ChatGPT学术文献考证---谁写的这篇论文
1 引言
在【ChatGPT---学术文献引用的反向校准】的讨论中指出，ChatGPT在大多数情况下不能正确给出文献的出处，推荐出来的论文绝大多数是"虚拟"的，假如输入一篇真实存在的论文，推荐出来的论文其“真实”的概率相对较大，真实出现的论文大概率是高引的、在本领域作出突出贡献的论文。
作一个试验，输入“A discrete numerical model for granular assemblies”，这是Cundall and Strack于1979年发表的DEM开创性论文。它推荐的论文是"Potyondy, D. O., & Cundall, P. A. (2004). A bonded-particle model for rock. International Journal of Rock Mechanics and Mining Sciences, 41(8), 1329-1364." 这是一篇真实的论文，连Vol.和页码都是完全正确的，这是因为训练库中包括了许多这篇论文。在下面的试验中，给出一篇真实的论文，问论文作者是谁？
2 谁写的这篇论文
测试(1) GSI: a geologically friendly tool for rock mass strength estimationChatGPT正确地给出了这篇论文的作者是P. Marinos and E. Hoek (2000)，当要求写一个Abstract时，它给出了如下回答：
The Geological Strength Index (GSI) system provides a rock mass characterization system for use in slope stability analysis. The GSI system is based on the characterization of the rock mass as an engineering material, taking into account the geological history, the degree of weathering, and the presence of discontinuities. The system provides a quantitative means of estimating the strength of the rock mass and the deformation characteristics of the discontinuities. The system has been used successfully in a wide range of rock mass engineering applications and has been accepted by many engineers and geologists as a valuable tool for the design of rock slopes, tunnels, and other underground openings. This paper describes the GSI system, its application in slope stability analysis, and its advantages over other rock mass characterization systems.This paper presents a review of the estimation of rock mass strength properties through the use of GSI. The GSI classification system greatly respects the geological constraints that occur in nature and are reflected in the geological information. A discussion is given regarding the ranges of the Geological Strength Index for typical rock masses with specific emphasis to heterogeneous rock masses. 这个测试显示ChatGPT的数据集中包含了这篇论文。
测试(2) Rock Mass Characterization for Underground Hard Rock MinesChatGPT不能给出这篇论文的作者是Milne, D., et al. (1998)，让其改正了两次之后，它认定作者是Nick Barton，并给出了他的一篇论文：Barton, N. (2002). "Some new Q-value correlations to assist in site characterization and tunnel design." International Journal of Rock Mechanics and Mining Sciences, 39(2), 185-216.这篇论文是真实存在而且完全正确的，但不是我们想要的答案。当要求再次修改时，ChatGPT给出了另外一篇论文：
Barton, N., Lien, R., and Lunde, J. (1974). "Engineering classification of rock masses for the design of tunnel support." Rock Mechanics, 6(4), 189-236.尽管这篇论文是真实的，但与原论文不一样，作为一种变换问法，使用论文的三位作者提问："这篇论文是Milne, D.写得吗"？即使这样，ChatGPT仍然坚持认为这篇论文是Barton(1974) 所作。这个测试显示这篇论文没有包含在ChatGPT的数据集中。3 结束语
尽管ChatGPT的自然语言处理功能已经非常强大，但在学术写作，至少在岩土工程领域的学术写作方面还存在着很大问题，绝对不能代替搜索工具，继续完善ChatGPT的数据集，特别是补充更多的Academic Database能够逐渐缩小虚拟与真实之间的差距。
