The researchers at OpenAI asked the GPT-3 neural network to write a scientific article about itself. It succeeded (and even better than the experts expected). It wrote a 500-word text in two hours. AI used scientific language, competently substantiated its theses, inserted quotes appropriately, and described in detail the principle of its work. What’s more, GPT-3, without too much modesty, called itself “an outstanding achievement in the field of AI.” Is it an objective reality or a manifestation of excessive selfishness? That is the question.
Perhaps the neural network simply waited too long for its abilities to be appreciated. GPT-3 has long been known as a tool that can create human-like text. During its existence, it has written news articles, generated content on behalf of other authors, and even created books in 24 hours. However, all these projects were implemented with the direct participation of a person – GPT-3 has never been the main author of its work. That’s why Almira Tunström, an employee at the University of Gothenburg, decided to give AI a new challenge: to write a scientific paper almost on its own. The specialists only directed the neural network and gave hints regarding the sections. There was no editing or selection of the best options,says SRSLY.
GPT-3 was asked to write an article about itself for two simple reasons. Firstly, this is a fairly new tool, and therefore little explored. This means that there is a minimum amount of data in open sources, and the neural network will have to figure it out on its own. And secondly, if mistakes were made, there would be nothing to worry about: the article is just a part of the experimental work.
Problems began when the text was published in a specialized publication. It was necessary to indicate the name and contact details of the authors. The neural network forgot to share this information, and the specialists had to get out on their own. In addition, there was an important legal issue: did GPT-3 even agree to be published? Tunström asked the AI directly, feeling like she was talking to a sentient being. GPT-3 didn’t mind. Subsequently, according to Scientific American, the article appeared on the international preprint site HAL. It is still being reviewed before being published in a scientific journal, notes NIXSolutions. The reason for this, probably, was an unusual author. Meanwhile, experts are pondering how promising this format is and whether a neural network can replace scientists.