The Turing test typically evaluates machine intelligence by asking whether a human judge can distinguish between human and AI conversational behavior. But the test also serves as an evaluation of the judge, upon whose discriminative capabilities the merit of the test depends. We investigate this dependency by replicating two variations of the Turing test: (1) a displaced test, where human participants judge transcripts of previously conducted interrogations, and (2) an inverted test, where AI systems make similar judgments. Comparing these with traditional interactive tests, we find that displaced judges perform similarly to interactive judges, and LLM judges perform significantly worse than humans. This challenges assumptions about the importance of real-time interaction, and suggests that accuracy is not significantly impacted by displacement, but may be impacted by differences in a judge’s model of human vs. AI behavior. Our results have implications for societal risks of AI, as systems that can consistently deceive both interactive and passive observers could enable large-scale online impersonation and manipulation.