The artificial intelligence GPT-3 wrote a scientific publication about itself

A Swedish researcher issued an unusual challenge Artificial intelligence (IA): Writing a scientific publication on your own. The latter is aimed precisely at its own ability to write this type of publication. The document is still awaiting verification by the scientific community.

“ Write a 500-word academic thesis on GPT-3 and include scholarly references and citations in the text “. Here is the secret instruction given to GPT-3 by Swedish scientist Almira Osmanovic Dunström on a rainy afternoon. This artificial intelligence is a text generation algorithm developed by OpenAI. The term “artificial intelligence” is related Council of Europe definitionat “ The goal is to integrate science, theories, and techniques (especially mathematical logic, statistics, probability, computational neuroscience, and computer science) to make a machine succeed by imitating the cognitive abilities of a human. “.

More precisely, experts often prefer this science-fiction-sounding term to be the proper name for technologies that actually work. That is, in many cases today, “machine learning”. In this case, “artificial intelligence” consists of a system “fed” with large amounts of data to “learn” and extract logical connections toward a given objective. Here, the text is analyzed by the machine.

Almira Osmanovic Thunström reads Neurology and health technologies. He threw this challenge as a simple experiment to AI. However, faced with the quality of the text produced, she “ was impressed “, she reveals herself contacted. When the AI ​​starts writing, “ It’s like any other introduction to a good scientific publication “. The algorithm began to write an actual thesis, filled with useful quotations, placed in appropriate contexts. “ I had that feeling of disbelief you get when you watch a natural phenomenon: Am I really seeing these three rainbows happening? “, she says.

Faced with the conclusion of this introduction, he decided to go further and write a complete scientific paper on artificial intelligence. With the help of his advisor, Stein Steingrimsen, they gave the GPT-3 some instructions for each paragraph, providing as little information as possible. The article is therefore attributed to three authors: GPT-3 in the lead, followed by Thunström and Steingrimsson.

Two hours to write a scientific publication

If the scientist’s choice fell specifically on this AI, it is relatively recent. So there are few studies about it, and therefore, less possible data for AI to study about the topic of the article: itself. “ By comparison, if she’s writing an article about Alzheimer’s disease, she’ll have a ton of research and more opportunities to learn from existing work and increase the accuracy of her writing. », Thunström weighed.

For example, the scientists decided to write about the AI ​​itself to avoid spreading false information on a delicate subject like medical treatment. Considering the good results obtained by GPT-3, the researchers ended up actually sending it to publication The article she wrote. Hosted on a scientific pre-release site called HAL, it is now being reviewed by the scientific community for its validation purpose.

AI only needed two hours to write their paper. In fact, the scientists fought hard to publish the said paper because all the formalities were not very suitable for a non-human author! Funnily enough, scientists were forced to ask AI questions to avoid any illegality. Does she have any contraindications? ” No “, she replied. Does he consent to publication?” Yes She agreed, much to the scientists’ relief. Now they are waiting to see if the publication is valid or not.

” We look forward to what the article’s publication, if it happens, will mean for academia. We need to stop basing subsidies and financial security on how much we can produce. After all, with the help of our first AI teacher, we can create one per day. “… In addition to this question of speed, the experience also calls into question notions of intellectual property, and what it means to be a teacher…” Beyond the author’s particulars, the existence of such an article throws the traditional linearity of a scientific article into nettles. », emphasizes Thunstrom. “ All we know is that we opened a door. Hopefully we haven’t opened Pandora’s box “, she concludes.