What we’ve learned from three years of testing AI

By Hadley Dixon 

AI Risk Assessment Analyst, Youth AI Safety Institute

On June 5, Robbie Torney and I gave the keynote talk at the 2026 Data Science Conference & Practicum Showcase, hosted by The Data Institute. It’s an interdisciplinary research and degree program at the University of San Francisco, which I’m an alumna of. Our theme: What Common Sense Media has learned from nearly three years of safety testing AI products — and how to get the data scientists of the future to care.

We started our talk in the social media era, highlighting mistakes the tech industry and society were too slow to recognize, and don’t want to repeat. Among these are lessons about prioritizing growth, engagement, and speed while ignoring the ways social media harmed its earliest (and most vulnerable) adopters: young people. Our research shows that teens have adopted AI even more rapidly than social media, but they are also uniquely developmentally vulnerable to its downsides. Kids still are forming their understanding of relationships and trust, still discovering their values, beliefs, and what kind of person they want to become, and still learning how they fit into the world around them. AI products which simulate friendship, short-circuit the cognitive struggle of learning, and coach users through dangerous behaviors exploit this vulnerability. We have the rare opportunity to act on these risks before damage scales.

Our work grew in response to these risks. In 2023, Common Sense Media created the first-ever AI ratings system designed to assess the use and impact of AI products on kids and teens. We began evaluating products against a set of principles, centered values, such as fairness, human connection, accountability, privacy and more. Since then, we have evaluated a range of AI products, including leading frontier models (ChatGPT, Claude, Gemini, Grok, MetaAI) to show parents, educators, lawmakers, industry leaders, and the general public what responsible AI looks like.

Three years later, we’ve identified a gap: Every AI company sets its own safety bar. There are no broadly accepted standards for what "safe" AI means and regulators are not moving very fast to fill that void. The Common Sense Media Youth AI Safety Institute is stepping into this gap. We are building rigorous, independent safety standards that not only set the framework of our internal evaluations at Common Sense Media, but also set an industry-wide bar for safety so that AI developers build better, safer products for young people. The goal is a world where AI companies race to the top on youth safety because the market demands it.

The audience at our talk was filled with aspiring data scientists who will build and deploy the AI systems of tomorrow. Many were weeks out from completing their advanced degrees and laser-focused on the newest emerging machine learning techniques. During the Q&A, an alumna in the audience raised an important question: “How do you convince current students, about to enter the AI industry, to care about the long-term effects of the technology they are building and how they’re shaping our futures?”

My answer: Don’t just follow the hype of building yet another chatbot. Be intentional in how you deploy AI. We have seen in our testing what a safer AI product looks like. It is one with a human in the loop, that supplements the workload of a real person rather than automates it. It knows when a child is using it, and prioritizes their safety.

AI cannot be separated from the developers that inform, shape, and influence its use. If we want to be building safe and accountable AI, we need more conversations like these.