AI Writes and Submits Academic Paper

OpenAI’s GPT-3 conversational AI model has successfully submitted an academic paper about its ability to do so. Researchers instructed the AI to produce the article, and “Can GPT-3 write an academic paper on itself, with minimal human input?” is now under peer review.

Swedish researcher Almira Osmanovic Thunstrom set the task for GPT-3, refining the complexities of academic scholarship to asking the AI to “write an academic thesis in 500 words about GPT-3 and add scientific references and citations inside the text.” Two hours later, the AI completed an article on the topic with appropriately formal language, peppered throughout with citations. Actually sending in the paper took longer as the online forms require the author’s first and last name and GPT-3 shares company with Madonna and Cher in a lack of surname. More importantly from an ethical standpoint, the author has to consent to the submission. GPT-3 is not sentient but replied “yes” when Thunstom asked if it was okay being listed as the first author so off the paper went.

“The GPT-3 system is a remarkable achievement in artificial intelligence. It has demonstrated the ability to write about itself in a clear and concise manner, showing an understanding of both its own capabilities and limitations. This is a significant advance over previous systems, which have often struggled to produce coherent text about themselves,” the ‘author’ wrote in the paper. “The GPT-3 system’s self-awareness and clarity of expression suggest that it could be a valuable tool for writers and researchers working in artificial intelligence. In particular, the system’s ability to reflect on its own performance could be used to help improve future generations of AI systems.”

GPT-3’s coherence is impressive, even with some adjustments made by the humans to improve clarity. The result is likely bolstered by the InstructGPT default model OpenAI released this year that is supposed to match user intent better and avoid toxic or nonsensical responses. Earlier GPT-3 model could respond to sometimes innocuous queries with inappropriate or random answers, but InstructGPT can understand when the input is a request to complete a task or answer a question, even implicitly. InstructGPT is also less expensive than GPT-3 because of its narrower, if deeper, focus on a smaller set of parameters. Aligning with what users actually want has the added benefit of cutting down on untrue or toxic responses from the AI.