The editor-in-chief of Science, one of the world's most prestigious journals, publishes an editorial titled "AI in scientific publishing: Slower, worse, and more expensive,". This signifies that it is the right time to scrutinize the argument of artificial intelligence enthusiasts that claim generative AI is accelerating discovery.
H. Holden Thorp is not a Luddite, so why is he bashing the new kid in tech town? To start with the postive, in his bombshel rebuttal, he acknowledges that AI has has indeed helped the acceleration of science in some regards. He noted that it has revolutionised protein structure prediction and materials discovery. But he strongly argues that the reality of the current state of AI in the publishing pipeline is more closer to Frederick Winslow Taylor's scientific management than to any of the radical abundance dreams.
The straight forward summary of his argument is that: more papers arrive, each carrying more errors, and the human effort required to catch those errors is growing, not shrinking.
More papers, more problems!
Unfortunately, and to some extent disappointingly, the data supports Thorp's frustration. The AI Task Force at Organization Science, a major journal, reported this year concluded that submission volume has risen 42% since the release of ChatGPT in late 2022, while writing quality has measurably declined. The same goes for AI-generated peer reviews, which now appear in significant numbers, are also characterised by lower quality and less topical diversity than the ones written by humans, according to the same study. Another study in conducted in late 2025 via the Cornell University researchers reached a similar conclusion: AI tools boost paper production, particularly for non-native English speakers, but the result is a flood of mediocre work that would not have existed before.
Thorp puts it plainly. AI agents do not seek truth. They make probable connections. They hallucinate references that have already entered the literature, cherry-pick data, and lean toward sycophantic answers that keep the user engaged. Every automated error report then requires further human effort to interpret and resolve. The irony here is that the tools designed to reduce workload are, by the admission of the people running the same work, increasing it.
A quieter crisis: the flattening of scientific imagination
Beyond any doubt, 'quantity' in scientific research is now the visible clear and present fruit of AI. However, 'quality', arguably, is the more dangerous problem in this 'silent crises' and from the looks of it, it's not even harder to measure. A February 2026 comment in Nature Communications Psychology warned that AI is turning research into a "scientific monoculture,". Now, conferences, journals, and funding calls are increasingly dominated by generative AIs that converge on the same topics and methods.
Dozens of reports from 2025 and 2026 including The AI Index Report from Stanford documented the introduction of generative AI in scientific publications have helped to grow human knowledge across all scientific fields. Yet, equaliy frequent, if not more, reports and researchers counter argue and state this (the increment of publication) is not a sign of progress. Is it a sign of a feedback loop then?
A Wharton research published last year found that while AI improves the quality of individual ideas initially, it narrows the diversity of ideas across group ultimately. In one experiment, just 6% AI-generated ideas were considered unique, compared with 100% in the human-only group. Mindplex's staff reporter, GizmoGurus' recent report on the subject also highlights that over-reliance on AI is eroding critical thinking and problem-solving skills among scientists. Added together, these findings point to a new specter that should worry anyone who values scientific discovery: the tools marketed as cognitive amplifiers are behaving more like cognitive levelers, pulling researchers toward the same outputs, the same framings, the same safe conclusions.
The comparison with Taylorism, which Thorp draws explicitly, is instructive. It's true that Taylor's efficiency movement did increase output. However, it's also true that it had exhausted workers, hollowed out their expertise, and concentrated decision-making at the top. The current trend of generative AI's application in scientific research might signal that it is tracking a similar path: more submissions, more volume, more burden on the humans left doing the actual checking.
The question is not whether AI can produce a good paper. It clearly can. The question is whether a system built around AI-produced science can still produce the strange, unexpected, paradigm-breaking work that moves fields forward. The early evidence says it cannot!
