The scholarly publishing industry finds itself at a pivotal juncture, transitioning from an initial phase of enthusiastic artificial intelligence (AI) tool implementation to a more critical and reflective stage of evaluation. Over the past few years, publishers have made substantial investments in AI technologies, driven by an urgent need to address escalating submission volumes and increasingly complex research integrity challenges. However, as noted by Sarah Hands, Chief Operating Officer at PA EDitorial, the immediate imperative is no longer merely adoption, but a comprehensive understanding of the real-world implications, successes, and failures emerging from these early deployments. This collective introspection, through shared experiences and evidence, is crucial to move beyond isolated experiments and forge a truly informed, industry-wide approach to AI integration.
The rapid pace of AI adoption is understandable, if not entirely warranted. Publishers grapple with unprecedented pressures, including a significant surge in manuscript submissions, exacerbated by global research output growth and, anecdotally, by the ease with which AI tools can generate initial drafts. For instance, a 2026 report from Organization Science highlighted a 42% increase in submissions since the public release of advanced large language models (LLMs) like ChatGPT in late 2022, placing immense strain on editorial and peer review systems. Simultaneously, the landscape of research integrity has become more challenging, with a rise in issues ranging from sophisticated plagiarism and image manipulation to the emergence of "paper mills" producing fraudulent research. In this environment, AI promises tantalizing efficiencies: automating initial manuscript screenings, identifying potential integrity breaches, streamlining peer review matching, and assisting with metadata generation. Technology vendors, eager to capture market share, aggressively promote their AI solutions, while editorial teams often feel compelled to embrace innovation to avoid being perceived as stagnant. Yet, amid this rush, the most valuable action the industry can take now is to create dedicated space for rigorous evaluation of what is truly being learned from these early applications.
A Strategic Pause: "Stop but Don’t Stop"
This sentiment echoes the compelling message delivered by Jake Okechukwu Effoduh in his 2024 keynote address to the Association of Learned and Professional Society Publishers (ALPSP) conference. Effoduh urged the industry to leverage the advent of AI as a unique opportunity to pause and critically examine its core principles and established practices, without abandoning its fundamental commitment to scholarly publishing. His call was not to halt AI development or deployment, but to infuse it with thoughtful deliberation. The underlying question is not if AI should play a role in editorial workflows, but how it should be integrated to genuinely enhance, rather than merely accelerate, the publishing process. As publishers accumulate real-world experience, the questions evolve from basic capability to nuanced impact: Which AI tools are delivering tangible, meaningful improvements? Where are they inadvertently introducing new risks, amplifying existing biases, or creating unforeseen complications? Crucially, how can human expertise be safeguarded and empowered to continue guiding the critical decisions that underpin the integrity of the scholarly record?
The Evidence Gap: Adoption Outpacing Evaluation
A significant concern across the industry is that the pace of AI adoption often seems to outstrip the capacity to adequately evaluate its true efficacy and impact. AI tools are frequently introduced based on perceived efficiencies, competitive pressures, or a general organizational mandate to keep abreast of technological trends. However, enabling a tool is merely the first step. The more profound and challenging questions emerge thereafter: Does the tool consistently deliver the anticipated benefits? Under what specific conditions does it perform optimally? Where does it fall short, introduce new problems, or create downstream issues in the workflow?
This "evidence gap" can manifest in several ways:
- Lack of Baseline Data: Many organizations adopt AI without first establishing clear baselines for existing processes, making it difficult to objectively measure improvement or quantify new risks.
- Insufficient Post-Implementation Metrics: While initial metrics might focus on speed or volume, deeper analysis into accuracy, false positives/negatives, impact on human workload quality, or user satisfaction is often overlooked.
- Siloed Experiments: Individual publishers conduct their own trials, often repeating similar experiments without sharing findings, leading to inefficient resource allocation and delayed collective learning.
- Focus on Anecdotal Successes: The industry, like many others, can be prone to highlighting isolated success stories while underreporting challenges, inconclusive results, or outright failures.
The solution is not necessarily to decelerate adoption entirely, but to cultivate a more robust and systematic approach to learning from it. As AI is implemented, publishers must prioritize data collection, evidence-based assessment, and the widespread dissemination of lessons learned, both positive and negative.
Knowing How, Where, and Under What Conditions AI Helps Most Effectively
The true value of AI in scholarly publishing will only be realized through comprehensive understanding. Industry-wide collaboration is paramount in this endeavor. Publishers need not, and indeed should not, repeat identical experiments in isolation. Just as in scientific research, there is immense value in making null, inconclusive, and even negative results visible, rather than exclusively showcasing successes. A shared repository of evidence detailing what has worked, what has not, and where outcomes have been ambiguous, could prevent redundant efforts and construct a clearer, more nuanced picture of AI’s genuine value proposition. Such collaborative efforts could illuminate where human oversight remains absolutely essential and define what truly effective implementation looks like across diverse publishing contexts. Initiatives like the cross-industry work on AI standards and interoperability, exemplified by reports from NISO (National Information Standards Organization) in 2025, represent critical starting points for establishing common ground and shared benchmarks.
From "Can It?" to "When Should We Trust It?"
A fundamental challenge in evaluating AI within publishing workflows is the disparity between establishing a tool’s capability to perform a task and understanding the conditions under which it can be reliably trusted to do so. For instance, while an AI might demonstrate the ability to screen manuscripts for certain characteristics, the deeper questions revolve around its performance variability: Where does it excel? Where does it consistently fail? Does its accuracy fluctuate across different disciplines (e.g., highly structured scientific papers versus qualitative humanities research), article types (e.g., original research, reviews, commentaries), or writing styles? What level of human oversight is minimally required to ensure accuracy and fairness?
An AI-generated flag, whether for potential plagiarism or a methodological flaw, should never be treated as a definitive finding. Instead, it is a signal – a prompt requiring contextualization, interpretation, and often, further human investigation. When AI and human judgment diverge, clear protocols are needed to determine how such disagreements should be resolved. Answering these complex questions demands sustained evidence collection over time, a willingness to scrutinize instances where the technology underperforms, and a commitment to continuous refinement based on real-world feedback.
The Persistence and Obscurity of Bias
One of the most insidious, yet less visible, risks associated with integrating AI into editorial workflows is not merely that the AI might make an incorrect decision, but that the human tendency to question such decisions may diminish. As AI outputs – scores, flags, recommendations – become embedded alongside human judgment, there is a distinct danger that these machine-generated signals may be perceived as inherently more objective or authoritative than they truly are. Editors, who might instinctively challenge a colleague’s subjective judgment, may be less inclined to question a system-generated flag, attributing an unwarranted level of infallibility to the algorithm.
This phenomenon is known as "automation bias": the tendency for humans to over-rely on or give undue weight to recommendations from automated systems, even when contradictory information or human intuition suggests otherwise. The danger extends beyond simple error; it lies in the potential for human operators to adopt a mindset of "the system flagged it, so there must be something there," rather than engaging in critical inquiry about the meaning of the flag and its contextual relevance. Conversely, the opposite bias can also emerge: a distrust of AI-generated signals simply because they originate from a machine, or an interpretation of these signals through pre-existing assumptions about a manuscript or author. It is critical to acknowledge that human judgment itself is not immune to bias. Therefore, the introduction of AI does not merely add a new, objective layer to a neutral process; it can interact with and potentially amplify existing human biases in ways that are often difficult to detect and mitigate.
This interaction of biases is particularly critical in the domain of research integrity. An AI-generated integrity flag should always be treated as a signal requiring expert human interpretation. Human judgment, experience, and nuanced understanding remain indispensable precisely because they can discern when something that appears unusual is actually legitimate, or conversely, when something seemingly innocuous warrants a much closer look. The pertinent question, then, is not simply, "Does the AI tool work?" but rather, "What happens to the quality and critical nature of human judgment when this tool is integrated into the workflow?" The goal is not to replace human bias with machine objectivity, but to understand how the two interact and to design workflows that foster better, more informed decision-making. Workflows, even prior to AI, were not perfect. However, there is little value in technology that inadvertently diminishes our capacity to question, investigate, and think critically. As AI becomes more prevalent, the ultimate test of an effective workflow may not be how little human intervention it requires, but how effectively it empowers people to make superior decisions.
Slowing Down Is Not Resisting Innovation
The pressure on publishers to optimize resources and enhance output is undeniable. The aforementioned 2026 Organization Science report illustrates the very real strain on editors and reviewers in managing burgeoning submission volumes. AI undoubtedly offers genuine avenues for efficiency, making the industry’s keen interest in technological solutions entirely understandable.
However, a deliberate approach, often perceived as "slowing down," should not be misconstrued as resistance to innovation. On the contrary, it represents a commitment to responsible, sustainable innovation. It means affording the necessary time and resources to properly evaluate AI tools, rigorously measure their outcomes against predefined objectives, and vigilantly monitor for unintended consequences. Most importantly, it necessitates asking whether a workflow has genuinely improved – not merely whether it has become faster or cheaper. This demands a move that is deliberate, not hesitant: adopting AI with a spirit of scientific curiosity to experiment, the intellectual discipline to rigorously evaluate, and the professional willingness to transparently share all learnings.
Lessons from Cautionary Tales: The "Should" Question
The analogy, though seemingly disparate, of Ian Malcolm’s famous line from Jurassic Park – "Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should" – resonates profoundly with the current discourse surrounding AI in scholarly publishing. We have moved beyond theoretical predictions about AI’s potential; we now possess real-world experience and emerging empirical evidence. This affords us the opportunity to test our initial assumptions, learn from what truly works, and be honest about what does not. The critical challenge is to address both fundamental questions: "Can AI do this?" and, perhaps even more importantly, "Should it?"
The Opportunity Ahead: Defining "Better" in Scholarly Publishing
If the answer to both "Can it?" and "Should it?" is affirmative, then a crucial third question logically follows: "How will we know whether AI has made things genuinely better?" Without agreed-upon measures, transparent reporting, and a collective willingness to disseminate disappointing or inconclusive results alongside successes, the pace of AI adoption will inevitably continue to outrun the accumulation of reliable evidence.
The current moment presents a profound opportunity to apply the same rigorous curiosity and critical thinking that underpins scholarly research to the practical implementation of AI. As these technologies become deeply embedded in real-world workflows, we have the chance to become equally curious about their actual impact in practice: where they genuinely add value, where they fall short, and what new risks or unintended consequences they introduce. To achieve this, the industry must collectively define what needs to be measured to confirm that these tools are indeed enhancing scholarly publishing.
Mere throughput, speed, or productivity, while useful metrics, are not inherently measures of improvement. Processing more manuscripts, more quickly, does not automatically equate to a better job. "Better" might encompass a range of more profound outcomes: more accurate editorial decisions, fewer false positives or negatives in integrity checks, reduced downstream workload for human editors (by genuinely automating repetitive tasks, not just shifting burdens), greater confidence in the publishing process, more equitable outcomes for authors, and, ultimately, a strengthened integrity of the scholarly record.
The overarching goal of scholarly publishing extends far beyond mere efficiency. It is to protect and enhance the scholarly record, support rigorous research, maintain unwavering quality and integrity, and facilitate sound decisions throughout the entire publishing lifecycle. If this fundamental purpose remains at the core of our endeavors, and if we become as intensely interested in the empirical evidence of AI’s impact as we have been in its tantalizing potential, the industry can transcend the simplistic question of "what can AI do?" to address the far more crucial inquiry: "where does AI genuinely make scholarly publishing better?" This shift in focus will pave the way for a future where AI serves as a powerful, responsible, and ethical partner in the advancement of knowledge.
