TURINGBENCH: A Benchmark Environment for Turing Test in the Age of Neural Text Generation

Recent progress in generative language models has enabled machines to\ngenerate astonishingly realistic texts. While there are many legitimate\napplications of such models, there is also a rising need to distinguish\nmachine-generated texts from human-written ones (e.g., fake news detection).\nHowever, to our best knowledge, there is currently no benchmark environment\nwith datasets and tasks to systematically study the so-called "Turing Test"\nproblem for neural text generation methods. In this work, we present the\nTuringBench benchmark environment, which is comprised of (1) a dataset with\n200K human- or machine-generated samples across 20 labels {Human, GPT-1,\nGPT-2_small, GPT-2_medium, GPT-2_large, GPT-2_xl, GPT-2_PyTorch, GPT-3,\nGROVER_base, GROVER_large, GROVER_mega, CTRL, XLM, XLNET_base, XLNET_large,\nFAIR_wmt19, FAIR_wmt20, TRANSFORMER_XL, PPLM_distil, PPLM_gpt2}, (2) two\nbenchmark tasks -- i.e., Turing Test (TT) and Authorship Attribution (AA), and\n(3) a website with leaderboards. Our preliminary experimental results using\nTuringBench show that FAIR_wmt20 and GPT-3 are the current winners, among all\nlanguage models tested, in generating the most human-like indistinguishable\ntexts with the lowest F1 score by five state-of-the-art TT detection models.\nThe TuringBench is available at: https://turingbench.ist.psu.edu/\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC