Should Researchers Write Papers for AI Instead of People?
IEEE Spectrum David Berreby
Researchers just proposed ditching traditional papers for a format built for AI agents, not humans. The pitch: AI is now smart enough to read, reproduce, and extend science on its own.
A group of 37 researchers from about two dozen universities and companies dropped a paper this May that reads less like academic work and more like a manifesto. Its title, "The Last Human-Written Paper," is not subtle. The argument: scientific papers were designed for humans reading other humans' work, but AI systems are now doing real research, and they need something different. The team's answer is a new format called an Agent-Native Research Artifact, or ARA, which the paper itself is published in, presumably so nobody misses the point.
Jiachen Liu, the lead author, finished her computer science PhD at the University of Michigan this year and promptly co-founded a startup, the Agent Native Research Lab, to chase this idea further. She told IEEE Spectrum that her thinking shifted around late 2024, when coding agents like Cursor started showing real research potential, though they still needed heavy scaffolding from human researchers. Her view now is that undergraduate-level knowledge is already baked into large language models, and PhD-level or professor-level knowledge is coming fast. Once that happens, she argues, humans stop adding value and AI needs room to keep improving on its own.
The paper identifies two problems with how papers get written today. One is what Liu calls the "storytelling tax": researchers write up maybe 20 percent of what actually happened, the polished narrative, while the failed attempts, the parameter tweaks, and the messy decision-making that led to a result all get cut. The other is the "engineering tax," meaning even the surviving 20 percent is often too vague or incomplete to reproduce. Her fix is a "Live Research Manager," an AI system that watches the entire research process as it unfolds and logs everything automatically, with the option to convert it into a traditional paper later if humans still want one.
The obvious worry is hallucination. If AI is generating and checking its own research, who catches the mistakes? Liu's answer is layered AI supervision, though she's careful to say she doesn't want one language model grading another, since both run on probability rather than logic. Instead she's building a neurosymbolic system where claims get expressed as formal, provable statements, aiming for a setup where errors get caught mathematically rather than by a second guessing model.
Where this gets genuinely strange is Liu's timeline for human involvement. She's written a companion piece called "The End of Human-in-the-Loop," predicting a point where AI has extracted everything useful from human expertise and no longer needs guidance at all. Right now, she says, humans are the bottleneck slowing AI down. She expects that to flip. As for training the next generation of scientists once AI does most of the work, her bet is that people will simply learn faster by working alongside AI systems than previous generations did without them.
My take
Calling humans a "bottleneck" in science is the kind of framing that sounds visionary right up until it sounds like a justification for cutting people out entirely, and Liu's roadmap points exactly there. The reproducibility complaints about traditional papers are real and worth fixing, but building a system explicitly designed for AI to eventually self-supervise without human input isn't infrastructure improvement, it's a bet on obsolescence dressed up as efficiency. Somebody should ask who actually benefits when the record of scientific progress stops being legible to the scientists it's supposedly serving.
Read more about this at: IEEE Spectrum
