37 Researchers Propose Replacing the Scientific Paper With AI-First Format

Thirty-seven researchers from roughly two dozen top universities and tech companies published a paper on ArXiv this May arguing that scientists should stop writing papers for humans. The paper, titled "The Last Human-Written Paper," proposes a replacement format called the "Agent-Native Research Artifact" (ARA), designed so AI agents can read, reproduce, and extend scientific work efficiently. The authors published the paper itself in ARA form on GitHub.

"AI agents are becoming first-class participants in research workflows, not tools that assist humans but autonomous contributors that read, reproduce, and extend scientific work. That transition demands infrastructure built around agents from the start," the authors write.

The proposal lands amid a split in the scientific community over AI's role in research. Some evidence suggests AI-enabled research can boost individual careers in a discipline while generating fewer new ideas and topics. Some biologists, meanwhile, have embraced AI as a "co-scientist." The ARA paper takes the most aggressive position yet: scientific documentation itself should be rebuilt around machines.

Lead author Jiachen Liu developed the ARA proposal while completing her Ph.D. in computer science at the University of Michigan, awarded in 2025. This May she cofounded the Agent Native Research Lab, an AI-for-science startup in Palo Alto, Calif. She laid out the case in an interview with IEEE Spectrum.

Two taxes on scientific knowledge

Liu says the traditional paper suffers from two fundamental flaws from AI's point of view. The first is the "storytelling tax." Once research is written into a paper, roughly 80 percent of the information about the work is lost, she argues. "We only write down the last 20 percent. All the process, a lot of important decision-making, the failures, the attempts that didn't work out, they are all gone." A researcher might spend significant effort fine-tuning a small component — a parameter or a few lines of code — and none of that appears in the final paper, leaving readers unable to learn what actually made the system perform better.

The second flaw is the "engineering tax": the paper is a lossy compression of the research process. Liu says she cannot reproduce work from papers because the language is too ambiguous or implementation and experimental details are missing.

Rather than training AI to adapt to human formats, ARA flips the burden. A component called the "Live Research Manager" acts as what Liu calls "a faithful AI observer of your entire research progress." Researchers would not document anything manually; the system would observe and record everything automatically. Converting back to a polished PDF paper would remain easy.

Liu frames the moment as comparable to the invention of the scientific paper 350 years ago, when scientists stopped hiding their research to avoid being scooped and science accelerated through archives, peer review, and conferences. "I think now is also a pivot point," she says. "Because now we have AI, we can unlock a lot of new opportunities."

She reports diverse but uniformly positive feedback on the paper. Industry responders see AI-native research and knowledge systems enabling collaboration across entire enterprises. Academic responders see a fix for a pain point in sharing results that has persisted for centuries, since breakthroughs come from communities, not lone geniuses. "If they're not positive, they probably don't bother reaching out to you, right?" she says.

AI supervising AI

The obvious objection is that large language models hallucinate, and human bandwidth to check AI-generated code, results, and analyses would create a bottleneck. Liu's answer is supervision by another layer of AI, judged by a formal system. Crucially, she says the supervisor would not be another language model: "A language model alone, no matter how smart it is, has the chance to hallucinate because it's a model based on probability, not logic."

She is building a formal system using neurosymbolic techniques — combining neural networks' handling of unstructured data with symbolic AI's structures of logic — to guarantee rigor. The result would make every research paper a formal system, so every claim could be written as a mathematical formula and proved by the system, keeping all claims self-consistent.

Dropping the "narrative tax" also means exposing failures publicly, which researchers may resist. "People don't want to be perceived as dumb," Liu concedes. She frames that reluctance as an opening for AI: when an AI wastes 12 hours on a dead end, the human steering the project can call out the mistake and appear smart for catching it.

How long humans keep that steering role is an open question. Liu argues a singularity point is coming. "Once AI has 'squeezed out' all the expert data from humans, it won't need any more input from humanity. That is the time AIs will start just self-evolving by themselves," she says. "Right now, the human is the bottleneck." She elaborated on this position in an article titled "The End of Human-in-the-Loop."

On training the next generation of scientists, Liu dismisses concerns that junior researchers will be left behind. "People will grow better by learning from AI. People's learning curve is very fast with AI," she says, predicting senior researchers will still exist — with learning experiences entirely different from earlier generations.

The stakes extend beyond one startup's protocol. If machine-readable artifacts replace the 350-year-old paper as the primary unit of scientific communication, the gatekeepers of science — journals, peer review, and the humans who currently control them — lose their structural role. Liu's venture is betting that transition has already begun.