Arezo Bodaghi remembers the moment her team’s AI flagged a particularly insidious form of online harassment—veiled threats wrapped in coded language—that most systems would have missed. It wasn’t just caught; it was caught in 0.0026 seconds. That speed and precision is the promise of PPO-CIS, the artificial intelligence framework Bodaghi and her collaborators at Concordia University have built to transform how social media platforms detect toxic content. In a digital world where millions of posts flood platforms every minute, the balance between free expression and safety has never been more delicate—or more urgent. PPO-CIS doesn’t just improve detection; it reimagines it, layer by layer, decision by decision, reward by reward.
Traditional content moderation systems face a brutal trade-off: scan quickly and risk missing harmful content, or scan deeply and slow down user experience. PPO-CIS breaks that binary. Using a method called proximal policy optimization—a reinforcement learning technique—the system learns to optimize both accuracy and speed by rewarding correct identifications and penalizing errors. It dynamically adjusts its scanning intensity based on the content, allowing it to process vast amounts of data without sacrificing precision. The architecture is cascaded: a fast initial filter clears the majority of benign posts, while potentially toxic ones are passed to deeper, more accurate models. The final ambiguous cases? Those go to human moderators, ensuring no critical judgment is left to machines alone.
But what sets PPO-CIS apart is not just its design—it’s its performance. Tested on two major datasets, including the team’s own AugmenToxic and the widely recognized ToxiGen, PPO-CIS achieved 2.1% higher accuracy than existing models. More strikingly, it processed 384 content samples per second—nearly nine times faster than the current average of 43. Even compared to CETRA, a leading reinforcement-learning model adapted from malware detection, PPO-CIS came out ahead. "The system can be adapted to prioritize criteria set by individual platforms so they can determine what is toxic," says Bodaghi, now a Ph.D. graduate from Concordia’s Department of Cybersecurity and Intelligent Systems Engineering. That flexibility means a platform in one country can tailor it to local hate speech laws, while another can focus on cyberbullying or misinformation.
For co-authors Ketra Schmitt of Concordia and Benjamin Fung of McGill University, the implications go beyond efficiency. In jurisdictions with strict content removal timelines—like Germany’s NetzDG law, which mandates takedowns within 24 hours—PPO-CIS could help platforms comply without over-censoring. The study, published in Knowledge-Based Systems, marks a leap not just in AI capability, but in digital accountability. As online spaces grow more complex, tools like PPO-CIS offer a path toward safer, more responsive communities—where technology doesn’t just react to harm, but learns to prevent it before it spreads.