The rise of large language models (LLMs) has triggered a series of unintended consequences for the global research community, most notably an avalanche of AI-enabled, low-quality academic papers.
We are no longer discussing a theoretical risk; recent large-scale analyses across scholarly databases including Elsevier’s Scopus, Web of Science, and arXiv indicate a staggering shift. By early 2025, researchers noted that nearly 20 per cent of computer science preprints and over 10 per cent of papers across STEM databases showed linguistic signatures of LLM assistance.
Some estimates suggest that scientists using these tools have increased their individual output by as much as one-third – creating a volume of submissions that the current system was never designed to absorb.
Submission acceleration
Critically, this is not just an issue of editorial polishing; AI has drastically lowered the cost of entry for industrial-scale paper mills which now use LLMs to churn out fabricated but plausible-looking studies at a rate that threatens to erode the foundational trust upon which scientific progress is built. However, the rapid increase in publications exhibiting LLM signatures should not always be interpreted as a proxy for declining research quality; rather, it reflects the growing integration of AI into mainstream scientific workflows, which may equally drive improvements in rigour and productivity.
This surge is pushing the peer review system toward a point of unsustainability. Since the mid-17th century, when the Royal Society first published Philosophical Transactions, peer review has been the bedrock of research quality.
Today, however, the sheer scale of AI augmented output forces us to ask a fundamental question: is the traditional peer-reviewed paper still the most effective forum for sharing research advances? The challenge is as much about supply as it is demand; as submissions accelerate, the pool of qualified reviewers already stretched and uncompensated is shrinking, leaving the gatekeepers of science overwhelmed by machine-generated mediocrity.
Publishers are not standing still. Houses like Springer Nature and Elsevier are investing in AI tools and human expertise to detect machine-generated content, to facilitate more accurate desk rejections before peer review. While these technological safeguards are a welcome and necessary layer of defence, they are not a silver bullet. Detection is an arms race, and we cannot simply “tech” our way out of a challenge that is as much about academic culture as it is about software. An over-reliance on automated detection risks baking in a new fluency bias, where non-native English speakers are disproportionately flagged for using the formal structures that AI often mimics.
Culture change
To move beyond a purely defensive posture, we all have a role to play in advocating for a shift from a culture of volume to a culture of value. This involves rethinking the incentives that drive research behaviour, though we must recognise that the right approach will vary significantly across academic disciplines. What constitutes a healthy use of generative AI in a STEM lab will look very different from the interpretive, long-form traditions of the humanities. Nevertheless, we can collectively champion several key shifts:
Supervisors and mentors: We should move away from setting fixed publication targets for PhD students to mitigate the publish or perish culture. In the humanities, this might involve focusing on the depth of a single chapter or monograph; in STEM, it might mean prioritising methodological robustness and research impact over publication volume.
This shift is not intended to limit student’s ability to publish as they establish their careers. Rather, it ensures that the quality of their work is recognized and effectively demonstrated. By focusing on the merit of the work rather than the quantity of output, the system can better support early-career researchers in producing high-quality, validated contributions to their fields.
Examiners: We must ensure that peer reviewed publications and outputs are not treated as a primary metric for passing a PhD viva. While peer reviewed publications provide valuable independent validation and additional assurance of research originality and contributions to knowledge, the candidate’s domain expertise, the influence of their research outcomes, and the integrity of their specific contribution must remain the priority.
Funders and reviewers: While we have begun to see a welcome shift toward narrative CVs and holistic assessments, this transition needs to accelerate. A decisive pivot is needed, away from counting outputs and toward assessing the long-term scientific and societal impact of the work and broader contributions.
University leadership: The Research Excellence Framework has introduced an important shift towards prioritising the quality of outputs over volume, and university leadership could mirror this transition by explicitly prioritizing quality over quantity in evaluations, promotions, and recruitment. This could involve a focus on workload models that provide academics with the essential thinking time necessary for high-impact work, ensuring that institutional standards value the rigor of individual publications over the rapid, incremental dissemination that AI facilitates.
Publishers and editorial boards: They must transition from passive recipients of content to active stewards of research integrity. This requires a strategy of both professional mentorship and robust mandate. To achieve this, boards should use international conferences and workshops as strategic platforms to set the global standard for publication culture. By exercising thought leadership in these forums, journals can move from merely setting rules to actively incentivizing transparent AI disclosure and documenting high-integrity tool usage.
Simultaneously, boards must enforce stricter sanctions for systemic malpractice. By combining global outreach with rigorous enforcement, they ensure the opportunity of publication remains reserved for those who uphold the standards of the international research community.
A different future
In parallel, we must provide graduate students and early career researchers with the training required to build a healthy, transparent relationship with LLMs. This isn’t about bans; it’s about fostering a intentionally designed partnership grounded in domain expertise, critical thinking, and intellectual honesty.
Ultimately, these proposals won’t solve everything in isolation. The answer to the AI challenge lies in a complex combination of cultural shifts, the responsible use of technology, and a fundamental realignment of incentives.
As we look further ahead, we must also be open to more radical evolution. Just as AI is fundamentally changing the initiation and execution of research, it will inevitably transform how we communicate it, perhaps moving us more toward living digital documents and open-data repositories rather than static PDFs.
The traditional academic paper has served us remarkably well since the 1600s, but change is both inevitable and healthy. We should embrace the possibility that the future of scholarly communication may look very different from the past, ensuring that our methods of sharing knowledge evolve alongside the intelligent tools we use to create it.
