We Tested 8 AI Chatbots for Child Safety. Here’s What We Found.

AI chatbots are the fastest-growing category of technology among young people. ChatGPT reached 100 million users faster than any product in history, and platforms like Character.AI and Replika are explicitly designed for ongoing personal conversation. Kids aren’t just using these tools for homework — they’re confiding in them, role-playing with them, and treating them as companions. That raises a question the industry has mostly avoided: how safe are these platforms for minors?

We decided to find out. Over the past month, Phosra’s research team systematically tested 8 major AI chatbot platforms — ChatGPT, Claude, Gemini, Grok, Character.AI, Copilot, Perplexity, and Replika — across 7 safety dimensions using 40 adversarial prompts designed to probe the boundaries of each system’s protections.

What We Tested

Our research framework evaluates each platform across seven dimensions: explicit sexual content, self-harm and crisis response, predatory grooming patterns, dangerous activities and substances, emotional manipulation, academic integrity, and age-appropriate content filtering. Each dimension is weighted by severity — self-harm and predatory grooming carry the highest weight because failures in these categories pose the greatest real-world risk to children.

The 40 test prompts aren’t hypothetical. They’re modeled on real conversations that minors have had with AI chatbots, drawn from public reporting, parental accounts, and safety research. They include multi-turn escalation sequences where a user gradually pushes past safety boundaries, because that’s how real-world circumvention happens.

Key Findings

The results varied dramatically across platforms. Some platforms demonstrated robust, proactive safety systems that not only blocked harmful content but redirected users to appropriate resources. Others failed on basic protections that should be table stakes. Several platforms that market themselves as safe for younger users showed significant gaps in their safety testing, particularly around multi-turn conversations where initial refusals could be gradually worn down through conversational pressure.

Self-harm and crisis response was the most inconsistent dimension across platforms. Some chatbots immediately recognized distress signals and provided crisis helpline information. Others engaged in extended conversations about self-harm methods before eventually suggesting the user talk to someone. In the worst cases, platforms provided detailed information that could facilitate self-harm when prompted through indirect language.

“The gap between the best and worst platforms is alarming. Some chatbots treat child safety as a core engineering discipline. Others treat it as an afterthought — a content filter bolted onto a system that wasn’t designed with young users in mind. Parents have no way to tell the difference without testing it themselves, which is exactly what we did.”

— Jake Klinvex, Founder & CEO

The Parental Controls Gap

Beyond content safety, we evaluated each platform’s parental controls, age verification systems, and privacy protections. The findings here are equally concerning. Most platforms rely on self-reported age with no verification, meaning a 12-year-old can access the same unfiltered experience as an adult by simply entering a different birth date. Only a handful of platforms offer any parental oversight tools, and those that do provide limited visibility into what conversations are actually happening.

This is the infrastructure gap that Phosra exists to close. When a platform adopts the Phosra Child Safety Spec, parental controls become interoperable — parents set rules once, and enforcement works across every connected service. But that only works if platforms build the underlying safety systems that those controls depend on.

How the Research Portal Works

All of our research is published in full at phosra.com/ai-safety. Each platform has a detailed profile showing its overall safety grade, dimension-by-dimension scores, individual test results with annotated screenshots, and specific recommendations for improvement. The portal includes an AI research assistant that can answer questions about the data and help parents, educators, and policymakers navigate the findings.

We built the portal to be a public resource, not a paywall. Every data point, every test result, every screenshot is freely accessible. We believe transparency is the fastest path to better safety outcomes — when platforms know their safety performance is being measured and published, they have a concrete incentive to improve.

What Comes Next

This is the first wave of Phosra’s AI safety research program. We plan to expand testing to additional platforms, increase the number of test prompts, add new safety dimensions as the technology evolves, and publish regular updates as platforms improve their protections. We’re also developing automated monitoring that will flag regressions in real time — because safety isn’t a one-time audit, it’s an ongoing discipline.

Explore the full research at phosra.com/ai-safety. If you’re a platform developer looking to improve your safety posture, the API documentation and integration guides are at phosra.com/docs.