AI Revolution or Revulsion? APS Journal Editors Weigh In

Arturo: AI is a new label for something that has existed for a long time. The autocompletion that occurs when we are typing is a product of AI. Autocorrect fixes our typing. Grammar check fixes our grammar. What the latest iteration of AI does is to take it one step further by generating text itself. That makes writing more appealing, but it also gives writing a very specific style. 

This new technology comes with benefits and costs. The benefits are particularly helpful to nonnative speakers who may have been held back by the cost of producing text in a less proficient language. The cost is that AI produces generic text, and often that text may sound good but not be meaningful or novel. Only humans can decide whether a contribution is genuinely novel or a very sophisticated hallucination. 

A manuscript written entirely in AI style with nothing original behind it shouldn’t be accepted. But a manuscript in AI style with truly original ideas—especially from a nonnative speaker—should be. Sounding good and being good are very different things, and only a human can tell them apart. 

One last note. There’s also an equity argument. Researchers in elite spaces hire editors and have talented students and postdocs who do a lot of writing for them. Not all researchers and professors have access to these resources. The editors, students, and postdocs who may be assisting in numerous ways are often coauthors or noted in the acknowledgements. So, perhaps acknowledging AI is an interesting parallel to the human acknowledgment that takes place in published research.  

Rachael: At a basic level, authors are already responsible for the content of their manuscripts, regardless of the tools used to produce them. In that sense, I am not convinced that AI requires a fundamentally new disclosure framework. We do not ask authors to disclose their use of other tools, and AI, in many respects, sits within that broader continuum. 

The central issue is accountability, and that has not changed. Authors are expected to stand behind the accuracy, originality, and integrity of their work. AI introduces new ways in which that accountability can be compromised—for example, through uncritical use, lack of understanding, or pressures that encourage overreliance. These are extensions of existing challenges rather than fundamentally new ones. 

A more productive focus is on responsible use. This brings us to AI education. If authors are using these tools without understanding their limitations, that is where the real risks emerge—errors, hallucinated content, or misplaced confidence in outputs. While journals are not primarily responsible for training, they can play an important role in signaling expectations and providing practical guidance on where AI adds value, where verification is essential, and where risks are highest. 

The goal is not to single out AI as exceptional, but to ensure that evolving tools are integrated in ways that preserve a core principle of scientific publishing: Authors are accountable for what they publish. 

Nicholas: The good news is that many journals are developing AI disclosure policies; the bad news is that most of these disclosure policies are largely incoherent in numerous ways. For instance, many journal policies require admission (or even confession) of AI involvement—these policies may be intended to “protect” the integrity of the literature, but they do not appear aimed at advancing scientific discovery. Asking something akin to, “Did you use AI?” is about as informative as asking, “Did you use software?” It collapses trivial and consequential AI application into the same category, and it produces little but noisy and uninformative disclosure at best. At worst, and I imagine more frequently, it drives AI-supplemented science “underground,” defeating the goals of enhancing science and improving transparency. That is, most policies thwart the very goals they attempt to achieve. 

The best framing of AI is not to what extent it was used as a tool but rather to what extent it contributed positively to the final scientific product. Transparency should be required when AI generates or materially transforms reasoning, analysis, interpretation, or representation—when it meaningfully contributes to the epistemic content of the work. Thus, we need to strive primarily not for restrictions on AI use but instead for auditability of AI use. Reviewers and readers always need to be able to inspect what entered the scientific record; it is critical they also can identify the origins of those contributions. 

In a recent policy I have developed for Clinical Psychological Science, requirements shift from a prompt model to an artifact-centered model: Researchers would archive the AI outputs that shaped their work, and they would only include prompts when they add interpretive context. That approach fosters transparency and seems scientifically defensible. Most journal policies are not. 

June: Transparency is a key ingredient to advancing rigorous psychological science. Disclosure of generative AI use is no exception. Authors should disclose all uses of AI whether great or small, including drafting text, generating summaries or figures, assisting with idea development, and proofreading. Clear disclosure benefits authors, readers, and the field itself by helping us understand how AI is shaping the research process and how it is reported. 

Felix: This is a difficult question to answer. I suppose using AI as a reviewer to summarize a paper would be OK, but at the same time, why would the reviewer need a summary when we expect them to read the manuscript in its entirety? I would not want reviewers to use AI to evaluate the paper (and essentially write the review)—in that case we could just do away with human reviewers and rely on silicon reviews. The only permissible use of AI for reviewers that I can see is by using AI to spellcheck and grammar-correct a review, written by the human reviewer. Maybe we will need to require a disclosure statement from our reviewers as well. 

June: Peer review is intended to obtain an individual expert’s feedback on a colleagues’ written work. Reviewers should strive to maintain this ideal and use their own judgment in the process. If AI tools are used in peer reviewing, disclosure should be required as to how it was used, consistent with expectations for the authors themselves. 

Jamie: First and foremost, regardless of how AI is/isn’t used in peer review, it’s critical that peer reviewers recognize that the final decision (and therefore, also responsibility) rests with them. Even if AI tools are used, they should only ever be used to support human decision making. 

With that said, personally I think there are interesting possibilities for using AI in peer review to help support peer reviewers. There are a few tasks in particular that are important in peer review, but we know are infrequently done: checking preregistrations against papers for consistency, checking code against manuscripts for both computational reproducibility and descriptive accuracy of reported analyses, etc. At Psychological Science, we do this checking at a much higher frequency because of the Statistics, Transparency, and Rigor Editor system, which has proven to be beneficial to improving the overall quality of publications. But such checking also takes a lot of time and effort.  

For this reason, as part of my own research I develop tools (like RegCheck) to help automate some of these processes. My personal opinion is that tools like RegCheck can be really useful in principle in the peer-review process, and indeed, RegCheck is now being used at Advances in Methods and Practices in Psychological Science to check submissions. At Psychological Science we are not yet using such tools as part of our editorial workflows, but it is an ongoing conversation.  

Arturo: Peer review is exactly where novelty detection matters. AI can polish empty ideas, and it can wrap original ones in generic style. Detectors flag the surface; only humans judge the thought. 

Reviewers should be permitted to use LLMs—many already do—to summarize, to scan for generic AI text, and to help structure an evaluation. But that work has to be seconded by the editor. The risk is reviewers offloading judgment to a tool that operates on form rather than thought. That’s why editors and reviewers now need to understand AI well: its tics, how it behaves across languages, what it gets right and wrong. That literacy is part of the job. 

I personally am OK with editors using AI to help assist with the process. However, the question of disclosure is an important one, and it might be wise to develop a policy or even a framework to ensure the appropriate use of AI. This might mean that publishers should develop their own in-house tools to protect intellectual property, maintain confidentiality, and ensure fair use of AI in manuscript evaluation. 

Rachael: My use of AI in editorial workflows has been relatively focused, primarily on improving communication with authors and others involved in the editorial process. The publication process is often high-stakes and, at times, emotionally charged, and the limited channel of email can amplify the potential for misunderstanding. Small differences in phrasing can meaningfully shape how decisions are received. 

In that context, AI can be useful for sharpening clarity and framing—helping ensure that feedback is precise, appropriately conveyed, and less likely to be misinterpreted. This is particularly important in communication with authors, where negative outcomes, even when justified, can feel unfair or discouraging if they are not handled carefully. These interactions sit within a broader, and often delicate, relationship of trust between authors and the journal. How decisions are communicated can shape not only perceptions of fairness in individual cases, but also confidence in the review process more generally, with implications for the journal’s reputation. 

In that sense, communication is not peripheral—it is central to how the fairness and integrity of the process are experienced. 

An additional benefit is efficiency. Clearer communication can reduce unnecessary back-and-forth and support faster, more effective responses, which benefits everyone involved. 

As with authors and reviewers, responsibility ultimately sits with the individual. AI can help shape communication, but the substance, tone, and intent remain entirely the responsibility of the editor. Used in this way, it can reduce friction without altering the basis on which decisions are made. 

Arturo: I do not use AI at all in my editorial workflow. I served as Editor-in-Chief at the Journal of Neurolinguistics for five years before joining Perspectives on Psychological Science. I have written two books and published numerous articles. At this point in my career, I have developed a good sense of what I do and do not know and good instincts about manuscript handling. However, a lot of this work was done when the volume was lower than it is now. With increasing volume, it may be difficult to manage without additional assistance of some sort, AI or human intelligence (HI). 

Nicholas: Secure, well trained AI models are profoundly useful when scale overwhelms resources. They can help compress large volumes of text, identify the core claims of a manuscript, and surface internal inconsistencies that might otherwise be missed. They can help generate counterarguments—to stress-test whether a paper’s claims actually hold up under alternative framings. 

One of the most helpful ways I have used AI systems during my term as Editor-in-Chief at Clinical Psychological Science was when I was forming my group of associate editors. I scraped the public information about all of the manuscripts published in the journal in the last 5 years off the web, and I used AI to help me identify content- and methodology-based themes that defined the majority of the papers the journal had published—the areas I knew I would need to recruit associate editors with at least some expertise. I cannot overstate how helpful this was when forming the Editorial Board for a broad coverage journal like Clinical Psychological Science. 

Felix: Students will necessarily be exposed to AI. As an instructor, it would be naive for me to believe that students are not using it, and, to an ever-greater degree, are relying on it. I teach quantitative methods, and a big part of my teaching and the associated assignments was to learn how to code in R to run statistical analyses. I think it is fair to say that learning to code (for many students for the first time) from scratch was a challenge. This challenge has gone away now, and I often find my office hours (which used to be very busy) quite lonely. AI is just extremely good at coding in R, and it allows students to make quick progress. Taking away the struggle does rob students of learning opportunities though. I recall getting my student evaluations last year, and one student stated that they felt that AI is a trap, and that their coding assignments took on the form of them pasting the question into AI, getting back R code that they paste into R to run it, and then going back and pasting error messages back into the AI—back and forth. In this scenario, actual learning and academic growth are replaced by rote copy and pasting, and that is clearly not the direction we want to go in with regards to student learning. AI is here to stay, and we need to embrace the fact that students will not only use it but will be expected to know how to use it responsibly. I still have not exactly figured out how to change my pedagogy to accommodate this new world. 

Ulf-Dietrich: Good answers to these ethical and educational questions seem to vary with topic, task, persons involved, AI and context, and change dynamically and rapidly, as for any new technological development. We hope to publish an evidence-based journal issue about these AI-related questions—and more about psychology and AI—soon. Importantly, we’ll make sure it has historical depth, as AI research has been around for a long time. 

June: The role of AI in higher education is complex, dynamic, and evolving. A key concern is that overreliance on AI may rob students’ engagement in critical thinking and diminish the sense of reward that comes from cognitively effortful tasks. We must proceed with caution and protect student learning as a priority. 

Nicholas: Once a student understands the fundamentals of a method or domain, they should be using AI extensively—and deliberately—to become more efficient and more expansive scientists. The goal is not minimal use; it is strategic use under supervision. AI can accelerate exploration, generate alternative approaches, and reduce the time spent on low-value mechanical tasks. Clearly, students need to internalize information themselves at a foundational level, and they need to develop their own skills in rhetoric and critique. But, even in those endeavors, properly used AI can be of paramount benefit. 

The real issue is not reliance, but where reliance occurs. Students should not outsource derivation, justification, or interpretation. They need to understand why a model is appropriate, what assumptions it makes, and how to interpret its outputs. But once those foundations are in place, there is no reason to artificially restrict AI use. In fact, doing so would put them at a disadvantage. The future will not be “humans versus AI,” but it will be “humans who partner with AI versus humans who don’t.” I have little doubt the former group will be better prepared for the next era of science. 

Jamie: The best way for students to use AI in their work is to first critically analyze the output of LLMs. For example, one exercise I like to do with students is to provide them with output from an LLM asked to do a specific task (e.g., write an essay in response to a title I have set) and then have them examine the outputted essay as if they were marking it themselves. They quickly see all of the ways that LLM outputs can fail, especially in terms of saying things that sound smart but actually are contentless. I think taking this type of didactic approach is better than outright banning its use or saying nothing at all about it, because it is simply a fact that most students are using it now anyway.  

In terms of reliance on AI affecting learning processes, I think there is not currently enough evidence to speak on this. Much of the current literature on cognitive offloading in the context of LLMs seems to me to be suboptimal in its implementation, and I think that we still need good investigations into this question to establish causal evidence one way or the other.  

Jamie: First, we need to be specific about what we mean by AI here. If we mean generative LLMs, then all the standard risks apply. Namely, the stochastic nature of LLMs means that the same text may be translated in slightly different ways across runs; important terminology might be butchered, mistranslated, or translated when it need not be. For example, if discussing the “Garden of Forking Paths” in a German text, this term should be preserved as-is rather than translated to “Garten der sich verzweigenden Wege.” In this sense, I would push back on the “low risk” aspect here: All of the same risks of generative LLMs in other contexts apply here, too. With that said, generative LLMs are also exceptionally good at representing the contextual meaning of language and may be particularly helpful for nonnative speakers to translate their work into English (or whatever the required language is). The ideal safeguard here is that these translations be checked by an expert who can speak the language that the text was translated into and confirm that there are no errors, mistakes, or mistranslations.   

Ulf-Dietrich: A human scientist native in the target language should read over it. And well-established methods like independent back-translation and comparison of multiple translations will be helpful in achieving high-quality output. 

Rachael: I agree that AI can play an important role in reducing language barriers. In a field that is largely English-language focused but global in scope, this is a valuable way to broaden participation and diversify psychological science beyond its historical reliance on WEIRD populations and concepts. 

That said, it is not without risk. Translation does not simply convert language—it can also shift meaning. This is particularly relevant for theoretically dense or culturally embedded constructs, where small differences in wording can affect interpretation. There is also a risk of false fluency, where text reads clearly but subtly misrepresents the intended meaning. These are general considerations, and I would be interested to hear more from expert translators on where the main risks lie in practice. 

As in the previous cases, the underlying principle is accountability. Authors remain responsible for the accuracy and meaning of their work, regardless of how it is produced. In the context of translation, this means ensuring that meaning is preserved and that outputs are carefully checked, rather than relied upon uncritically. AI can support this—for example, by enabling comparison across translations or highlighting ambiguities—but it does not replace the need for careful author oversight. 

Effective use of these tools depends on understanding their limitations—where they are reliable, where they are not, and where additional verification is needed. Safeguards, in this sense, are less about restriction and more about informed use. 

Arturo: I’m optimistic. AI will force us to stop publishing so much and start asking what is worth publishing. In a world where people can submit several AI-generated manuscripts a day, more volume hits diminishing returns. My hope is that people will turn to producing less not more. I already believed there was too much text before AI; now I’m certain.  

Nicholas: I am more than optimistic, but not because AI will make science easier. In many ways, it will make science more difficult, in fact. But, AI, if used properly, will make science more explicit and expansive. AI forces us to confront ambiguities that have always existed but were easier to ignore—what counts as a method, what counts as a result, what level of transparency is actually required for evaluation. 

What excites me is that AI pushes us toward a more rigorous conception of scientific accountability. When a system can generate text, code, or analytic strategies, the relevant question becomes not who produced something, but how we know the product itself is valid. That is a healthy foundation for science. 

What concerns me are current journal policies, which are often performative rather than principled. Journals are implementing rules that do not map onto how these systems actually function—overemphasizing prompts, relying on invalid detection tools, and imposing restrictions that are neither enforceable nor scientifically meaningful. The journals that begin to embrace AI, and require transparency and rigor in doing so, almost certainly will publish the best work in the years ahead. 

Rachael: I am both optimistic and excited, while also anticipating inevitable challenges. What I find most exciting is the potential for AI to increase accessibility to science. This includes reducing language barriers, broadening participation, and transforming how research is communicated—for example, by enabling papers to be translated into formats such as podcasts or other more accessible outputs for both students, teachers, and the public. There is also potential to improve efficiency and, importantly, the clarity of communication across the publication process. 

At the same time, there are important risks. One concern is the emergence of unintended consequences that may create strain on the system. By lowering the barriers to producing manuscripts, AI may substantially increase submission volumes. While this is positive in terms of access, it places additional pressure on already limited editorial and reviewer capacity, which may in turn lead to restrictive policies that counteract those benefits. 

I am also concerned about a gap between use and understanding. As AI becomes more embedded in everyday practice, there is a risk that it is used uncritically—whether through lack of awareness or increasing pressures within the academic landscape. This raises the risk of an abdication of responsibility. Over time, this can lead to increased errors and ultimately an erosion of trust within the research community, among the public, and in science more broadly. 

Ultimately, the central challenge is not whether AI should be used, but how it is integrated. As in the previous cases, this comes back to accountability and education: ensuring that individuals remain responsible for their work, and that they have the understanding needed to use these tools effectively. 

Handled well, this is not a threat to the system—it is an opportunity to strengthen it. This also calls for a collective effort to support the transition. That includes investing in education and developing shared practices that enable the field to benefit from these tools while maintaining core standards. 

Ulf-Dietrich: AI is a method that can be applied to many tasks. I am both optimistic and pessimistic. Elsewhere (Reips, 2023), I predicted that right now it is exciting and useful. Its usefulness and quality likely will go down from here. And it has a lot of biases in the way it is constructed and works. However, I find it exciting to apply it to what we do in research and scholarly publishing and explore its potential. Of concern to me is that publishing and AI development are on incompatible timelines. Attempts to describe “current AI” are per se outdated once we get them published in our time-tested scholarly peer-reviewed fashion. We simply will have to try coping.