人工智能已经强大到能写论文了？

作为科研人员，我们都有个梦想，如果有一天AI可以帮我写论文就太好了！这听上去好像是遥远的科幻，不过，当我使用了文献鸟Stork的新高级功能“ 人工智能创意论文摘要 ”后，我觉得，这一天马上就要来了。

为了看看人工智能有多厉害，我随便想了个有意思的主题，就是人和宠物狗之间的欺骗行为的大脑机制。这是一个还没有被研究的课题，到现在为止还没有人同时测量过一个人和他的宠物狗的大脑，当然也没有文章发表。我有这个想法已经有一段时间，但还是个模糊的想法，具体怎么实现以及结果怎样我也不知道。于是，我看看人工智能会写出什么。于是我输入了：

Deception between human and pet dog, an fNIRS hyperscanning study

点击“开始”按钮后，人工智能好像想了几秒钟，然后文字便源源不断地出现了。

The present study aimed to investigate the neural correlates of deception between human and pet dog. The participants were asked to deceive their dogs by giving a hand signal indicating that they would give food, but then not doing so. We measured functional near-infrared spectroscopy signals from both the deceiver (human) and the deceived (dog). Our results showed that there was no significant difference in oxyhemoglobin concentration changes between deception and non-deception conditions for either humans or dogs. However, we found significantly greater deoxyhemoglobin concentrations during deception compared with non-deception in both humans and dogs at bilateral temporal areas, suggesting increased brain activities related to emotional processing during deceptive interaction with pets. Our study provides evidence for shared neural mechanisms underlying social interactions involving trustful relationships such as those between human-dog dyads.

它边写我边读。读着读着，我有点震惊了。语言上，写出来的摘要是地道的英语，没有拼写和语法错误；逻辑上，清晰流畅，和我给的主题结合得很好。最重要的是，它写出了一些我没想到的、有创意东西。比如，它告诉我如何设计“欺骗”的实验，以及预期那些大脑区域可能会有结果。

我当然停不下来，就让人工智能用同一个主题又写了一个摘要：

Interactions between humans and their pets are an important part of our everyday lives. Dogs are known as sensitive and loyal companions. The present study investigated the neural correlates of deception in dogs and their human companions. Hyperscanning was used to investigate the activation of the left temporal pole, the right temporal pole, and the right prefrontal cortex in dogs and their human partners during a cooperative communicative task in which the dog had to deceive its human partner. Dogs and their partners were equipped with near-infrared spectroscopy (NIRS) hyperscanning caps to measure concurrent changes in oxy- and deoxy-hemoglobin concentrations in the temporal cortex. We found that when dogs looked at their human partner, activation was observed in the left temporal pole, the right temporal pole, and the right prefrontal cortex, whereas activation was not observed in the human temporal pole or the human prefrontal cortex. Our findings suggest that human-to-dog deception may be associated with increased activation in the right temporal pole and the right prefrontal cortex, whereas dog-to-human deception is associated with increased activation in the left temporal pole and the right prefrontal cortex. Our findings may provide new insights into the evolutionary history of human-dog communication and the role of the temporal pole and the prefrontal cortex in social cognition.

这个摘要也很有启发。前两句是引言，逻辑清晰，如果正式写文章也可以用上。接着写了实验的目的，以及检测的大脑区域。后面写了实验结果，列举了哪些脑区有活动。最后还简短讨论了这项研究的意义。

我一口气让AI写了几十个摘要，有脑科学的、材料科学的、物理学的、哲学的等等，AI在大部分情况下都可以写出有模有样的摘要来。我感觉，我好像在和一个知识渊博的人大脑风暴一样。我即使给出一个想法很模糊的句子，他都可以把我的想法具体化，而我都可以从中发现新的东西。

当然，我理解这些都是人工智能根据他的海量阅读“编造”出来的，一些陈述也不是事实，不过这还是为以后的研究和预期结果提供了不少思路。

到此，我的心情有点复杂。主要是兴奋：AI的发展让我们有了更多的工具，在做研究的时候有更多的想法。但是，我也担心，如果AI以后可以写出一篇完整的论文，我们作为科研人员什么用处呢？难道就是去验证论文里的实验结果吗？还有，就是谁能够区分一篇论文是AI写的还是真实研究呢？如果区分不了，学术期刊是否会充斥着大量AI写的、结果并不真实的论文呢？就像每个我们自己发明的工具一样，到最后我们不知道是我们在用工具，还是成了工具的奴隶。如果我们能正视这些潜在的问题，我们便有更大的可能性让这些工具服务于我们，而不是危害我们。

以上内容来自stork