When AI helps write papers, polish no longer reliably signals that a scientist did the work, and even experts struggle to tell.

This explores what happens to scientists' own judgment (how they write, read, generate ideas, and evaluate each other's work) when LLMs sit in the middle of all of these steps. The corpus doesn't show one big failure. It shows several small ones, each at a point where science depends on a person checking something for themselves. The first is about signals. Across 2.1 million preprints, scientists who adopted LLMs published 23.7–89.3% more papers. Over the same period, the old link between complex, careful prose and paper quality reversed Does LLM writing assistance change how scientists publish?. Reviewers and readers have long used polish as a rough sign that the author did the work, and that sign is fading. Even ML experts can't reliably tell LLM-written abstracts from human ones, and when authorship was disclosed, readers preferred the LLM-edited versions 55% of the time Can readers tell LLM abstracts from human ones?. Polished writing is getting cheap, so it no longer shows whether a human thought hard about the work.

The second concern is that LLMs quietly change what a finding says. In a study of 4,900 summaries from ten models, most dropped qualifiers and stated findings more broadly than the source did. LLM summaries were nearly five times more likely than human ones to overgeneralize. The surprising part is that asking the model to be accurate made the problem worse Do LLMs overgeneralize when summarizing scientific research?. If scientists increasingly learn about each other's work through LLM summaries, the field's sense of what has been shown can drift without anyone noticing. A related problem shows up in research methods. When one researcher tweaks prompts until the output looks right, the evaluation criteria start bending toward what the LLM can do rather than what the task needs. The result is a feedback loop that confirms itself Does iterative prompt engineering undermine scientific validity?.

Third, there's the question of whose ideas these are. In a study with over 100 NLP researchers, LLM-generated research ideas were rated more novel than ideas from human experts, though slightly less feasible Do language models generate more novel research ideas than experts?. Fine-tuned LLMs also beat neuroscientists at predicting which experimental results actually happened Can LLMs predict novel scientific results better than experts?. These findings cut both ways. LLMs may widen the range of ideas researchers consider. But if many researchers draw ideas and hunches from the same few models, science could lose the independent viewpoints it relies on to catch mistakes. The corpus raises this risk but doesn't measure it directly.

Finally, peer review itself is being handed partly to machines. A structured LLM pipeline matched human reviewers' reasoning about novelty 86.5% of the time on ICLR submissions Can structured pipelines make LLM novelty assessment reliable?. That's impressive, and it's also a reason to worry: the remaining gap is where reviewer judgment is most needed. ICLR 2026 offers one way to handle this. The program chairs treated imperfect LLM detectors as one input for human area chairs rather than as automatic filters. They desk-rejected papers only for something they could verify, namely fabricated references How can conferences detect and handle LLM misuse in peer review?. The lesson across the corpus is that independence depends less on whether scientists use LLMs and more on keeping a human check at each point where a model could substitute its judgment for theirs.