Date created: | Last Updated:

Category: Project

Description: In previous studies, we evaluated a variety of AI and humans in both a classic Turing test (https://arxiv.org/abs/2310.20216) and an inverted Turing test (https://doi.org/10.17605/OSF.IO/DTA89), where AI models judge whether witnesses are human or AI on the basis of transcripts from the classic test. In the classic Turing test, we found that humans were not better than chance at identifying the best-performing GPT-4 prompt as AI. In the inverted Turing test, we found GPT-4 is less accurate than chance at identifying both AI and Human witnesses. In the present study, we will run a "displaced Turing test," analogous to the the previously-run inverted Turing test but with human adjudicators. The interrogators are "displaced" in the sense that they are not interacting with the witness but making judgements on pre-existing Turing test transcripts.