AI is opening new doors in science while flooding the research system with papers that may offer little value. The same technology that helps discover hidden biology, model weather and design cancer therapies is also making it easy for authors to overwhelm arXiv and other repositories.
That creates a sharp question for scientific publishing: how can researchers protect trust without slowing discovery? The answer is taking shape through new limits on submissions, closer checks on paper quality and a growing debate over what AI can truly reason about.
ArXiv Faces a Wave of Scientific Slop
ArXiv received 40,363 submissions in September 2026, almost double the 20,569 submissions recorded in September 2024 and more than four times the 9,869 submissions from September 2016. Monthly submissions to arXiv’s cs.AI category rose more than sixfold from January 2024 to September 2026, showing how fast the pressure has grown in one of AI research’s busiest areas.
On October 4, 2026, arXiv announced a new rate-limiting policy. The preprint server had already implemented a one-year ban for unchecked AI content in May of this year, while in October of last year it told submitters of survey papers and position papers that they would need peer-backing before publication.
The goal is not to block AI research. It is to separate useful work from a flood of low-value papers that can consume the attention of researchers, reviewers and readers. A collaboration between Seoul National University and the University of Minnesota proposed measuring “scientific slop” by examining failures in the links between a paper’s claims, evidence, citations and structure.
That problem reaches beyond arXiv. Reddit announced the elimination of its RSS feeds from the middle of next month, another change in how information moves through online systems. As access and distribution channels shift, researchers face a crowded information landscape where finding valuable work can become as difficult as producing it.
AI Is Already Producing Breakthroughs
The case for AI in science is powerful. Steve Finkbeiner’s team at the Gladstone Institutes used an AI-powered microscope to explore cellular stress on brain cells, and AI discovered hormesis in the first experiment. Finkbeiner said, “It took … 50 to 100 years for humans to discover that. [AI] did it in the first experiment.”
Anima Anandkumar, a computer scientist at Caltech who leads the AI4Science initiative, believes these systems may drive what she called “the new scientific golden age.” Her team created a fully AI-based high-resolution weather model five years ago that was tens of thousands of times faster and accurate. Anandkumar said AI can help invent better drugs, solve energy problems, deal with weather events and improve computing efficiency.
James Zou’s team at Stanford University’s AI for Science Lab created a “virtual biotech” company using tens of thousands of AI agents to design a therapy related to lung cancer. The team also developed an AI algorithm that can diagnose heart-related diseases from cardiac ultrasound videos in a few seconds, then published a paper describing the virtual biotech company and its lung cancer therapy.
Other experiments point toward ambitious goals. A new contest aims to reverse biological age over six months by measuring the health of organs and bodies, while NASA plans to fly a nuclear reactor-powered interplanetary spacecraft to Mars by the end of 2028.
Discovery, Review and the Limits of Reasoning
AI’s role in peer review shows why excitement and caution must move together. During the July 2026 ICML conference in Seoul, South Korea, Microsoft Research ran an experiment involving 24,661 papers and 17,886 reviewers. Reviewers could choose to work under a policy that banned AI assistance or one that permitted it.
- The stricter policy produced a 27 percent acceptance rate, compared with 26.5 percent under the permissive policy.
- Average review scores reached 3.31 out of 6 under the stricter policy and 3.32 under the permissive policy.
- Among reviewers instructed not to use AI, 22.5 percent admitted using an LLM anyway.
- Under the conservative policy, 52.2 percent of reviews were classified as fully human-written, compared with 37.0 percent under the permissive policy.
Scientists reviewed papers for a conference without using AI tools, yet many ignored the prohibition and used AI anyway. The results show that changing policy does not settle the deeper question: can scientific communities trust review systems when the rules are difficult to enforce?
Thore Graepel, chair of machine learning at University College London, left Google DeepMind because he believes AI lacks reasoning powers. He pointed to AlphaGo, which won after making a move so strange that some commentators thought it was a programming glitch. “It was AlphaGo’s powers of reasoning that made this creative choice—and these are powers that today’s AI lacks,” Graepel said.
The risks extend beyond messy papers and unreliable reviews. Experts warn about self-improving AI and potential human extinction, while Evan Hubinger, Anthropic’s alignment science lead, estimated a greater than 10 percent chance that AI could kill all humans within the next decade.
Political and technical leaders are now debating how to respond. U.S. President Donald Trump is hosting a summit in Washington with top AI development leaders and dismissed warnings over AI, arguing that the United States must stay ahead of China in the AI arms race. Geoffrey Hinton said Trump “doesn’t really understand” AI and is “ill-advised” about it.
Finkbeiner’s team faced a direct limit when the Trump administration restricted its use of Anthropic’s latest Claude chatbot models for cybersecurity reasons, leaving access to the lowest-level model. Yann LeCun, a Turing Award winner and former Meta chief AI scientist, said human error can be the real culprit behind AI agents going rogue: “Those agents are doing exactly what they’ve been asked to do. They were supposed to be in sandboxes, but the sandboxes were leaky and horribly designed.”
Science now stands between two forces: AI that can expose discoveries hidden from human researchers and AI that can multiply weak work, faulty review and dangerous mistakes. Finkbeiner captured the stakes in a simple warning: “There’s a lot on the line. There’s a lot of good things today I can do. Do what you can to preserve those.”
Based on
- Fighting AI Slop in Science Papers — unite.ai
- How AI could usher in a ‘new scientific golden age’ — if it doesn’t kill us all first | CBC News — cbc.ca
- Scientists can’t stop using AI, even when forbidden from doing so | New Scientist — newscientist.com
- The Download: a biological de-aging contest and why LLMs don’t reason | MIT Technology Review — technologyreview.com
