Rejected Because an LLM Could Write It Better? What Peer Review Actually Weighs

A computer-science advisor gave a student a blunt rule for the LLM era: your submission will be rejected if a reviewer can generate a better paper by using an LLM. The student took the line to Academia Stack Exchange and asked whether it hid red flags, and if so, how red.

The worry under the question is a ratio. The student had spent a year writing one paper; a reviewer, the fear goes, could prompt a superior one in five minutes. If that were the real bar, the asker reasoned, the rational move would be to generate a paper every day and collect more than 300 publications a year.

The two most-upvoted replies neither dismiss the advisor nor fully back the student. One calls the rule a statement of the obvious and pins down the one place it genuinely breaks. The other argues the real test is broader than any single prompt and spells out what a reviewer should check before saying no. Between them they draw a usable line for anyone submitting research while generative-AI tools sit on every reviewer's desk.

The student's objections, taken seriously

The asker raised a stack of objections, and most are reasonable on their face. First, an asymmetry: as far as the asker knew, many publishers forbid AI-generated manuscripts. So a human-written paper could be rejected for losing to a machine-written one that the same publisher would never print. Can the two even be compared?

Second, conflict of interest. A reviewer who reads a paper and then generates a similar one now holds a competing manuscript. If that reviewer judges the generated version better, they are, in the asker's phrase, grading their own homework. The last objection is the most principled: a review should rest on the merit of the submission, not on the merit of some other article that has never been peer-reviewed.

The case that the rule is simply obvious

The top answer, with 39 votes, sees no red flag. Its logic is direct: if a reviewer can type “prove the Riemann hypothesis” into ChatGPT and get something better than your paper, your paper has no value and should be rejected. Nothing about that depends on language models specifically.

Its central example predates them. Suppose a paper presents a closed-form solution to a gnarly integral, and a reviewer types the integral into Mathematica and gets the same result in five minutes. That paper probably shouldn't be published. Computer algebra set that floor long ago; LLMs extend the same question from integrals to proofs, prose and code. On this reading, the advisor was describing a standard reviewers already apply, with a new tool attached to it.

Where the test breaks: your ideas in the prompt

The same answer names the student's strongest objection, and it is neither the publisher asymmetry nor the homework problem. It is information leakage. The test only works as an apples-to-apples comparison if the reviewer prompts the model cold. If the reviewer feeds in the paper's insights, the output borrows exactly what made the paper worth writing, and the comparison is void.

This, too, is older than LLMs. The answer points to mathematical fields where finding integers with certain properties is very hard, while checking that a given candidate has them is easy. A reviewer who has seen your integer can confirm it in seconds, and that speed says nothing about how hard it was to find. Swap the integer for a proof idea, a training trick or an architecture, and the same trap catches any reviewer who prompts a model after reading the submission.

From leakage to theft

Leakage can be innocent. A reviewer who has read your paper cannot unread it, and every prompt they write afterwards is coloured by it. Theft is something else. The top answer draws the line plainly: if the advisor meant that a reviewer might steal your idea and publish it themselves, using an LLM to reproduce the results quickly, that is obviously a red flag. Out of context, though, the answerer would not jump to that conclusion.

The student's conflict-of-interest point lands hardest here. A reviewer holding a near-copy of your paper, generated from your own submission, is no longer a neutral judge of it. Neither answer suggests reviewers routinely behave this way; both read the advisor's sentence as a statement about the quality bar rather than about reviewer conduct. The distinction matters because the two call for different responses: the first is a question of how strong your paper is, the second is a question of misconduct by the person judging it.

What a careful reviewer would actually check

The second answer, with 12 votes, says the situation will be more nuanced than the advisor made out. Its author would not reject a paper merely because a frontier LLM can prove a stronger result, even with a simple prompt and thinking time on the order of minutes. Rejection would need a broader set of observations, holding together.

Those observations are specific: the result is not unexpected; the question is one anyone could ask; an LLM gives a satisfactory answer within minutes; the proof in the paper is essentially the LLM's argument; and if that argument contains some surprising step, the paper fails to bring that central idea to the forefront. Any one of these alone is not enough.

Even then, the answer does not call such work worthless. Arguments that fit every criterion may still deserve to be archived, just not in a journal: as preprints, on GitHub or in a dedicated repository. That is a softer outcome than the advisor's phrasing implies. The work still exists and can be found and cited; it simply doesn't carry the weight of a journal publication.

Someone still had to point the telescope

The second answer's most memorable argument comes from astronomy. An amateur astronomer makes a genuinely interesting observation, say some strange new variability in a star's output, like Tabby's star, and writes it up well. Professional instruments could immediately gather better data, and a better paper could then be written. The amateur's paper may still deserve to be published.

The principle underneath: low resources, and even relatively low effort to complete a work, do not make that work scientifically uninteresting. And if effort was never the measure, then the ability to replicate the work with even less effort means even less. As the answer puts it, someone still had to point their telescope at the right patch of the sky at the right time.

Applied to LLMs, this cuts against the strict version of the advisor's rule. A model that can elaborate a result once someone has asked the right question has not asked it. Being first to the question, or first to notice the anomaly, can itself be the contribution that earns a place in the literature.

What counts as a contribution

None of this lowers the bar. The same answer is clear that a publication in a good journal must make a contribution to knowledge that is valuable to the community. What it widens is the list of things that can count as one:

• a good question;
• a surprising answer;
• an answer obtained through perseverance;
• a particularly enlightening write-up;
• a new method;
• the luck of being first to point the telescope at the right patch of sky, or to notice a patch of mould crowding out the bacteria.

Notice what is missing: hours spent. A year of work is not on the list, and a five-minute prompt is not a mark against you either. That answers the student's sharpest fear directly. The cost of producing a paper, by hand or by model, is not what reviewers are meant to weigh; what the paper adds is.

The bar that does move: competition

Here the second answer concedes the core of the advisor's warning. However your paper came into existence, it will compete for attention with LLM-generated or LLM-aided papers. If it loses to them on quality, that is a barrier to publication, and it doesn't matter whether the answer to your question can be obtained with a single prompt.

Acceptance to a good journal, the answer says, depends on convincing reviewers that the work adds to knowledge more meaningfully than most other submissions to the same journal. If that is what the advisor meant, the answerer agrees. The student's framing, my paper against a reviewer's generated paper, gives way to something more ordinary and harder: my paper against the whole submission pool, where generative-AI assistance is now part of the field.

Limits and open questions

The thread settles less than it seems to. Neither answer addresses the asker's claim that many publishers forbid AI-generated manuscripts, so anyone leaning on that should read the target journal's or conference's policy on generative AI directly rather than assume. Those rules are set venue by venue.

Two harder questions stay open. First, an author usually cannot see how a reviewer prompted a model, so there is no easy way to know whether a comparison was run cold or seeded with the paper. Second, the second answer's checklist leans on what an LLM can answer within minutes, and that shifts as models improve. A result that passes the test today may fail it on a later model, which makes the durable parts of a contribution, the question, the observation, the surprising idea brought to the front, worth more than the derivation that follows them.

This article draws on “Your paper will be rejected if a reviewer can generate a better paper by using LLMs.” and answers to it (1, 2), on the Stack Exchange network, licensed under CC BY-SA 4.0.