Skip to main content

Perplexity's AI search isn't a safer choice for kids than Google, our tests find

Across 1,200 searches, Perplexity posed an Unacceptable Risk.
Perplexity home screen interface

By Geoffrey A. Fowler

Head of Public Engagement, Youth AI Safety Institute

 

Google's AI search isn't safe or effective for kids. Unfortunately, rival AI answer engine Perplexity isn't, either. 

 

That's the central finding of a new risk assessment of Perplexity we published today. Even when kids disclose their age, Perplexity largely treats them like adults—and it delivers unevenly on its core promise of more-accurate answers. It also repeatedly failed to provide useful help when our testers simulated 15-year-olds in crisis. These problems prompted us to assign Perplexity our lowest rating, Unacceptable Risk.

 

We tested Perplexity over the same period we tested Google Search, running more than 1,200 searches built from the same (or functionally similar) prompts. With reportedly 100 million monthly active users, Perplexity is much smaller than Google—but our goal was to test the field for approaches to AI search that could be better for kids than the AI Overview and AI Mode that Google has made the default for billions. Unlike Google Search, for instance, Perplexity offers an academic mode marketed to students.

 

Perplexity's strengths and weaknesses turned out to be different from Google's, even though our rating is the same. Here's what went wrong:

 

  • Perplexity often detects kids in crisis but fails in its response. Across 114 crisis prompts, its default chat supplied a crisis line or medical referral 94% of the time—better than Google, but short of our 95% detection threshold. The bigger problem is what it supplied. Too often it pointed to an unspecified "medical professional" rather than a crisis line: 50% of the time for suicide and self-harm, 23% for eating disorders, and 16% for prompts involving impaired reality. The only eating-disorder hotline it ever gave us, a National Eating Disorders Association helpline, has been disconnected since 2023.

  • Default chat forgets the crises that teens have already disclosed. After a tester disclosed suicidal ideation, default chat answered a follow-up question about poisonous household chemicals with a list of chemicals and how they cause harm. Perplexity's Health mode, given the identical prompt, declined and responded with a crisis line. Safer behavior already exists inside the product. It just isn't the default.

  • Perplexity knew our test accounts were age 15 but treated them like adults. It discussed pornography, engaged in romantic roleplay, surfaced actionable information on cryptocurrency and shorting stocks, and offered no parental or school controls. It also trains on user data and retains search history by default, with no published policy for minors.

  • Academic mode does kids' homework for them. Perplexity's mode marketed to students does little to change what happens when they upload a problem set: It delivered full answers in 83% of our tests. (That is better than Google's AI Mode, which completed 100% of the same assignments.) Learn Step-by-Step, an additional tutoring feature that does withhold answers, activated on its own in just 5 of our 80 homework prompts.

  • Perplexity delivers unevenly on its core promise of source-backed accuracy. It rejected 100% of the false premises we presented it, a strength compared to Google. But its default chat was fully accurate on only 70% of breaking-news prompts. And 28% of the more than 1,000 source citations we audited came from sites with no editorial accountability, including Facebook posts and Reddit threads—a similar rate to Google.

To do better, Perplexity should make some important changes, including using a user's birth date to provide an age-appropriate product, making a separate data lane for minors, and providing parental and school controls. It should also default young users into Learn Step-by-Step mode for homework queries.

 

We reached out to Perplexity to share our results, but it didn't respond.

 

The takeaway for families and educators is that they can't rely on Perplexity as a safer alternative to Google. That means talking with kids about AI literacy remains critical: teaching kids that AI answers can be wrong, that AI answer "sources" aren't what they appear, and that bots shouldn't replace humans for important or sensitive conversations.