With AI, scientists are empowered to do more - worse. A model of how research gets done predicts that language models will make scientists produce more papers, but at a lower quality. Carl Bergstrom at the University of Washington and colleagues split the research process into stages and asked which ones an LLM speeds up. There's a DISCOVERY STAGE where you generate hypotheses and run enough preliminary work to judge whether a project is worth pursuing. Then a DEVELOPMENT STAGE in two halves: * required work - the figures and the drafting you can't publish without; and * discretionary work - the follow-up experiments and the polishing In coding terms, the discretionary work is the testing and bug-fixing before a release. LLMs accelerate discovery and required work - but they don't touch the discretionary half, because that half is defined by not being required. Multi-agentic feedback loops or not, the discretionary half is typically the "human in the loop" part - lots of "oh hang on, what's it even doing??" reality sandwiches. So the time saved goes into starting the next paper rather than improving the current one, and the prediction falls out of that: more output, a smaller fraction of it good. Bergstrom's point is about incentives: scientists are assessed on volume - so a tool that raises volume gets used to raise volume. His suggested fix is to judge researchers on their best work instead of counting papers and citations. (Note: Bergstrom's paper went up on arXiv on 19 July and hasn't yet been peer reviewed) #ScienceOfScience #AI #Research