Main content
Inverted Turing Test - GPT
Date created: | Last Updated:
: DOI | ARK
Creating DOI. Please wait...
Category: Project
Description: In a previous study, we evaluated a variety of AIs and humans in a Turing test (https://arxiv.org/abs/2310.20216). We found that humans were not better than chance at identifying the best-performing GPT-4 prompt as an AI. With the present project, we aim to determine GPT models' ability to distinguish between humans and AI as the observer/”adjudicator” in an Inverted Turing Test. We will present models with a sample of 500 transcripts of Turing Tests, and elicit judgements about the identity of each transcript’s witness.