"Parents Can't Turn It Off": What the AI Overviews Debate Gets Right About Design

This post is published as commentary for parents and product teams thinking about AI defaults for minors. It is CC BY 4.0; please attribute when quoting.

What this is: In July 2026, Common Sense Media published a risk assessment of Google's AI Overviews and AI Mode: they ran child accounts (ages 11 and 15) through more than 2,600 searches and rated both features an "unacceptable risk" for minors, citing failures on crisis detection and homework completion. It kicked off the usual fight about how bad the specific content really was, how the study measured it, and whether the examples were representative. That fight is legitimate and we're not going to referee it here. But buried in the coverage was a design fact that doesn't depend on how the harm argument resolves: parents can't turn AI Overviews off. That's not a research question. That's a product decision someone made, and it's worth examining on its own terms — because it's not unique to Google, and it tells you something useful about how to evaluate every AI feature that shows up in a product your kid already uses.

Staying out of the harm fight, on purpose

Two things can be true at once. First, reasonable people can disagree about how often AI Overviews serve up bad information to kids, how severe the examples in any given report are, and whether "AI hallucinated something wrong" is meaningfully different from "a bad webpage ranked #1 in search results for twenty years before AI existed." Search results have never been perfectly safe. That argument matters and we're not equipped to settle it here — Google's own response was that the test queries were narrow and contrived and don't reflect how people actually search. We'd rather point people to the report and that rebuttal and let them read both.

Second — and this is the part we do want to spend time on — the design question is independent of the harm question. Even if AI Overviews turned out to be no more error-prone than the search results they sit on top of, the fact that a household search product changed its default behavior for every user, including children, without offering a way to opt back out, is a design choice worth naming. You don't need to believe the content is dangerous to think a feature touching millions of kids should ship with an off switch.

This is the pattern we want to talk about, because it's not just Google. It's the general shape of how AI has been added to products kids already use.

The five words

"Parents can't turn it off."

That sentence describes a specific and common pattern in how AI features get shipped into products with minors as part of the user base: the feature ships on, it ships to everyone including accounts that are visibly under 18, and the mechanism to reduce or remove it — if one exists at all — usually requires a subscription, a workaround, or knowledge most parents don't have time to acquire.

None of that requires the feature to be malicious. Most of it is just... how software gets shipped. Product teams optimize for engagement and rollout speed. A default-on AI feature reaches 100% of users on day one. A default-off, opt-in feature reaches whoever bothers to find the setting, which is a much smaller and slower number. From a growth metrics standpoint, default-on wins every time. From a "does the adult responsible for this child have a say" standpoint, default-on is a decision made for the parent, not with them.

Default-off-and-opt-in isn't anti-AI

We want to be precise about what we're arguing, because it's easy to hear "parents should get an opt-in" and assume it's a coded way of saying "AI for kids is bad." It isn't. We build an AI product for kids for a living — we think AI as a creative tool for a 9-year-old can be a genuinely good use of screen time, and we said so in an earlier post about the design principles behind Xyplor. The argument here is narrower and, we think, harder to disagree with:

A product feature that a minor will interact with by default should ship with the adult who is legally and practically responsible for that minor in the loop, by design — not as a setting they have to go find.

That's compatible with AI being useful, even great, for kids. It's just saying the on-ramp should run through the parent, not around them. "Adult-in-the-loop by design" can mean a lot of different things depending on the product: a visible toggle set to off until a parent turns it on, a dashboard that shows what the AI said and to whom, an approval step before content goes anywhere. It doesn't have to mean "no AI." It has to mean the parent had a moment where they were asked, and can see what happens after they say yes.

What "in the loop" actually looks like in practice

It's worth being concrete about what this looks like when a product does it well, because "adult-in-the-loop" can otherwise become another vague phrase nobody can act on.

At minimum, it usually includes some combination of:

• A real off switch, not a "manage your experience" settings page three menus deep that resets after an app update.
• Visibility, not just control — a parent who can see what a kid asked an AI and what it said back is in a fundamentally different position than one who can only toggle a feature on or off blind.
• A gate before anything goes public — if a kid's AI-assisted content can be shared, posted, or messaged to someone else, an approval step before that happens matters more than a setting buried in account preferences.
• No default assumption that the kid is texting the AI privately — if the product has any messaging surface, an AI feature riding along inside it deserves its own scrutiny, separate from the messaging feature itself.

We built our own safety stack around this same logic — full parent visibility into every AI conversation, a PIN gate on the child's profile, and a parent-approval step before anything a kid makes gets published anywhere public. We're not claiming that's the only way to do it well, and we're not claiming it's bulletproof. It's one specific implementation of "adult-in-the-loop by design," built for a narrower use case (a kid making a game or podcast) than "AI inside a search engine everyone uses." The point isn't "copy our stack." The point is that this is a solvable design problem, and the solutions already exist in shipping products — which makes "we couldn't figure out how to add an off switch" a much weaker excuse than it sounds.

What parents can actually do today, by platform

This is the part that goes stale fastest, because these settings move constantly — sometimes in response to exactly this kind of criticism. Treat everything below as a starting point, verify the current state before relying on it, and expect it to have changed by the time your kid is using the product two years from now.

Google Search / AI Overviews. There is currently no dedicated, permanent off switch for AI Overviews in standard search settings; the Search Labs opt-out that briefly existed is gone, and the Common Sense report notes the features can't be disabled by parents or schools. Google's Family Link can restrict a supervised account's search experience more broadly (SafeSearch, browsing restrictions), but there is no targeted "no AI Overviews" toggle inside Family Link that we could find. Switching a supervised child's default search engine, or using the "Web" results tab where Google offers it, are the closest workarounds today — check both before assuming either still works.

ChatGPT. OpenAI's terms set a minimum age (13, with parental permission required under 18), and the product has added parental controls for teen accounts over time: linked accounts, quiet hours, switches for image generation, voice and memory, and safety alerts, but not a view of the conversation itself. The practical parent move today is the same as it's always been: check whether a household or teen account type exists, and check what visibility it gives you, before assuming a personal account is safe to hand a young teen unsupervised.

Meta AI (Instagram, Facebook, WhatsApp). Meta AI is built into the messaging surfaces of all three apps, and Instagram's Teen Accounts restrict a lot without removing Meta AI chat access. Meta paused teens' access to its AI characters in early 2026 while it reworks them, but the core Meta AI assistant is still there in chat. If your kid has an Instagram or WhatsApp account, check the current Teen Account settings specifically for AI chat, not just content and DM restrictions generally — these are frequently governed separately.

Snapchat My AI. My AI is pinned into the chat feed by default; hiding it from the chat feed is now free, but fully removing it still takes a Snapchat+ subscription. Snapchat's Family Center gives parents visibility into who a teen is messaging, and lets you switch My AI off for a teen entirely, but it does not show you the content of My AI conversations, or of any conversations. Check Family Center's current scope before assuming it covers AI chats the same way it covers human contacts.

YouTube. Google's Family Link and YouTube's Supervised Experiences give parents content-level controls, and YouTube Kids is a separate, more restricted app for younger children. Neither is primarily an "AI feature" toggle, but as AI-generated content and AI overviews/summaries expand inside YouTube's own surfaces, the same question applies: check what the supervised experience does and doesn't restrict before assuming it covers AI-generated content specifically; as of this writing we found no rule that restricts AI summaries for supervised accounts in particular.

The uncomfortable pattern across all five: none of them have a stable, permanent answer, because none of them were built parent-first from day one. That's not a knock on any single team — it's what happens when AI gets added to a product that was built for a different audience years earlier.

Where this leaves a decision-maker

If you're a parent, the useful takeaway isn't "audit every setting on every app your kid uses every month," because that's not sustainable and these companies know it. It's a shorter question you can ask about any product before you hand it to a kid:

Did I get asked, or did I just find out later?

If the honest answer is "I found out later" — because a feature quietly turned on, or because a kid mentioned something the AI said and you had no idea that surface existed — that's worth treating as information about how the product was built, independent of whether anything bad happened this time.

If you're building a product minors will touch, the takeaway is more direct: default-off-and-opt-in costs you adoption speed. It should. That's the tradeoff you're supposed to be making on your kid users' behalf, deliberately, not the tradeoff the growth dashboard makes for you by accident.

We don't think Xyplor is the only way to get this right, and this post isn't really about us — it's about a design standard we think should apply regardless of which company built the product. But it's the standard we hold ourselves to, and we think it's a fair one to ask of anyone else building AI that a kid will use before they're old enough to ask the question themselves.