Repeated exposure to unchecked submissions may make reviewers more suspicious of the next paper. An ambiguous explanation or a correctable citation error could then be treated as evidence that the authors did not check their work. The concern is that careful authors could lose the benefit of the doubt when they make an ordinary mistake.
Here, AI slop means low-effort AI-generated material submitted without adequate author checking. Poor-quality work alone does not establish AI use. The hypothesis examined here is that the burden of checking such material may leave reviewers with less time and patience for other submissions.
The public accounts examined here cannot establish whether AI submissions caused a particular harsh score or made sound work more likely to be rejected. This note examines what those accounts and the official guidance support, then sets out how authors can make factual corrections easier to assess.
1. Why some reviewers say they are losing patience
The ICLR program chairs' explanation of the 2027 submission policies[1] describes a shortage of expert attention in a rapidly growing field. It also raises the problem of submissions whose authors cannot verify their own claims. The chairs acknowledge that restrictions intended to manage this burden impose costs on some legitimate researchers.
There are concrete examples of work that needs checking. Nihar Shah's collection of placeholder statements and unfinished material[2] links ICLR submissions containing provisional results and TBDs. He explicitly warns that these observations do not establish AI use, whether experiments happened, or intentional misrepresentation. The examples are a reason to inspect the evidence, rather than a count of fabricated papers.
Reviewers describe that effort in their own words. An October discussion[3] includes complaints about missing baselines and unsupported results, alongside reports of strong papers. In another thread,[4] a reviewer says they spent four hours reviewing before learning the paper had been desk rejected. The reason for that rejection is not established.
A reviewer spending that time has a reasonable interest in whether the authors checked the work before submitting it. The concern is what happens when justified caution follows them into the next assessment.
2. When suspicion starts to shape the assessment
One exchange illustrates the concern. In an ICLR discussion,[5] a commenter describing their reviewing experience says suspected AI-generated writing is off-putting:
“even if the findings are legitimate, the writing puts me off”
A reply[6] argues that legitimate findings still deserve assessment. Another participant in that exchange makes a useful distinction: writing that is genuinely difficult to understand warrants revision, regardless of how it was produced.
This exchange does not establish the papers' authorship or the correctness of their findings. It does show a disagreement about whether suspected AI use should influence the assessment when the underlying work might be sound.
One hypothesis is that repeated exposure to poor submissions makes that suspicion harder to set aside. A reviewer might spend less time investigating an ambiguous explanation, or interpret an ordinary citation mistake as proof that the authors never understood the underlying work. Those are plausible behavioral consequences to examine, not measured outcomes from these comments alone.
An honest mistake can still be serious. A result without adequate evidence remains a problem even if nobody used AI. What matters is whether the specific error undermines the claim, whether it can be corrected, and whether that correction receives consideration.
3. Fatigue does not always mean harsher scores
In their reflection on the NeurIPS 2025 review process,[7] the program chairs describe finite reviewing bandwidth and report some reviewers raising public scores to end prolonged exchanges while remaining critical privately. They interpret that behavior as a sign of fatigue.
That is an organizer observation from a different conference cycle. It does not establish an effect of AI submissions. It also points in a different direction from the concern about excessive harshness.
The risk worth examining is whether workload and distrust weaken the connection between evidence, discussion and the final recommendation. We should not assume the effect always favors acceptance or rejection.
4. What authors say they are dealing with
One author reports 4/2/2/2 and says a review describes a different paper.[8] The underlying review is not included, so we cannot check that claim here. The same person repeats it in another thread; that remains one report.
Another author says a reviewer asked for a comparison with an earlier version of their own work.[9] They had changed the title and method name, which could explain the confusion. We cannot establish the underlying review or the earlier work's publication status from the post. It is still a useful example of a question that needs clarification before a scientific comparison can proceed.
An author reporting 3/1/3[10] describes a different situation: the lowest-scoring reviewer questions novelty, while the other concerns seem easier to resolve. A novelty disagreement requires an argument about the contribution. It is not necessarily a factual mistake, and disagreement between scores does not identify which assessment is right.
There are good experiences too. An author reporting 3/3/2 thanks a reviewer[11] for feedback specific enough to open a new research direction. That is a useful standard: criticism that engages with the work and helps the author understand what matters.
These are selected public accounts, not a survey. They illustrate concerns about the process but do not establish how common poor reviewing is or whether AI submissions caused it.
5. What careful assessment should look like
A reviewer who has made a factual mistake should acknowledge the correction and explain whether it changes the recommendation. Other objections may still justify rejection. The author should be able to understand which objections remain and how their response was considered.
The ICLR reviewer guide[12] asks reviewers to concentrate on issues that affect their recommendation, participate in discussion, and explain what did or did not change their final judgment. Requests for additional experiments should remain limited in scope. The AC guide[13] asks area chairs to encourage responses to rebuttals, intervene when discussion receives insufficient attention, and respond promptly to reasonable author inquiries.
A useful reply might confirm that a concern was resolved, explain why it remains, or identify a separate objection. Even a rejection becomes easier to understand when the reasoning follows the evidence.
Scores cannot supply that explanation. A mean or percentile does not show whether a reviewer considered a correction, and a percentile is not an acceptance probability. Guessing whether a review was AI-written cannot settle its scientific objections either. ICLR's reviewer AI policy[14] requires disclosure when reviewers use LLMs to generate or edit review text. Reviewers must also report their original self-written assessment and their interactions with the model; authors are not shown that initial assessment.
6. Make the disputed point easy to check
For rebuttal, organize each disputed point by the kind of response it needs.
If the review gets a fact wrong, point to the submitted version. Give the exact statement, the location of the contrary evidence, and the question you want reconsidered. Showing that a result was already present is different from adding it after the review.
If your own explanation caused confusion, acknowledge that. Say what was unclear, how you changed it, and whether the clarification affects the claim. Calling it an honest mistake does not explain its consequences.
If the reviewer identifies a real evidence gap, address the gap. Provide the analysis you can support, narrow the claim where necessary, or explain the limitation. A complaint about the process cannot substitute for missing evidence.
If the objection concerns novelty, identify the comparison. Establish which earlier work and version the reviewer means, then explain the precise difference in claims, assumptions or results. A changed title alone does not establish a new contribution.
The following are constructed examples, not excerpts from actual reviews or rebuttals:
An overlooked result: “The review asks for an ablation without component X. Table 3 in the submitted version reports that experiment. We have made the reference clearer in Section 4. Does this address the concern, or is a different ablation needed?”
An omission in the paper: “We omitted the preprocessing detail needed to reproduce this result. We have added it in Section 3 and explained which experiments it applies to.”
A novelty comparison: “The cited manuscript is an earlier version of this work. The revision adds Y; the original result remains X. We would appreciate clarification on whether the concern is about the contribution itself or the relationship between the versions.”
Use such wording only when it accurately describes your paper. The point is to leave a disagreement that another reader can inspect.
7. If the correction gets no response
Ask the area chair for a specific intervention. Include the disputed claim, the evidence location and the response already posted. Explain why the unresolved point matters to the assessment.
An illustrative message would be:
“Reviewer 2 states X. Our response points to Section 3 and Table 2 of the submitted version, which address that statement. Could you help clarify whether this concern remains part of the recommendation? We would appreciate the reviewer's assessment of the evidence.”
That request gives the AC a concrete issue to consider. It does not guarantee another review, a score change or acceptance.
Use the feedback that improves the paper, and explain the changes you made. Favorable scores do not remove the need to answer a substantive objection.
8. Check the references you add while responding
A rushed response can introduce another source of confusion. Before adding a requested citation, open the work, confirm its identity and version, and read the passage relevant to your argument. A real DOI can still be paired with the wrong title. A correct bibliographic record does not establish that the paper supports your sentence.
The existing rebuttal citation guide covers those checks. StrictCite can compare bibliographic details with scholarly records; deciding what a source proves still requires reading it. The tool cannot determine whether a review is fair or predict a conference decision.
Inspect the evidence returned for each finding.