report argues that stopping there is actually a bit of a half-truth.

Listener: Wait, if they aren't building models, then what exactly is myOrbit? I thought the whole point was this 'AI twin' that learns who I am. How does it learn if there’s no model being trained?

Host: That's exactly the distinction they want to make. They use what they call 'orchestration.' Basically, your twin, Aura, isn't a single model they've built in a lab. Instead, it draws on the heavy hitters—Anthropic, OpenAI, and Google—and picks the best one for whatever you’re asking in that moment.

Listener: Okay, so my words are being sent to these big labs. But isn't that just moving the problem? If I'm talking to OpenAI through them, isn't OpenAI just training on my data instead?

Host: This is where the document gets refreshingly honest. They actually walked back an earlier, stronger promise they made. They used to say these labs were contractually barred from training on your content, full stop. But now they’re saying: look, these labs have their own terms, and those terms can change.

Listener: So they can't actually guarantee what the labs do with my data?

Host: They can't make a blanket, forever-promise for third-party giants. They configure the services for privacy, but they want you to know that these providers might retain data for things like safety or legal reasons. It's a 'weaker' claim, but they argue it's the only one that’s actually honest.

Listener: I appreciate the honesty, but let's go back to my twin. If it's not 'training,' how does it remember me? If I tell it I hate cilantro on Tuesday, does it remember that on Wednesday?

Host: It does, and that’s the key. They draw a hard line between 'training' and 'personalization.' Training is when a company takes everyone’s data and bakes it into a general model that serves strangers. Personalization is holding your context specifically for you.

Listener: So one makes a product *out of* me, and the other makes a product *for* me?

Host: Exactly. Your twin learns you so it doesn't have to start from scratch every morning, but that data stays attached to your account. It doesn't make the underlying AI smarter for anyone else.

Listener: What if I decide I'm done? Can I just hit a 'delete' button and have all that 'learning' vanish?

Host: Yes and no. You can delete your conversation history and your twin's memories—those are separate things in their policy—but they’re very clear that they can't 'untrain' a general AI model. No one can. It’s technically impossible right now. So they focus on what they *can* control: deleting the data they store on their end.

Listener: And what about the 'asterisk'? You know there’s always a catch in the fine print.

Host: You're right, there is one. They mention in their terms that 'de-identified' data might be used for product improvements—like aggregate signals on how the app is being used. It’s not your name attached to a conversation, but it’s not nothing either, which is why they include it.

Listener: And the security? Is this all encrypted so they can't see it either?

Host: This is another place where they refuse to use marketing fluff. They use standard encryption for storage and transit, but they admit they *could* technically read it. There’s no 'zero-access' or 'end-to-end' encryption here. They say the keys are theirs, which protects you from outside hackers, but not from the company itself.

Listener: It sounds like they're going out of their way to tell me exactly where the walls are, rather than pretending the room is infinite.

Host: That’s the whole philosophy. They believe the imprecise version—'we don't train on your data'—fails the moment you notice your twin actually remembers you. By being precise, the story holds up even when you look closely at the privacy policy or the deletion routes. It’s worth checking out the full document for the technical details on encryption and the specific deletion schedules they follow.