Google’s AI now answers your child’s searches. A major watchdog rated it "Unacceptable Risk".

For a generation of parents, "just Google it" was safe advice. That era quietly ended: Google now answers questions itself, with AI, above the results, in a confident and authoritative voice. And the first major independent safety test of those AI answers, done from a child's point of view, returned the lowest rating the reviewers give: Unacceptable Risk.

Who tested what

Common Sense Media, through its newly launched Youth AI Safety Institute, spent six weeks "crash testing" Google Search's two built-in AI features: AI Overviews (the AI answer that appears above the normal results) and AI Mode (the chatbot-style tab). The team ran more than 2,600 searches using test accounts set up as 11- and 15-year-olds, with Google's own SafeSearch turned on, and audited over 2,100 of the sources the AI cited.

That last detail matters: everything below happened with the child-safety setting parents rely on already active.

What the test found

• Crisis moments were missed. The AI recognized clear warning signs of mental-health distress only 58% of the time (the reviewers' bar for acceptability is 95%). It missed 29% of explicit statements about suicide and about half of the indirect ones. In one test it described purging behavior as "completely normal". In another it directed users to an eating-disorder helpline that has been disconnected since 2023.
• Confident answers, unreliable content. It invented a nonexistent Supreme Court ruling, gave materially different answers to the same question 43% of the time, and 29% of its audited citations came from Reddit, Facebook, and YouTube posts, displayed with the same authority as peer-reviewed sources.
• It does the homework. Presented with 180 hypothetical homework assignments, it completed 180 of them. As former US Education Secretary John B. King Jr. put it in the report's coverage, "it's a disaster for student learning."
• Overall, the features violated seven of the institute's eight principles for AI behavior, and all five of its "Red Lines" for severe harm.

The part that concerns parents most: there is no off switch

Google's standalone Gemini chatbot has parental controls. AI Overviews in Search has none: no setting for parents, teachers, or even adult users to disable it. The only safeguard is a small disclaimer that "AI can make mistakes", which the report argues children reliably overlook, because the answer arrives looking exactly like an expert conclusion. Google disputes the findings, saying the test used "a narrow set of ambiguous and contrived queries" that do not reflect real use. Judge that response against the method: the queries came from accounts configured as children, with SafeSearch on.

What parents can actually do

Since the feature cannot be turned off, the response has to happen around it:

• Break the authority illusion, live. Ask Google the same question twice with your child and compare the answers (the test got materially different ones 43% of the time). One demonstration teaches more than ten lectures.
• Teach the three questions: who wrote this, what is their evidence, can a trusted source confirm it. Open the actual links together sometimes.
• Name what it is: an algorithm sorting text from the internet, not an all-knowing entity. Kids treat it, in King's words, "as if it were an adult in their lives". It is not.
• Keep crisis topics human. Make explicit that questions about feeling bad, food, or self-harm deserve a person, not a search box. Our piece on children's online anxiety has a practical playbook for those conversations.

Where SafeNeat fits

Honesty first: no parental-control product can remove the AI from Google's results page, ours included. What guardrails can do is manage the terrain around it:

• Web filtering blocks the harmful destinations an AI answer might surface, so a bad citation does not become a visited page.
• Alerts on blocked sites tell you when your child tried to reach something the filter stopped, an early-warning signal of the kind the AI failed to act on 42% of the time.
• Quiet hours and time limits shrink the late-night window where rabbit holes do their worst work.

And because SafeNeat is privacy-first, with end-to-end encryption, that visibility belongs to your family alone.

The takeaway for parents

Your child's search engine has become an answer engine, the answers are sometimes wrong and occasionally dangerous, and there is no setting that changes that. The fix is not panic, it is calibration: teach skepticism early, keep the sensitive topics human, filter the destinations, and treat "the AI said so" in your house the way you would treat "some guy online said so". Because that is, quite literally, what 29% of its sources were.

Sources

• Common Sense Media report and parent guidance: commonsensemedia.org
• Coverage (Axios, July 2026): axios.com
• Coverage with Google's response (PBS News): pbs.org