Counter Turing Test (CT2) evaluation results: AI-generated text detection

In the Counter Turing Test (CT2) evaluation, the best F1 for distinguishing human-written from AI-generated text reached 1.0000, while the best model attribution was only 0.9531. Top teams used fine-tuned Transformers, ensemble learning, and hybrid methods, with DeBERTa and BART performing prominently. The low attribution scores highlight the difficulty of distinguishing outputs from different LLMs, and adversarial robustness and cross-domain generalization still require research.