US and Canada Advance New Child Safety Bills
Lawmakers in North America are tightening regulations around AI and child protection. In Florida, an “AI Bill of Rights” is advancing to a Senate vote, aiming to balance free speech with child safety by requiring AI transparency and parental controls. Concurrently, Canada’s Parliament has introduced Bill C-63, the Online Harms Act, which mandates reporting of online child exploitation and holds service providers accountable.
- Florida’s bill, a priority of Governor Ron DeSantis, would grant parents the right to opt their children out of AI-based instructional tools and review their child's activity on a platform. It also mandates that users must be notified when they are interacting with an AI system versus a human. - Canada's Bill C-63 would establish a new Digital Safety Commission to enforce the law, including the power to order the removal of content that sexually victimizes a child within 24 hours. The bill also creates a duty for platforms to act responsibly to mitigate the risk of users being exposed to harmful content. - At the federal level in the U.S., the Kids Online Safety Act (KOSA) has been proposed, which would establish a "duty of care" for online platforms to prevent and mitigate harms to minors, such as content promoting eating disorders or suicide. This is part of a broader legislative push that includes proposed updates to the Children's Online Privacy Protection Act (COPPA). - To create adaptive learning systems that personalize education, engineers use a machine learning task called Knowledge Tracing (KT). Deep learning models like DKT use Recurrent Neural Networks (RNNs) to model a student's knowledge state over time based on their interactions, predicting their future performance on learning activities. - A significant technical hurdle in building AI tutors for young children is the poor performance of Automatic Speech Recognition (ASR) systems. Models like OpenAI's Whisper, which achieve a 3% word error rate on adult speech, have a 25% error rate on child speech because they are primarily trained on adult voices, which have a different pitch and rhythm. - Reinforcement Learning (RL) is a technique used to optimize adaptive learning systems in real-time. By defining a student's learning progress as a "reward," an RL agent can learn a policy for sequencing educational activities to maximize learning gains while minimizing the number of items assigned. - To balance recommending new educational content with reinforcing existing concepts, engineers can use multi-armed bandit (MAB) algorithms. This approach allows a system to manage the "explore-exploit" tradeoff by experimenting with new content (exploration) while continuing to serve content it knows is effective (exploitation). - A key concern reflected in the new legislation is the potential for AI to foster emotional dependency or be used for exploitation. Some AI safety frameworks propose "non-negotiables," such as prohibiting AI from generating fake images of children or being designed as manipulative "companion experiences".
