VIEWPOINTS

The fatigue of the AI referee

Using AI to critique our papers and grants promises relief. Why does it often leave us more exhausted and detached, and less like the authors of our own work?

I don’t know about you, but when I use AI for a task, I rarely come away relieved. More often I am more tired than when I started, and more detached from what I set out to do. I believe this is AI fatigue, and here I describe how it crept in and what I think we can do about it.

From party trick to referee on demand

Like many scientists, I was lured by large language models (LLMs) such as Claude and Copilot, especially over the last couple of years, largely out of fear of missing out (FOMO). I began with routine documents (letters, admin tasks and the like), and I was amazed, as everybody was, that a document could be written in seconds. Then came the era of hallucinations and inaccuracies. I had to check everything, re-prompt, and fine-tune, and in the end I didn’t save much time compared with drafting from scratch. I assumed this was a teething problem, and largely it was.

As the models improved, they became good at mimicking my writing style and dropped the tell-tale em-dashes and “delve deeper” jargon. So I became bolder and started using them for something I thought required real intelligence: assessing my own drafts of papers, grants and presentations. Basically, I was using them as a referee on demand. This ability has not gone unnoticed, and tools like Q.E.D. [1,2] are now built precisely to review manuscripts and grants as if the fiercest reviewers on the planet were reading them. I tried it a few times, and I was genuinely impressed. What was truly tantalising was the possibility of having someone with whom to bounce ideas and check consistency without waiting for my colleagues’ feedback, which I know impinges on their already brutal schedules: an omniscient reader, always available. I was not alone in this: in a large analysis, the points raised by GPT-4 overlapped with those of human reviewers about as much as the points of two human reviewers overlap with each other, and more than half of the surveyed authors found its feedback helpful [3]. Then again, LLMs are trained on human-written text, so perhaps this is not surprising.

Down the rabbit hole

Then something started to creep in. Some critiques were humanly feasible. Others relied on an “excess” of logic, asking for experiments that would be extremely difficult to carry out and that perhaps no scientist would have thought of, either because they require a superhuman grasp of the grant or because they are notoriously infeasible. You may argue that this is a good thing: human brains are limited, and we do miss things, so it is good that something helps us fill these gaps. I thought so too, and I tried to address these critiques, adding the suggested experiments and caveats. Each critique was logically impeccable, so it could not be dismissed, yet satisfying it would take time (for experiments, months of work) without changing the central message for a human reader. So I answered it in the text, with another control, another qualification, another sentence, and that triggered the next critique. I spent hours refining the narrative so that the model would stop raising major concerns, and I was left with a piece so complex that only another LLM could distil its message. Still more critiques kept coming: it was a never-ending job. As often happens with digitalisation, instead of saving time, which was the initial purpose, we end up spending more time on activities we never envisioned: the illusion of optimisation and the perfectionist trap, leading to overcomplication.

Whose paper is it?

The worst part was not the lost time. It was the loss of ownership. In trying to satisfy every logical objection, the piece I was going to write morphed into something else. It was no longer “mine”. It had become a potpourri of sentences that make perfect sense to an LLM and little to a person reading them. Part of the problem was cognitive offloading: I often added corrections and new passages without fully digesting them because I trusted the LLM, even when its critiques were hard to grasp. That cognitive offloading matters because, for scientists, writing is not merely a transactional activity; it also helps us digest concepts, structure thoughts, and distil ideas. After going back and forth with model-driven changes, I realised I had lost the meaning of the content, and I felt drained and, truly, powerless. You may now argue that what I was writing was flawed and that the AI was largely correct. Most of the objections were indeed valid. The problem is that answering all of them turned a text written for people into a text written for a model. Correctness and communication are different goals, and without noticing, I had swapped one for the other.

I see the same signs more and more in the papers and grants I review: documents that are no longer accessible, strings of perfectly logical sentences with dehumanised content, and figures and schemes that would take minutes to decipher. At the beginning of the LLM era, documents were beautifully written but largely meaningless, beautified noise, as most of us experienced. Now they are much better, but still difficult to grasp. Maybe I am using LLMs wrong, but this makes me think it is not just a personal quirk, since several colleagues have told me the same. The minute we concede writing, experimental design and critique to them, we no longer have a product that is based on us. What I felt was alienation.

Taking back the pen

I am not arguing that we should abandon these tools. I am arguing for using them deliberately, and here are a few habits I am trying to adopt:

  • Write the first full draft yourself, so that the voice and the logic are yours.
  • Limit the number of LLM review rounds, and decide in advance when you will stop.
  • Triage every critique by scientific relevance and feasibility before acting on it. A correct objection is not necessarily one worth answering.
  • Use a human reader as the final test of clarity.

These are personal observations, not data, and whether such habits are enough is a question the community could usefully study. There is, however, related evidence: a survey of 319 knowledge workers found that higher confidence in generative AI was associated with less critical thinking, whereas higher self-confidence was associated with more [4]. Owning our writing may be one way of keeping that self-confidence. What I do believe is that we still write, and do science, for humans rather than for LLMs. If we want to keep our agency and our attachment to these activities, we should be extremely cautious about over-relying on LLMs. Have you felt the same? I would love to hear how you are using AI in your own writing, and where you draw the line.

References

  1. Tran L. Q.E.D.: an AI tool for smarter manuscript review. The Scientist. 2025 Nov 18.
  2. QED Science. Available from: https://www.qedscience.com/ [accessed 15 November 2026].
  3. Liang W, Zhang Y, Cao H, Wang B, Ding DY, Yang X, et al. Can large language models provide useful feedback on research papers? A large-scale empirical analysis. NEJM AI. 2024;1(8):AIoa2400196. doi:10.1056/AIoa2400196
  4. Lee H-P, Sarkar A, Tankelevitch L, Drosos I, Rintel S, Banks R, Wilson N. The impact of generative AI on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In: Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). ACM; 2025. doi:10.1145/3706598.3713778

Transparency note: a large language model was used for proofreading and for checking references; the ideas and the voice are my own. Still, it is worth noting that I included some information, for instance some of the references, because Copilot suggested it.