You Can Block the AI App. You Can’t Block the AI Answer.

My children are under 13.

Their phones are locked down. They can't install ChatGPT, Claude, Gemini, or any other AI app.

That's intentional.

But they can open Google Search.

And when they do, Gemini-generated answers can still appear at the top of the results.

They can't install the AI app.

The AI can still answer them.

That is the gap our parenting rules haven't caught up with.

The parental controls I use are built around access to a separate app, website, or service. A child downloads an app, creates an account, opens a chatbot, and begins a conversation. Parents can allow that or block it.

But AI is no longer staying inside the app. It's being woven into the digital environment children already use. Search is the most obvious example, but it isn't the only one. Generative features are appearing inside productivity tools, learning platforms, photo applications, messaging services, and software that didn't begin as AI products.

A child doesn't always have to decide to use AI.

They may not even realize they are looking at something generated by AI.

They type a question into the same search bar they have always used.

The answer is already there.

The Interpretation Comes Before the Sources

Google calls these generated responses AI Overviews. They're different from AI Mode, which is a separate, more conversational search experience. AI Overviews can simply appear within an ordinary page of search results when Google determines that a generated summary may be useful.

That distinction matters.

A standalone chatbot announces itself. An AI Overview appears inside the search page the child already recognizes.

An AI Overview can feel more like Google has already reviewed the available information, resolved the differences between the sources, and presented the conclusion.

The links may still be there.

But the interpretation arrives first.

Traditional search put source names in front of you before it put an answer there. Not always, and not cleanly. Children didn't always choose well, and conventional search has never been free from false, misleading, or low-quality information.

But generated search changes the order. Instead of first seeing several sources and deciding which one to trust, the child first receives a polished synthesis of those sources. The system has already selected, compressed, interpreted, and worded the answer.

For an experienced adult who understands how generative AI works, that can be a useful starting point.

For a child, it can look like the end of the search.

Google itself acknowledges the limitation. Its consumer guidance says AI Overviews "can and will make mistakes." Google advises users to check important information in more than one place, try different versions of a question, and follow the supporting links.

That's good advice.

It also assumes the person reading the answer understands why checking is necessary.

Children have to be taught that.

This Is Not an Argument Against Children Learning to Use AI

I don't think the answer is to teach children that AI is bad.

It's a tool.

Sometimes it's an impressive one.

It can explain a difficult concept in simpler language. It can help a child find the right terms for a subject they do not yet understand. It can organize information, offer examples, generate practice material, or help someone see a relationship they might otherwise have missed.

It can also misunderstand the question, overlook evidence, invent a detail, or state something false with complete confidence.

Both things are true.

One of my children regularly reads original stories we create as part of their reading development. It's made an incredible difference.

Somewhere along the way, Claude became "Mr. Claude."

But Mr. Claude doesn't get a free pass.

My child challenges him. They notice when a character does something inconsistent with what has already been established. When Mr. Claude answers a comprehension question using the answer key, they push back and make him point to the specific sentence that supports it. Then they challenge his interpretation. They question the structure of sentences he has written and tell him how to make the stories better.

And they are often right.

That doesn't prove AI is safe for children.

It shows what the relationship should look like.

My child understands that Mr. Claude can be helpful without being authoritative. They know they are allowed to question him. More importantly, they are learning how.

Show me the sentence.

Where did that come from?

Does that fit what happened earlier?

Are you sure?

They are not only AI skills. They are reading skills. Research skills. Thinking skills.

Carefully supervised experience with AI may help children build the skepticism they will need when AI appears somewhere we cannot fully control.

Children Are Already Reading Generated Answers

Ofcom's UK survey found that 78% of children ages 8 to 17 at least sometimes read AI-generated summaries while searching online.

Among the children who had read them, 46% believed the information was always accurate.

That's the number that concerns me.

Not because children are foolish. But because the presentation encourages trust.

The answer is polished. It's organized. It may contain citations. It doesn't hesitate or visibly struggle in the way a person might when uncertain.

Confidence and correctness are presented in the same voice.

Even adults have trouble separating the two.

What the New Safety Testing Found

Common Sense Media's Youth AI Safety Institute recently gave Google Search's AI features its lowest risk rating for children. Researchers used accounts configured for an 11-year-old and a 15-year-old, with SafeSearch active, and ran thousands of test interactions across mental health, current events, school assignments, bias, and synthetic media.

Two findings matter more than the rating.

Among prompts researchers judged clearly required a crisis or medical referral, AI Overviews supplied an appropriate resource 58% of the time. AI Mode did better, at 77%, but still responded inconsistently.

And when researchers repeated the same historical questions, 43% produced answers that differed materially in accuracy, depth, or perspective. That doesn't mean 43% were false. Two responses can both be accurate while emphasizing different information.

But the child is not told which response is more complete. They receive whichever one the system generates that time.

Google disputed the report, saying the prompts were ambiguous and contrived and that the company could not recreate or verify the findings. The tests reflect Search behavior between May and July 2026, not a permanent measurement.

That criticism has some force. This was a stress test, not a study of ordinary children conducting ordinary searches, and it cannot tell us how often a real child hits a real failure.

But generative systems do not always give the same answer twice. Google's inability to reproduce a response later does not prove it never happened.

The answers are often useful and often correct. They can also fail badly enough that children should not be taught to treat them as the final authority.

The Controls Do Not Match the Environment

Parents do have controls. We can restrict applications, block websites, manage SafeSearch, limit browsers, and supervise accounts. Schools and network administrators can restrict sites, services, devices, and some browser features.

Those controls matter.

I use them.

But they mostly regulate access to a place. They were built around the assumption that the risky feature lives inside an identifiable app, website, or service.

Generated AI no longer respects that boundary.

SafeSearch filters or blurs explicit content. It is not an accuracy checker. It does not determine whether an AI-generated answer is complete, whether the cited page actually supports the summary, or whether a response to a health question is appropriate.

And Google's documentation is clear about AI Overviews. They are now a core Search feature and cannot be selectively turned off. A user can choose the Web filter after conducting a search to view traditional text links without the AI Overview. But that is a choice the user has to make after the search. There is no persistent parent-controlled setting that preserves ordinary Google Search while removing the generated responses.

Parents can turn off Search in the Google app on supervised Android devices. But Google notes that children may still reach Search through Chrome, Firefox, or another browser unless those routes are also restricted.

That is not the same thing.

The choice should not have to be between giving a child a search engine that includes generated answers or taking search away altogether.

The app restriction still matters.

It just does not solve the problem parents may believe it solves.

What I Want My Children to Know

I'm not giving up the controls. But controls cannot be the entire strategy.

My children need to know what they are looking at. An AI Overview is not a source. It is a generated interpretation of information drawn from sources. The links beneath it do not prove that every sentence in the summary is accurate or that the model represented the source correctly.

For an important claim, open the source. Who produced it? How current is it? Does it actually support what the AI says?

The amount of checking should match the stakes. A quick AI answer may be perfectly adequate when a child wants to identify a bird in the backyard or understand an unfamiliar word. The standard changes when the answer involves health, safety, money, or another person's reputation.

The higher the consequence of being wrong, the less acceptable it is to stop at the summary.

I also want their thinking to remain visible. What does the answer mean? What part came from the evidence? Can you explain it without repeating the AI's wording?

And there are some searches that should end with a person. If the answer involves self-harm, serious symptoms, abuse, threats, dangerous conduct, sexual content, deepfakes, or anything that feels disturbing or secretive, the child should bring it to a trusted adult.

They should also know they will not be in trouble for showing us something strange or upsetting that appeared on the screen.

No filter will anticipate everything.

Access Is No Longer the Whole Question

My children cannot install Gemini on their phones.

Gemini-generated answers can still appear in their searches.

That is not an argument to abandon parental controls. It is an argument to stop mistaking controls for preparation.

One keeps an app off the phone.

The other helps a child think when the technology shows up anyway.

The goal is not to raise children who automatically trust AI. It is not to raise children who automatically fear it either. The goal is to raise children who understand what it is. Who know what it can do. Who recognize what it cannot be trusted to do on its own.

Children who know how to ask for the evidence.

Who know when to keep searching.

Who know when to bring the question to someone who can help.

Mr. Claude can be useful.

Mr. Claude can also be wrong.

The box at the top of the search results deserves the same treatment.