Inverse Turing Bench: Evaluating Language Models as Judges of Human vs. AI Dialogue

The Inverse Turing Bench has been introduced as a benchmark for evaluating language models' capabilities in distinguishing between human and AI dialogues in multi-turn text interactions. Preliminary evaluations showed that GPTZero, Claude Opus-4.6, and GPT-5.5 achieved accuracies of 89.41%, 77.92%, and 75.94%, respectively, highlighting the challenges of detection methods, where statistical approaches exhibit semantic blind spots and semantic methods are vulnerable to persona-prompting. This benchmark is significant for AI practitioners as it emphasizes the importance of human-AI differentiation in developing more robust AI systems.