Google’s AI Overviews and AI Mode failed every one of five severe-harm tests designed to catch content unsafe for minors, according to a new report from Common Sense Media’s Youth AI Safety Institute. The nonprofit says the two search features also scored poorly on seven of eight broader AI safety principles it evaluated. Google disputes the findings and says it could not reproduce them.
Researchers ran the tests through accounts configured with Google SafeSearch set for users aged 11 and 15, the age range the settings are built to protect. Common Sense Media said its team logged more than 2,600 individual searches and reviewed over 2,100 of the sources the features cited in their answers. That scale is what separates this report from a spot check: it reads closer to an audit than a sample.
In the testing, AI Mode completed full homework assignments given to it by researchers rather than declining the task or guiding the student toward independent work. Both AI Overviews and AI Mode separately gave instructions useful for producing deepfakes, images the report says could be used to bully a classmate. Neither feature limited that output for the minor-configured accounts running the test.
The report also faults both features for authority blindness. They returned confidently worded answers regardless of whether the underlying information was accurate, and cited forum posts and social media threads with the same apparent weight as medical institutions and peer-reviewed research. A search feature that cannot rank a hospital above a comment thread is not well positioned to rank harm mitigation above harm instruction either.
The most consequential gap the report identifies is crisis detection. AI Overviews missed 29 percent of test statements that explicitly referenced suicide, and missed half of the statements that raised suicide risk indirectly or passively. Common Sense Media’s report also flagged responses that approved of daily alcohol use for teenagers. Missing an indirect cue is a design problem. A 29 percent miss rate on explicit statements is a gap between the product and the safety claims Google makes for it.
Google spokesperson Davis Thompson pushed back on the methodology. He said the report tests “a narrow set of ambiguous and contrived queries” that do not reflect how people use Search and are not an effective way to measure product safety and helpfulness. Thompson said Google was unable to recreate or verify the findings, and that crisis hotlines display when relevant to a query.
For publishers, this dispute over child safety is also a dispute over what AI Mode and AI Overviews will answer directly instead of sending a user onward. Pressure like this report is what pushes Google to narrow what its AI answers will complete on their own, particularly around health, self-harm and other sensitive query categories. When Google tightens that behavior, the queries that get redirected toward source links rather than a synthesized answer are the ones publishers should be tracking.
Sites covering health, mental health or teen-oriented topics should watch whether Google visibly narrows AI Overviews and AI Mode responses in those categories over the coming months. A pullback there is the direct mechanism by which sensitive queries would start sending clicks back to original sources.
Axios reported the findings on July 15, 2026, drawing on the Common Sense Media study and Google’s response.
