Based on the underlying paper, the key innovation is not simply better toxicity detection, but an adaptive cascade of AI models. Instead of sending every post through a large, computationally expensive detector, the system starts with faster, lightweight classifiers. If a piece of content is clearly benign or clearly toxic, a decision is made immediately. Only ambiguous cases are passed to increasingly powerful models. A reinforcement-learning algorithm decides which model to consult next, balancing speed, accuracy and computing cost. This approach dramatically increases throughput while slightly improving detection accuracy. (ScienceDirect)

For a 50–60 word EurekAlert summary:

Concordia researchers have developed an AI system that detects toxic online content faster and more accurately by combining multiple detection models in an adaptive sequence. Simple cases are screened by lightweight classifiers, while only difficult content is sent to more powerful AI tools. The approach improves accuracy while processing content up to nine times faster than conventional methods.