The research center says the AI system, Carl, can write peer-reviewed papers without human assistance.
Autoscience Institute, a research center focused on improving AI systems, has built an AI tool that can write research papers. The AI system, called Carl, can conduct research, perform experiments, and write papers to present the findings. It can do all this without human assistance.
Carl has become the first AI system to write a peer-reviewed paper. The paper was accepted at a workshop in the International Conference on Learning Representations (ICLR), after passing double-blind peer reviews. In double-blind peer reviews, the authors and the reviewers remain anonymous to each other. Hence, the reviewers did not know that the paper was written by AI.
The ICLR is an international body of professionals dedicated to advancing humanity’s knowledge in the branch of AI called deep learning. Deep learning is a specific focus in AI that tries to get computer programs to simulate human decision making
Carl is designed to conduct research specifically in the field of AI. It first searches existing literature and generates a research topic. It then formulates hypotheses, creates experiments to test them, and then writes its findings to the highest level of academic standard.
Autoscience Institute says that it takes just seconds to read and grasp published papers. This allows it to work much faster than any human ever can. As an AI tool, it also “works nonstop,” meaning that it can take on projects at all times of day.
Carl needed human assistance in only a few instances. A human was required to give it a go-ahead signal at specific points during research. This is a feature purposely built in by Autoscience Institute to control computational costs. A human team also manually edited the citations to match the referencing style required by ICLR. In addition, humans had to help Carl copy and paste prompts to and from OpenAI’s o1-pro and Deep Research models, which it uses for its work.
However, Autoscience Institute says this was only because those models do not have APIs available yet. Once they do, there would be no need for humans to assist Carl with this, according to the research center.
Carl is the latest example of AI models designed to help scientists with research. Last month, Google released a model, called Co-Scientist, that can help scientists rapidly conduct biomedical research. Experts believe that AI assistants can advance scientific research and help make discoveries that would otherwise take humans years, if not decades.
Autoscience Institute has since withdrawn Carl’s paper from the ICLR workshop. The center believes that the administrators need time to establish procedures that will apply to research conducted by AI. However, the institute plans to propose another workshop where AI systems like Carl would be given an opportunity to submit papers alongside humans.






Leave a comment