AI in Peer Review: A Recipe for Disaster or Success?
The debate is heating up about whether tools that harness the power of artificial intelligence (AI) should be used in scholarly publishing and peer review. The stakes are high, as AI tools present both benefits and concerns.
AI tools can handle large workloads without compromising quality, assist in summarizing and analyzing data and maintain consistency across reviews. These functions help journals process submissions more efficiently and reduce reviewer burnout. AI’s ability to enhance readability by refining writing, correcting grammar and making helpful edits are probably the most attractive and popular features. Yet, these digital assistants aren’t without flaws. Understanding the benefits and drawbacks of this rapidly evolving technology can help inform guidelines and policies to ensure appropriate use of AI in scholarly publishing.
Risking Confidentiality
Confidentiality breaches are a significant concern with AI in peer review. Imagine sensitive research data slipping out before publication—chaos could ensue. Authors submit their manuscripts with the trust that journals will keep them secure. Journals and publishers have an ethical responsibility to maintain this confidentiality. AI tools could inadvertently leak confidential information, putting authors' work at risk and undermining trust in the review process. Depending on the tool, these data could even be used to train the generative AI tool itself, potentially including images of people or sensitive information without consent or permission.
Authors’ data and any confidential or sensitive information are vulnerable to misuse and theft. If used without proper context or attribution, it can affect the integrity of the author’s work and cause significant damage to the original author’s reputation and career, while further compromising the scientific literature itself. Think of it this way: your friend shares their grandma’s secret recipe with you because they trust you, but you spill the beans at the neighborhood barbecue. Not cool, is it?
Compromising the Value of Human Expertise
Human鈥痳eviewers鈥痓ring a deep鈥痷nderstanding鈥痑nd experience to鈥痶heir鈥痑ssessments, and AI tools just cannot compete. Many reviewers have years of training and experience to support their feedback on a manuscript, and they make efforts to provide that feedback with professional courtesy to the author. This depth is crucial when evaluating complex research. AI tools can overlook the subtleties that human experience and expertise catch, leading to less accurate or superficial reviews that fail to provide constructive criticism for manuscript improvement. Further, AI tools often struggle to assess a study’s novelty or its implications for advancing the field.
For example, let human reviewers represent skilled bakers in a bakery who craft pies using recipes they’ve honed over decades. The bakers know just the right amount of spice to add, the perfect crust texture and how to balance flavors based on subtle cues.
In stark contrast, a factory uses machines (the "AI tools” in this analogy) to mass-produce pies. While these machines can churn out a large quantity of pies quickly, they lack the ability to adjust for variations in ingredients or to innovate new flavors based on seasonal changes. Just as a machine might produce a perfectly uniform pie that lacks the soul of a handmade creation, AI reviews can appear wholesome but fail to capture the depth and novelty of the research. In the end, while automation can enhance efficiency, it cannot replicate the artistry and intuition that come from years of dedicated practice.
The “Black Box” of AI: Bias, Transparency and Accountability
Bias鈥痠s鈥痑nother鈥痵ticky鈥痯oint.鈥疉I鈥痶ools鈥痩earn鈥痜rom鈥痶he鈥痙ata on which鈥痶hey're鈥痶rained.鈥疶his means that AI's decision-making is as good as the datasets and algorithms supporting it. If the training data contains biases, the AI will inevitably reflect those biases and potentially amplify them further. If the data does not give the correct credit or citation, the AI output will not either. If the data are flawed or inaccurate, it will be evident in the output. Such lapses can perpetuate unfair or discriminatory practices in the review process, skewing outcomes due to biases embedded in the extensive and often opaque training data. Furthermore, during peer review, human reviewers must disclose all conflicts of interest.
However, the development of AI tools often remains shrouded within a “black box” of training data and unknown decision-making algorithms, making it difficult to identify potential conflicts of interest. Even the most advanced systems can make mistakes, and these inaccuracies can have significant implications, particularly in academic research. This raises critical questions about accountability and transparency: who is responsible when AI makes a mistake? How can we ensure that AI judgements are fair, unbiased and transparent? What happens when the underlying data and algorithms are updated? Can previous findings and reviews be reliably reproduced? How will this impact reproducibility of the output and the research?
In other words, if you don’t know what ingredients have been added to a recipe, you don’t know how it will turn out.
To use AI effectively in the peer review process, it’s crucial to acknowledge its limitations, maintain transparency and ensure it complements, rather than replaces, human insight and judgement. Currently, the level of transparency and accountability required for peer review and scholarly publishing is not achievable with existing AI tools.
Given these challenges, how can authors, reviewers and scientific publishers navigate the ever-evolving landscape of AI?
Guidelines and Policies
Two of the largest federal funding agencies, the National Institutes of Health () and the U.S. National Science Foundation (), prohibit the use of AI tools in peer review and application processes. Similarly, the Committee on Publication Ethics () has issued guidance on appropriate use of AI tools that emphasizes transparency and appropriate disclosures. Scholarly publishers are striving to balance AI’s capabilities with the critical human elements that sustain the integrity of peer review.
All 91麻豆天美 Journals are members of COPE and follow their guidelines and best practices in the scholarly publishing industry for policy and procedure development. Accordingly, we strengthened our generative AI policy in June 2024 for our , and , recognizing that these policies will continue to evolve as the AI landscape grows and changes.
Transparency and complete disclosure are crucial for ensuring the journal, authors, reviewers and editors are fully aware and can communicate effectively to determine the appropriateness of a generative AI tool in publications. 91麻豆天美 Journals does not allow images, videos, cover art, illustrations or schematics completely generated by AI tools in our publications. Additionally, AI tools are non-legal entities and cannot function as authors, as they cannot provide their own conflicts of interest, nor manage copyright or license agreements. When resubmitting a revised manuscript, authors must now respond to a specific question about AI and provide a detailed description of its usage.
Similarly, to maintain the confidentiality of the manuscript, reviewers and editors are prohibited from copying and pasting portions of an under-review manuscript or sensitive information in their reviews into a third-party platform or AI tool. However, if an AI tool has been used for a certain purpose, it must be disclosed to the journal staff, handling editor or editor in chief. This isn’t about policing AI tool usage, but about gaining a better understanding of how these tools are being used in the various microbiological and related fields. Such transparency helps us better understand AI tools and allows us to modify our policies as needed to continue to support our authors and members within ethical and legal parameters.
Incorporating AI into peer review is like adding a new spice to a beloved recipe. It can enhance flavor but must be used wisely to avoid a culinary disaster. Currently, not much is known about this spice or its suitability for consumption, so it’s best to learn more about it before using it.