Why We Stopped Calling Nova a 'Friend'

This post is published as an honest account of a product decision. It is CC BY 4.0; please attribute when quoting.

What this is

In late September 2026, we changed one piece of wording in Xyplor for kids 10 and under. Nova — the AI helper that coaches kids through building their projects — used to be introduced in-app as "a friend made of code." We changed that to "an AI helper." We also updated the language on xyplor.com/safety to state explicitly: Nova is a helper, not a companion. It doesn't role-play as a friend or family member, and if a kid brings it something heavier than a stuck project, it routes them to a trusted grown-up instead of trying to be one.

That's a small wording change on paper. We think it matters more than it looks, and we want to walk through why — including the part where we had it wrong first.

What we had wrong

"A friend made of code" was written to be warm. It was meant to make a 7-year-old feel comfortable asking Nova for help instead of feeling like they were talking to a machine that might judge a bad first draft. As a tone-setting phrase for a kids' product, it did that job fine.

But "friend" is not a neutral word, and we didn't think hard enough about what it implies when it's said to a child by an AI system. A friend is someone you tell things to. A friend remembers you, takes your side, and is there when you're upset. None of that is what we built, and none of it is what we want a 6-year-old to expect from a chat box — no matter how good our intentions were when we wrote the copy.

Nova's actual job is narrower than that: help a kid describe what they want to build, explain what the AI just did, suggest a next step, and answer the questions kids are curious about. It is explicitly built not to be a confidant. Calling it a friend set an expectation the product was never designed to meet, and the gap between the label and the behavior is exactly the kind of gap that matters most with young kids, who are still learning to tell the difference between a responsive program and a person who cares about them.

What changed, specifically

As of 2026-09-29, the in-app introduction for kids 10 and under says Nova is an AI helper — not a friend, not a buddy, not a companion. On the safety page, we state plainly that Nova won't role-play as a friend or family member. Its system prompt is built around that boundary: if a kid brings it something that sounds like real distress, Nova acknowledges it briefly and then encourages them to talk to a parent or another trusted adult, rather than trying to be the one who comforts them. This sits alongside the input filter that already does pattern-matching for unsafe content in every kid AI conversation — the wording change didn't add a new safety mechanism, it made the existing one's framing match its actual behavior.

We're not claiming this change altered how any specific kid uses Nova. We don't have before-and-after data on that, and we're not going to pretend we do. What we can say is that the product now describes itself accurately to the kids using it, which it didn't before.

Why the word carries more weight than it seems to

Two things pushed us to make this change, and both are worth parents' attention even outside our product.

Common Sense Media's age-and-stage guidance. Common Sense Media publishes guidance on AI chatbots and kids, by age band. For children 5 and under, the recommendation is no AI chatbots at all, because young kids can't reliably tell AI apart from a real person. For ages 6 to 12, the guidance is limited use, alongside a parent or caregiver. And across every age group under 18, the guidance is explicit: "No AI for emotional support, mental health advice, or 'friendship.'" That's not aimed specifically at us — it's general guidance about AI chatbots, not a review of any particular product, and Common Sense Media hasn't evaluated Xyplor. But Nova is a chat-style AI interface that a 7-year-old talks to regularly, and "friendship" is precisely the frame we'd put in the product's own mouth. Once we read that guidance next to our own copy, the mismatch was hard to ignore.

What investors are starting to ask. At an Edtech Insiders webinar on AI governance in K-12 in late September, Amit Patel of Owl Ventures described what investors now ask founders building for elementary and early-learning kids: whether the company has consulted safety experts, and specifically what guardrails prevent young children from forming emotional attachments to the product. That's a sign of where the category's scrutiny is heading — not just "is the content filtered," but "does the product's own framing invite a kind of attachment a 6-year-old isn't equipped to have with software." We'd rather answer that question before we're asked it in a diligence call than after.

Neither of these sources is about Xyplor specifically, and we're not citing them as validation of anything we've built. We're citing them because they're the two clearest external signals we found that the industry is converging on the same concern we'd started to have internally: the words a product uses to describe its own AI matter to a young child's mental model of what that AI is.

What this doesn't solve

Changing a label doesn't change what the underlying model says in a long, winding conversation. Language models are good at picking up on tone and matching it; that's true of every AI product a kid uses, and it's why a label can't be the whole answer. The system prompt boundary and the input filter are the actual backstop here — the wording change is the part that sets the kid's expectation correctly from the start, not the part that enforces it. We think both matter, but they're doing different jobs, and we don't want to overstate what a label alone can do.

We also haven't measured whether this change affects how kids actually behave in the product — whether they ask Nova fewer personal questions, or none at all, or exactly the same as before. We don't have that data, and if we get it, we'll report it honestly rather than assume the change worked just because it was the right thing to do.

What this means if you're evaluating any AI product for a kid

If your child uses an AI chatbot, tutor, or helper — ours or anyone else's — it's worth actually reading how the product introduces itself to the kid, not just what the safety page says in the fine print. Does the onboarding copy call the AI a friend, buddy, or companion? Does the product's own language invite the kind of relationship Common Sense Media's guidance says kids under 18 shouldn't have with an AI? That's a five-minute check, and it tells you something the privacy policy won't.

It's also worth asking the question investors are now asking founders: what happens when a kid brings something heavier than a task to the AI? Does it try to comfort them in character, or does it route them to a person? That one behavior — the redirect — is, in our view, more load-bearing than any label.

We got the label wrong for a while. Fixing it didn't require new engineering, just the discipline to notice that the word we'd chosen to make Nova feel warm was also the word that could make a 7-year-old believe something about it that wasn't true.