Elamur Research
☰
home / papers / 2023 / How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection

How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive license to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/2301.07597

Summary

ChatGPT can produce fluent answers to almost any question, so researchers need ways to measure how far it is from human experts and to recognise text it generates. The authors built the Human ChatGPT Comparison Corpus (HC3), which pairs human answers with ChatGPT answers to the same questions. It has English and Chinese parts. The human answers come from existing QA datasets (expert or highly voted replies) and from wiki-style concept definitions. The ChatGPT answers were collected by hand through its web interface, starting a fresh thread for each question. Using HC3, the team ran Turing-style and helpfulness tests with volunteers, did a linguistic analysis, and trained three detectors, among them RoBERTa-based classifiers and GLTR. They also tested which factors change detection quality. People who use ChatGPT often can usually spot its answers. Newcomers mostly cannot. Fine-tuned RoBERTa detects full ChatGPT answers almost perfectly, but it does much worse on single sentences and on unfamiliar sources. The released corpus and models are an early, widely cited baseline for work on detecting AI-generated content.

Key points

  • HC3 size: the English part has 24,322 questions, 58,546 human answers and 26,903 ChatGPT answers in 5 splits (reddit_eli5, open_qa, wiki_csai, medicine, finance). The Chinese part has 12,853 questions, 22,259 human answers and 17,522 ChatGPT answers in 7 splits, including baike, nlpcc_dbqa, law and psychology.
  • Human evaluation (English, 17 volunteers): experts given a human/ChatGPT answer pair picked out ChatGPT 90% of the time. Experts shown a single answer were right 81% of the time, and amateurs only 48%, which is about chance. The helpfulness score was 0.57.
  • Detection: RoBERTa reaches about 99.8% F1 on full answers. On isolated sentences F1 falls to roughly 81–92%. Models trained on question+answer input (the QA setting) are more robust on sentences than models trained on the answer alone.
  • Hard cases: for English, wiki_csai and open_qa were the hardest sources and reddit_eli5 and finance the easiest. For Chinese, nlpcc_dbqa was the hardest. The Chinese results held up better across test settings.
  • Limitations: the sources are unbalanced, the answers were collected without special prompts (so style prompts could get past the detectors), and the Chinese conclusions may be less reliable. Data, code and models are at github.com/Hello-SimpleAI/chatgpt-comparison-detection.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

257 papers of our library cite it (of the 140673 whose bibliographies we read); 11 of them have a page here, 246 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2301.07597 · arXiv 2301.07597

@misc{guo2023how,
  title = {{How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection}},
  author = {Guo, Biyang and Zhang, Xin and Wang, Ziyuan and Jiang, Minqi and Nie, Jinran and Ding, Yuxuan and Yue, Jianwei and Wu, Yupeng},
  year = {2023},
  doi = {10.48550/arxiv.2301.07597},
  eprint = {2301.07597},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2301.07597}
}

References

40 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861, 2021. find our card
  2. Francesco Barbieri, Luis Espinosa Anke, and Jose Camacho-Collados. Xlm-t: Multilingual language models in twitter for sentiment analysis and beyond. In Proceedings of the Language Resources and Evaluation Conference, pages 258–266, Marseille, France, June 2022. European Language Resources Associatio… find
  3. Steven Bird, Ewan Klein, and Edward Loper. Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O’Reilly Media, Inc., 2009. find
  4. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. find
  5. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. find our card
  6. Shu Chen, Zeqian Ju, Xiangyu Dong, Hongchao Fang, Sicheng Wang, Yue Yang, Jiaqi Zeng, Ruisi Zhang, Ruoyu Zhang, Meng Zhou, Penghui Zhu, and Pengtao Xie. Meddialog: a large-scale medical dialogue dataset. arXiv preprint arXiv:2004.03329, 2020. — [arXiv:2004.03329](https://arxiv.org/abs/2004.03329) find
  7. Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. neural information processing systems, 2017. find
  8. Nan Duan. Overview of the nlpcc-iccpol 2016 shared task: Open domain chinese question answering. In Natural Language Understanding and Intelligent Applications, pages 942–948, Cham, 2016. Springer International Publishing. find
  9. Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. Tweepfake: About detecting deepfake tweets. Plos one, 16(5):e0251415, 2021. find
  10. Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. ELI5: long form question answering. In Anna Korhonen, David R. Traum, and Lluís Màrquez, editors, Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, J… find our card
  11. Zhihui Fang. The language demands of science reading in middle school. International journal of science education, 28(5):491–520, 2006. find
  12. Yao Fu, Hao Peng, and Tushar Khot. How does gpt obtain its ability? tracing emergent abilities of language models to their sources. Yao Fu’s Notion, Dec 2022. find
  13. Sebastian Gehrmann, Hendrik Strobelt, and Alexander Rush. GLTR: Statistical detection and visualization of generated text. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 111–116, Florence, Italy, July 2019. Association for Com… find our card
  14. Ariel Goldstein, Zaid Zada, Eliav Buchnik, Mariano Schain, Amy Price, Bobbi Aubrey, Samuel A Nastase, Amir Feder, Dotan Emanuel, Alon Cohen, et al. Shared computational principles for language processing in humans and deep language models. Nature neuroscience, 25(3):369–380, 2022. find
  15. Biyang Guo, Yeyun Gong, Yelong Shen, Songqiao Han, Hailiang Huang, Nan Duan, and Weizhu Chen. Genius: Sketch-based language model pre-training via extreme and selective masking for text generation and augmentation. arXiv preprint arXiv:2211.10330, 2022. — [arXiv:2211.10330](https://arxiv.org/abs/2211.10330) find
  16. Biyang Guo, Songqiao Han, and Hailiang Huang. Selective text augmentation with word roles for low-resource text classification. arXiv preprint arXiv:2209.01560, 2022. — [arXiv:2209.01560](https://arxiv.org/abs/2209.01560) find
  17. Songqiao Han, Xiyang Hu, Hailiang Huang, Minqi Jiang, and Yue Zhao. Adbench: Anomaly detection benchmark. Advances in Neural Information Processing Systems (NeurIPS), 2022. find
  18. Jennifer Hu, Sammy Floyd, Olessia Jouravlev, Evelina Fedorenko, and Edward Gibson. A fine-grained comparison of pragmatic language understanding in humans and language models. arXiv preprint arXiv:2212.06801, 2022. find our card
  19. Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks Lakshmanan, V.S. Automatic detection of machine generated text: A critical survey. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2296–2309, Barcelona, Spain (Online), December 2020. International Committee on C… find
  20. Katharina Jeblick, Balthasar Schachtner, Jakob Dexl, Andreas Mittermeier, Anna Theresa Stüber, Johanna Topalis, Tobias Weber, Philipp Wesp, Bastian Sabel, Jens Ricke, et al. Chatgpt makes medicine easy to swallow: An exploratory case study on simplified radiology reports. arXiv preprint arXiv:2212.1… — [arXiv:2212.14882](https://arxiv.org/abs/2212.14882) find
  21. Michael R King. The future of ai in medicine: a perspective from a chatbot, 2022. find
  22. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. find
  23. Macedo Maia, Siegfried Handschuh, Andr’e Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. Www’18 open challenge: Financial opinion mining and question answering. In Companion Proceedings of the The Web Conference 2018, WWW ’18, page 1941–1942, Republic and Canton of Geneva… find
  24. William Nagy and Dianna Townsend. Words as tools: Learning academic vocabulary as language acquisition. Reading research quarterly, 47(1):91–108, 2012. find
  25. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. find
  26. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830… find
  27. Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. Deepfake text detection: Limitations and opportunities. In Proc. of IEEE S&P, 2023. find
  28. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. find
  29. Mary J Schleppegrell. The language of schooling: A functional linguistics perspective. Routledge, 2004. find
  30. Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. find our card
  31. SophonPlus. Chinesenlpcorpus. https://github.com/SophonPlus/ChineseNlpCorpus, 2019. ↗
  32. Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F. Christiano. Learning to summarize from human feedback. neural information processing systems, 2020. find
  33. Teo Susnjak. Chatgpt: The end of online exam integrity? arXiv preprint arXiv:2212.09292, 2022. — [arXiv:2212.09292](https://arxiv.org/abs/2212.09292) find
  34. Alan M Turing. Computing machinery and intelligence. In Parsing the turing test, pages 23–65. Springer, 2009. find
  35. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. find
  36. Junjie Wang, Yuxiang Zhang, Lin Zhang, Ping Yang, Xinyu Gao, Ziwei Wu, Xiaoqun Dong, Junqing He, Jianheng Zhuo, Qi Yang, Yongfeng Huang, Xiayu Li, Yanghan Wu, Junyu Lu, Xinyu Zhu, Weifeng Chen, Ting Han, Kunhao Pan, Rui Wang, Hao Wang, Xiaojun Wu, Zhongshen Zeng, Chongpei Chen, Ruyi Gan, and Jiaxing… — [arXiv:2209.02970](https://arxiv.org/abs/2209.02970) find
  37. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoe… find our card
  38. Bright Xu. Nlp chinese corpus: Large scale chinese corpus for nlp, September 2019. find
  39. Yi Yang, Scott Wen-tau Yih, and Chris Meek. Wikiqa: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. ACL - Association for Computational Linguistics, September 2015. find our card
  40. Lili Yao, Nanyun Peng, Ralph Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. Plan-and-write: Towards better automatic storytelling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7378–7385, 2019. find