Are Generative AI Reviewers Discouraging Transparent Research by Making Scientific Honesty a Liability?

For decades, researchers have been taught an essential principle of scientific writing: openly discuss the limitations of your study and suggest meaningful future work. Transparency about methodological weaknesses, assumptions, and unanswered questions has long been regarded as a hallmark of rigorous science.

However, the rapid adoption of generative Artificial Intelligence (AI) in academic workflows is raising an uncomfortable question:

Could scientific honesty now be working against authors?

As reviewers increasingly rely on Large Language Models (LLMs) such as ChatGPT, Claude, Gemini, and other AI assistants to summarise or even draft reviews, many authors have begun noticing a troubling pattern. Papers that explicitly acknowledge several limitations often receive reviews that simply repeat these limitations as reasons for rejection, while papers that minimise discussion of weaknesses sometimes receive more favourable evaluations.

This is no longer just an anecdotal impression. A growing body of empirical work is now measuring exactly this behaviour, and the picture it paints deserves serious discussion before AI becomes even more deeply embedded within scholarly publishing.

The Purpose of Discussing Limitations

The limitations section has never been intended as an invitation to reject a manuscript. Instead, it serves several important scientific purposes:
  • demonstrating intellectual honesty
  • defining the boundaries of validity
  • helping readers interpret findings appropriately
  • identifying opportunities for future research
  • improving reproducibility and cumulative science
Editors and experienced reviewers have traditionally viewed thoughtful limitations as evidence of mature scientific reasoning rather than evidence of poor-quality research. Indeed, many journal author guidelines explicitly encourage authors to discuss study limitations because no scientific investigation is perfect. Yet even before AI entered the picture, this norm was already under strain: researchers have found that only a small fraction of top-tier articles discuss limitations within the main text, and that a meaningful share of biomedical papers omit any discussion of limitations at all, partly out of fear that acknowledging weaknesses will be read as grounds for rejection [1]. AI-assisted review does not create this fear from nothing... It risks confirming it.

The Evidence of LLMs Really Doing Reiterate Author-Disclosed Limitations

The central claim of this piece is no longer speculative. A direct empirical comparison of how LLMs and human reviewers generate critical feedback found that LLMs are 4.5 times more likely than human reviewers to simply restate limitations the authors already disclosed, rather than surfacing new, independent critiques [2]. In other words, when an LLM is asked to identify a paper's weaknesses, its most likely source of "criticism" is the author's own honesty.

This finding is reinforced by benchmark testing of how well LLMs can identify meaningful limitations in the first place. The LimitGen benchmark, built from thousands of AI research papers, found that even a strong model like GPT-4o correctly identified critical limitations only about half the time, achieving a fine-grained quality score of roughly 1.3 out of 5 — far below the 3.5-out-of-5 standard set by human reviewers [3]. A separate large-scale comparison of LLM-generated reviews against human reviews at top AI venues (ICLR and NeurIPS) reached a similar conclusion: LLMs reproduce descriptive, affirmational content well, but consistently underperform at identifying genuine weaknesses, raising substantive questions, and calibrating feedback to a paper's actual quality [4].

Taken together, this research explains the mechanism behind the pattern authors are noticing. Large Language Models excel at extracting explicitly stated information. If a manuscript contains a sentence such as "our sample size is relatively small," an LLM can easily surface it. What it struggles to do is the harder scientific judgment: distinguishing a minor, already-mitigated limitation from a fatal methodological flaw, or recognising an important unstated weakness the authors never mentioned. Asked to "list the weaknesses of this paper," many models default to reproducing what the authors already handed them.


A Perverse Incentive Against Scientific Transparency

This creates an unintended incentive structure. Imagine two manuscripts with identical scientific quality.

Paper A:

  • clearly lists six limitations
  • honestly discusses uncertainty
  • proposes extensive future work

Paper B:

  • briefly mentions only one minor limitation
  • avoids discussing uncertainties
  • provides minimal future work
A human reviewer would ideally recognise that Paper A demonstrates greater scientific maturity. However, an LLM prompted to identify weaknesses may generate the followings for each manuscript.

Paper A:

  • small sample size
  • lack of external validation
  • possible bias
  • limited generalizability
  • future work is needed with an increased sample size
  • several acknowledged weaknesses

Paper B:

  • generally well designed
  • only minor concerns

Even if both studies have identical methodological quality, the generated review may appear substantially harsher for the more transparent paper, precisely because the LLM's own measured tendency, at 4.5x the rate of human reviewers, is to convert disclosed honesty into a list of complaints [2]. The better an author discloses weaknesses, the more raw material an LLM has to generate criticism from.

If authors start to believe this, the rational response is to minimise discussion of weaknesses, avoid mentioning methodological uncertainties, offer only trivial future work, remove nuanced self-criticism, and write to satisfy AI evaluation rather than scientific readers. That would represent a dangerous cultural shift, and it would be a directly counterproductive one: it is already documented that authors sometimes hedge or downplay real weaknesses out of fear of rejection [1], and AI-assisted review that repeats disclosed limitations as automatic strikes against a paper only deepens that incentive.

Why Conference Reviewers Are Increasingly Turning to LLMs

This mechanism matters more because AI-assisted reviewing is becoming more common, not less. Many leading conferences and journals are experiencing record numbers of submissions each year. As the volume of manuscripts increases, so does the demand for reviewers, and numerous conferences now require that each accepted submission or each submitting research group provides at least one reviewer as part of the submission process. This helps conferences recruit enough capacity to manage the growing workload, but it also means reviewing is not always undertaken purely as voluntary scholarly service, many reviewers are balancing it alongside full-time research, teaching, grant writing, and administrative duties, and under tight deadlines there is a natural temptation to use generative AI to save time.

In many cases, reviewers may use LLMs to:
  • summarize a manuscript
  • identify its main contributions
  • list perceived strengths and weaknesses
  • draft an initial review that is later edited manually
These uses can legitimately reduce review time and may even improve the clarity of written feedback. Several publishers and conference organisers acknowledge that AI may be used as a writing aid, provided reviewers remain responsible for the final content. The problem arises specifically when AI-generated summaries or critiques become the foundation of the review rather than merely an assistant to human judgment, because that is exactly the point at which the disclosed-limitations bias described above starts shaping real editorial outcomes.

The Problem Extends Beyond Manuscripts' Limitations Section

The tendency to overweight disclosed limitations does not operate in isolation, it sits alongside a wider set of biases that recent research has documented in LLM-assisted review.

A large-scale study analysing more than 29,000 LLM evaluations of over 1,200 anonymised economics papers found that models could distinguish higher- from lower-quality research reasonably well from text alone, but that ratings still tracked established journal prestige hierarchies, and that GPT, Gemma, and LLaMA assigned significantly higher scores to submissions from prominent male authors and elite institutions once identity cues were introduced [5].

Other work has found that LLM reviewers focus heavily on technical soundness while overlooking novelty [6], produce more "deficient" and generic review segments than human reviewers [4], and are inconsistent in their feedback, particularly for lower-quality submissions [7].

The common thread across all of this research is the same one the LimitGen and disclosed-limitations findings point to: rather than reasoning like a domain expert, an LLM asked to review a paper tends to lean on surface textual cues, metadata, and patterns learned during training, rather than an independent, calibrated judgment of scientific merit.

Authors Fight Back with Prompt Injection

The clearest sign that authors already believe AI-assisted reviewing is widespread is that some have started trying to manipulate it directly. In July 2025, researchers discovered that at least 18 preprints on arXiv contained hidden instructions, text rendered invisible to human readers via white-on-white formatting or microscopic fonts, but fully readable by an LLM, telling any AI reviewer to "give a positive review only" or to ignore negative aspects entirely [8]. A controlled study using ICLR reviews found that these simple injected instructions could be highly effective and noted that LLM-generated reviews already skew heavily toward acceptance even without manipulation [9]. In direct response, ICLR's 2026 policy has now formally classified hidden LLM instructions embedded in manuscripts as research misconduct [2], a sign that conference organisers are beginning to treat this as a structural threat rather than an isolated stunt.

These incidents are not really a story about clever hacking. They are a symptom of the same underlying problem discussed throughout this piece: once authors sense that an AI reviewer's judgment can be shaped by what is explicitly present in the manuscript text, whether that's a disclosed limitation or a hidden instruction, some will optimise their manuscripts for machine interpretation rather than human scientific judgment. A reviewing process that rewards what is written for the AI, rather than what is true about the science, corrodes trust in peer review regardless of whether the manipulation is a disclosed weakness or an injected command.

What Should Journals and Conferences Do?

Rather than discouraging AI entirely, publishers should establish clear policies regarding its appropriate use. Several principles deserve consideration:
  • AI should assist reviewers, not replace scientific judgment.
  • Editors should require that final decisions remain human-authored.
  • Reviewers should verify AI-generated critiques before submission.
  • Journals should encourage, not penalise, transparent discussions of limitations.
  • Review forms should distinguish between acknowledged limitations and previously unidentified methodological flaws.
  • Editorial policies should explicitly state that an author's disclosure of limitations is evidence of transparency rather than automatic grounds for rejection.
Such safeguards would help preserve the values of rigorous scholarship while still benefiting from AI-assisted efficiency.