Should Researchers Write Papers for AI Instead of People?

Thirty-seven researchers from about two dozen universities and tech companies just put out a paper arguing that the scientific paper, as a format, is basically obsolete. Not because the writing is bad, but because the intended reader has changed. Their claim, laid out in a piece called "The Last Human-Written Paper," is that AI agents are becoming genuine participants in research rather than assistants fetching citations, and that the infrastructure around science needs to be rebuilt with agents in mind from the ground up.

Lead author Jiachen Liu, who finished her computer science Ph.D. at the University of Michigan this year and co-founded the Agent Native Research Lab in Palo Alto in May, walked IEEE Spectrum through the thinking. Her proposed replacement is something called an Agent-Native Research Artifact, or ARA. The team even published their own paper in that format as a demonstration. Liu says the traditional paper suffers from two flaws when viewed through an AI's eyes: a "storytelling tax," where roughly 80 percent of the actual work, the failed attempts, the fiddly parameter tweaks, the dead ends, never makes it into the final draft; and an "engineering tax," where even what survives is too vague or incomplete to let anyone actually reproduce the work.

Her fix isn't to make AI adapt to human habits but the reverse. A piece of the ARA system called the Live Research Manager would sit alongside a researcher and passively log everything as it happens, so nothing needs to be manually written up. If someone still wants a polished PDF for a journal, Liu says converting the raw record back into a conventional narrative would be trivial. As for the obvious worry, that AI hallucinates and can't be trusted to self-report accurately, Liu argues the answer isn't a second language model checking the first one, since that just stacks probability on top of probability. Instead she's building a neurosymbolic system where claims get expressed as formal, provable statements rather than plausible-sounding prose.

Liu traces her own shift in thinking back to late 2024, when coding agents like Cursor first suggested to her that AI might eventually do the job of a researcher outright. Back then she still believed humans were indispensable in the loop. She's since written a follow-up piece, "The End of Human-in-the-Loop," arguing that once AI has absorbed essentially all expert-level knowledge, humans stop being useful inputs and start being the bottleneck. She frames this as a second pivotal moment in science's history, comparable to the invention of the paper itself roughly 350 years ago, when open sharing replaced secrecy and research sped up dramatically.

Asked about junior scientists losing the chance to learn by doing the grunt work AI will absorb, Liu isn't worried. She thinks people will simply learn faster by working alongside AI, producing a different kind of senior researcher rather than a worse one. Whether the rest of the scientific community shares that optimism, especially the part that still has to write, review, and trust these papers, is very much up for debate.