By Jim Shimabukuro (assisted by Claude)
Editor
The scientific enterprise stands at an inflection point. Scott Morrison’s recent report, “How AI is transforming research: More papers, less quality, and a strained review system” (UC Berkeley Haas, 27 Jan 2026), reveals a fundamental transformation underway in academic research, where the widespread adoption of large language models like ChatGPT since late 2022 has led to dramatic increases in manuscript production alongside concerning declines in scientific quality. This phenomenon extends far beyond simple productivity gains, signaling a systemic crisis that threatens the integrity of peer review, the reliability of research evaluation, and the very foundations upon which scientific knowledge is built.
The empirical evidence is striking. Research examining over two million papers across three major preprint servers—arXiv, bioRxiv, and the Social Science Research Network—found that scientists using AI tools increased their output by more than fifty percent on bioRxiv and SSRN, and by over one-third on arXiv. These gains were most pronounced among researchers for whom English is not a first language, with some groups experiencing productivity increases approaching ninety percent. The technology appears to be democratizing access to high-level scientific communication, potentially shifting global research power dynamics away from traditionally English-dominant institutions.
Yet this productivity surge masks a troubling reality. The Berkeley study discovered an inverse relationship between writing sophistication and quality for AI-assisted papers—while human-written papers with more complex language tended to be stronger, AI-assisted manuscripts with sophisticated prose were actually less likely to be published in peer-reviewed journals. The implication is profound: the traditional signals that editors and reviewers have relied upon to identify rigorous research—clear, precise, sophisticated language—have become unreliable or even actively misleading in the age of AI-generated text.
This decoupling of writing quality from scientific merit creates what might be called an evaluation crisis. For decades, the peer review system has operated on certain assumptions about the relationship between how research is communicated and how it was conducted. Papers with polished, sophisticated language have historically been cited more frequently, based on the premise that teams capable of articulating complex ideas precisely likely have command of their subject. Now that AI can produce prose that is more linguistically complex than many human scientists can write, this heuristic has broken down entirely. Reviewers face an unprecedented challenge: distinguishing between papers that are merely well-written by machines and those that represent genuine scientific contributions.
The scale of this challenge becomes clearer when we examine how AI has permeated the peer review process itself. A December 2025 survey of sixteen hundred academics found that more than fifty percent now use artificial intelligence tools while reviewing manuscripts, often in direct violation of journal guidelines. Most of this usage remains superficial—drafting reports, polishing language—rather than enhancing rigor through statistical validation or methodology checks. The result is a recursive loop where AI-generated papers are evaluated by AI-assisted reviewers, with human judgment increasingly marginalized in both creation and evaluation.
Recent experimental evidence demonstrates both the promise and peril of fully automated research production. The AI Scientist system generated papers entirely end-to-end, from formulating hypotheses through writing manuscripts, with one paper passing peer review at an ICLR 2025 workshop and receiving acceptance scores clearly above the threshold. Yet this milestone raises as many questions as it answers. If AI can independently produce work that passes traditional peer review, what does this say about either the reliability of AI-generated research or the adequacy of current evaluation standards? The answer likely involves uncomfortable truths about both.
The infrastructure of scientific publishing is already buckling under the strain. Preprint servers are feeling the crush, with arXiv announcing in October 2025 that it would no longer host computer science review and position papers unless they had already undergone peer review, citing an uptick of suspect papers. Similarly, bioRxiv and medRxiv have implemented AI review tools to generate rapid feedback on preprints, attempting to use technology to manage the flood of technology-enabled submissions. These are defensive measures, attempts to stem a tide rather than harness it productively.
Beyond the immediate crisis lies a deeper question about the future structure of scientific communication. Some researchers propose restructuring scientific papers to include machine-readable structured appendices that would transform manuscripts from static documents into queryable, executable research environments, enabling AI systems to directly validate computational reproducibility and cross-reference findings. Such infrastructure would fundamentally enhance AI-assisted peer review by grounding evaluations in verifiable, executable claims rather than ambiguous prose. This represents a potential path forward: adapting the publication system itself to make AI assistance more transparent and verifiable rather than attempting to exclude it.
The political and economic dimensions of this transformation cannot be ignored. The National Institutes of Health announced in 2025 that a federal public access policy would take effect six months earlier than planned, with NIH director Jay Bhattacharya calling for a crackdown on article processing charges that can reach nearly thirteen thousand dollars. The collision of AI-driven publication growth with ongoing debates about open access and publication costs creates additional instability. If AI dramatically increases paper output while funders simultaneously restrict spending on publication fees, the economic model supporting much of scientific publishing faces an existential challenge.
Looking toward 2026 and 2027, several trajectories appear likely. First, we should expect continued rapid growth in submission volumes, but with increasing stratification in how different venues respond. Prestigious journals may implement more stringent screening mechanisms—potentially including specialized “reviewer agents” that filter submissions before human review—while lower-tier venues become overwhelmed. Some conferences have already begun implementing primary paper initiatives that charge submission fees unless papers represent an author’s only submission, attempting to discourage high-volume, low-quality submissions.
Second, transparency requirements will likely expand significantly. The question has shifted from whether researchers use AI to how they use it and for what purposes. Publishers are developing specialized tools like Veracity from Grounded AI, which checks whether cited papers exist and analyzes whether cited work corresponds to authors’ claims. More journals will likely mandate detailed disclosure of AI assistance, potentially requiring authors to specify which AI systems were used, for what tasks, and with what level of human oversight. Some disciplines may move toward requiring that all code, data, and AI prompts be made available alongside publications, making the research process more transparent and reproducible.
Third, we may see the emergence of parallel publishing ecosystems. There are calls for dedicated preprint servers specifically for AI-generated research, treating AI as a legitimate research contributor while ensuring complete transparency about its role. Such infrastructure could preserve the integrity of traditional venues while creating appropriate channels for machine-assisted research. The alternative—attempting to exclude AI entirely—appears both impractical and ultimately futile given the technology’s ubiquity.
Fourth, the criteria for evaluating research quality and researchers themselves will need fundamental revision. Universities and funding agencies that continue to rely primarily on publication counts will inadvertently reward exactly the behavior that undermines research quality—using AI to maximize output regardless of contribution. As one Berkeley researcher noted, institutions that begin experimenting now with new evaluation criteria, funding models, and approaches to verification will be better positioned than those that wait for the full impact to become undeniable. This likely means moving toward evaluation systems that prioritize demonstrated impact, reproducibility, and genuine novelty over raw publication metrics.
Fifth, the role of peer review itself will continue evolving. The progression from content generation to oversight is neither new nor unique to AI—academic careers already follow this trajectory, with junior researchers writing papers and senior researchers reviewing them. What changes is the scale and speed. Human peer reviewers may increasingly serve as meta-evaluators, overseeing AI systems that perform initial technical checks while reserving judgment about significance, novelty, and broader implications for human expertise. This division of labor could make peer review more sustainable if implemented thoughtfully, but risks creating a detached, perfunctory evaluation process if human engagement becomes too minimal.
The implications extend beyond academic publishing to the broader scientific endeavor. When traditional quality signals become unreliable and evaluation systems struggle to separate valuable research from polished noise, funding decisions become more difficult, replication efforts become harder to prioritize, and the accumulation of reliable knowledge slows even as paper production accelerates. The Stockholm Declaration, endorsed by scientists across disciplines in June 2025, called for fundamental reform of academic publishing, noting that hundreds of thousands of low-quality or fraudulent publications annually threaten scientific and economic progress.
The path forward requires acknowledging several uncomfortable realities. First, AI assistance in research writing is now ubiquitous and attempting to exclude it entirely is likely impossible. Second, current peer review mechanisms were not designed for and cannot effectively handle the volume and sophistication of AI-assisted submissions. Third, the incentive structures that drive researchers toward high-volume, low-value publication predate AI but are dramatically amplified by it. Fourth, the economic models supporting scientific publishing are increasingly unsustainable, particularly as submission volumes grow while public funding comes under pressure.
What emerges during 2026-2027 will likely be messy, experimental, and contentious. We can expect ongoing tensions between researchers who embrace AI as a powerful tool and those who view it as fundamentally corrupting the research process. We will see continued evolution of detection methods alongside AI systems designed to evade detection, a technological arms race that serves neither science nor society. We will witness experiments with new publication formats, evaluation criteria, and infrastructure—some promising, many failing. The question is not whether the research publication system will change, but whether it can change rapidly and thoughtfully enough to preserve scientific integrity while adapting to technological realities.
The “arms race” framing may be too pessimistic, or at least incomplete. A dialectical or recursive process—LLMs improving, guidelines adapting, LLMs adjusting again—could actually be productive rather than merely adversarial. This optimistic trajectory is plausible, though far from guaranteed.
The historical precedent is illuminating. When word processors replaced typewriters, when statistical software replaced manual calculation, when search engines replaced library card catalogs, each technological shift initially created anxiety about degraded quality and lost skills. Yet in each case, the scientific community eventually developed norms, practices, and infrastructure that integrated the new tools while preserving or even enhancing research integrity. The question is whether AI-assisted research writing follows this pattern or represents something fundamentally different.
There are several reasons to think a productive equilibrium might emerge. First, the transparency that comes from ongoing scrutiny could actually improve the research process. If journals require detailed disclosure of AI usage, if reviewers can query exactly how claims were generated, if readers can trace the provenance of arguments back through machine-readable structured data, we might end up with research that is more accountable than what preceded it. The current system relies heavily on trust that authors accurately represent their methods and findings. A future system where AI assistance is documented and verifiable could reduce rather than increase opportunities for deception.
Second, LLMs learning to meet evolving publication standards could drive genuine quality improvements if those standards are well-designed. Imagine guidelines that require papers to explicitly state hypotheses before methods, that demand clear logical connections between claims and evidence, that insist on appropriate statistical power calculations, that flag overclaimed conclusions. If AI systems are trained to meet these requirements, they might actually enforce rigor more consistently than human authors do. The problem with current AI-assisted papers isn’t necessarily that machines wrote them, but that the machines were optimizing for superficial sophistication rather than substantive quality.
Third, the economic incentives might eventually align productively. AI companies have reputational stakes in ensuring their tools aren’t primarily known for flooding academia with garbage. Academic institutions and publishers have financial and prestige incentives to maintain research quality. If these parties recognize mutual interest in a functional system, they could collaborate on standards that serve everyone. We might see AI developers creating specialized research-writing models trained specifically on peer-reviewed literature, designed to meet disciplinary norms, perhaps even including built-in guardrails against common research integrity violations.
The key insight is that LLMs “continually learn, adjust, and improve.” This adaptability could be leveraged constructively. Consider how this might work in practice: journals publish updated guidelines specifying what constitutes appropriate AI use and what disclosure is required. AI developers train models to follow these guidelines, perhaps even to flag when authors are attempting to use them in ways that violate norms. Papers submitted with proper AI-assistance documentation become the norm rather than the exception. Reviewers develop expertise in evaluating not just the final product but the documented process by which it was produced. Over time, this creates an ecosystem where AI assistance is transparent, accountable, and quality-enhancing rather than quality-undermining.
We’re already seeing early moves in this direction. When bioRxiv and medRxiv implemented AI tools to provide rapid feedback on preprints, they weren’t simply trying to detect and reject AI-written papers—they were using AI to help authors improve submissions before formal peer review. This suggests a model where AI serves as a kind of first-pass quality control, catching obvious problems and helping authors strengthen their work before it reaches human reviewers. If this approach scales effectively, it could make the entire system more efficient while maintaining standards.
The structured appendix proposals that some researchers are advancing represent another piece of this puzzle. If papers include machine-readable components that specify data, methods, analyses, and results in formats that AI systems can directly verify, we create infrastructure for algorithmic quality checks that don’t depend on detecting whether a human or machine wrote the prose. The focus shifts from “was this written by AI?” to “can these claims be independently verified?” This is fundamentally healthier for science.
However—and this is crucial—reaching this productive equilibrium is far from inevitable. Several conditions would need to be met for this optimistic scenario to unfold rather than the darker alternatives.
First, the incentive structures in academia must change. As long as universities primarily evaluate researchers by publication counts, the pressure to use AI for high-volume, low-quality output will overwhelm any norms about appropriate use. The shift toward valuing research quality, impact, and reproducibility over raw productivity metrics is essential but faces enormous institutional inertia. If we arrive at a functional AI-integrated publishing system while still rewarding quantity over quality, we’ll simply have a more efficient mechanism for producing and credentialing worthless research.
Second, the commercial AI developers must genuinely commit to research integrity rather than just maximizing product adoption. There’s a real risk that competition among AI companies leads them to market research-writing tools based on their ability to help authors publish more papers faster, regardless of whether those papers advance knowledge. If AI developers instead compete on helping researchers produce better, more rigorous, more impactful work—even if that means fewer papers—the dynamic becomes quite different. This requires either genuine ethical commitment or regulation that makes it unprofitable to prioritize volume over value.
Third, we need much more sophisticated ways to evaluate both AI capabilities and research quality. The fact that papers can pass peer review isn’t necessarily evidence that they should—it might be evidence that peer review is broken. As AI systems become better at mimicking the surface features of good research, we need evaluation methods that probe deeper: Can the findings be replicated? Do they generate novel predictions? Do they lead to practical applications or theoretical insights? Do they advance their fields? These questions are harder to answer but more meaningful than whether a paper reads well or conforms to genre conventions.
Fourth, there must be sustained investment in the infrastructure that makes accountable AI-assisted research possible. Machine-readable structured data, reproducibility platforms, verification systems, specialized review tools—these all require resources to develop and maintain. If the response to AI-driven submission growth is primarily cost-cutting and automation, we’ll get the cheapest possible evaluation system, which probably won’t be an effective one. Investment is needed not just in technology but in training reviewers to work effectively with these new tools and paradigms.
The timeline matters enormously. If the productive equilibrium takes five to seven years to emerge while we spend those years drowning in low-quality AI-assisted papers that make it through peer review, the damage to scientific credibility could be severe and lasting. Public trust in scientific findings, already strained by controversies over reproducibility and fraud, might not survive a period where the research literature is widely perceived as AI-generated noise. The window for getting this right is not unlimited.
There’s also a question about whether AI-assisted research should look different from human-generated research in ways beyond just disclosure. Perhaps we need distinct venues for AI-assisted exploratory studies versus traditional confirmatory research. Perhaps certain types of claims—particularly those with immediate policy or clinical implications—should require additional verification when AI was substantially involved in their generation. The goal wouldn’t be to stigmatize AI use but to match the level of scrutiny to the nature and stakes of the work.
In short, this doesn’t have to be a story about humans versus machines or tradition versus innovation. It could be a story about co-evolution, where AI systems learn to support the actual goals of science—knowledge production, understanding, insight—rather than just mimicking its surface features, and where scientific institutions learn to leverage AI’s capabilities while maintaining the values and standards that make research trustworthy. That’s genuinely possible.
The question is whether we can get there intentionally, through thoughtful design and collective deliberation, or whether we’ll stumble toward it through years of dysfunction and crisis. The difference matters not just for scientists but for everyone who depends on scientific research—which is to say, all of us. The medical treatments we receive, the technologies we use, the policies we live under all ultimately rest on research findings. If the system that produces and validates those findings becomes unreliable, the consequences extend far beyond academic publishing.
This optimistic scenario is achievable. The arms race could become a productive feedback loop. But it requires intentionality, investment, and institutional courage that aren’t yet clearly evident. The next two years will likely tell us which trajectory we’re on.
The Cornell symposium scheduled for March 2026 (https://www.sciencedaily.com/releases/2025/12/251224032347.htm) represents the kind of collective deliberation needed, bringing together scientists and policymakers to guide these changes rather than simply reacting to them. Similar efforts across institutions, disciplines, and countries will be essential. The transformation of scientific publishing in the AI era is not a problem to be solved but a transition to be managed—and managed soon, before the gap between how research is produced and how it is evaluated grows so wide that scientific credibility itself becomes the casualty.
Sources
Berkeley Haas Newsroom: https://newsroom.haas.berkeley.edu/how-ai-is-transforming-research-more-papers-less-quality-and-a-strained-review-system
Nature: https://www.nature.com/articles/d41586-025-04066-5
Sakana AI: https://sakana.ai/ai-scientist-first-publication/
ArXiv (AI and Peer Review): https://www.arxiv.org/pdf/2509.14189
Medium (Impact Newswire): https://medium.com/@impactnews-wire/how-ai-is-moulding-the-future-of-peer-review-in-science-research-2a9f7fe5bc58
ScienceDaily (Cornell University): https://www.sciencedaily.com/releases/2025/12/251224032347.htm
Science Magazine (aiXiv): https://www.science.org/content/article/new-preprint-server-welcomes-papers-written-and-reviewed-ai
Undark: https://undark.org/2026/01/07/apc-science-publishing/
Royal Society Open Science (Stockholm Declaration): https://royalsocietypublishing.org/rsos/article/12/11/251805/234088/Reformation-of-science-publishing-the-Stockholm
IJCAI 2026: https://2026.ijcai.org/ijcai-ecai-2026-call-for-papers-ai4tech/
[End]
Filed under: Uncategorized |
Educational Technology and Change Journal 
Leave a comment