Susan King (Credit: Chris Taggart)
Artificial intelligence is changing the way that scientists write their research, and the way that journals publish those papers. But despite the changes brought by technological advances, Susan King, executive director of Rockefeller University Press, says that collaborative human relationships are at the heart of publishing more than ever before.
We spoke to King about what her team is doing now to both implement and safeguard against AI, and where she thinks we'll be in five years.
What are some of the ways you've seen AI impact scientific publishing?
One of the most established uses of AI in publishing is in protecting research integrity. That's not actually new, and didn't start with AI. Rockefeller University Press has long been a pioneer in that area. My predecessor, Mike Rossner, started screening for image manipulation more than 20 years ago, using Photoshop to adjust brightness, contrast, and levels to reveal abnormalities. RUP staff members Laura Smith and then Rochelle Ritacco joined Mike in screening for image manipulation-they continue to screen images, now empowered by AI. That's been a boon. AI makes it possible to check images in a submitted manuscript against images in preprints and in published articles across journals and publishers, to safeguard for inappropriate duplication.
We also participate in the STM Integrity Hub, a cloud-based environment created by scientific, technical, and medical publishers. It allows us to check submitted manuscripts for a range of integrity issues, including whether they've been simultaneously submitted to another journal. It's designed to be compliant with applicable laws and industry best practices, and it helps us detect things like fraudulent papers produced by paper mills that offer false promises of career advancement to desperate scientists.
Do you see AI affecting how readers find and use journal content?
We're beginning to see a change in traffic patterns as AI-generated summaries are becoming more ubiquitous and researchers are using AI search engines more frequently. The shift from classic search to AI-based systems reflects in the growth in referrals from AI search engines and a decline in traffic from traditional tools like Google Scholar.
It's not yet clear how this will play out, but it has real implications. Libraries, for example, rely heavily on usage statistics to decide whether to subscribe to a journal. Their calculations answer questions like "Is this journal heavily used at my institution?" or "Is it worth the cost per use?" If more reading happens through AI-generated layers rather than direct visits to the journal site, it could become harder for libraries to assess value using traditional metrics. Of course, when a journal's content is incorporated into an AI summary, that's a significant form of usage. So there's an open question in our field as to what kinds of metrics we can use in the future to capture a journal's true value.
That's one reason we're starting to think more about licensing agreements that assist AI agents. We recently partnered with a company called Cashmere, which surfaces journal content to third parties using AI tools and allows us a better view into AI use.
The other benefit of our AI licensing agreements is that we lower the risk of flawed AI summaries by allowing AI systems to access content directly via MCP (Model Context Protocol) instead of relying on incomplete training data. AI systems try to fill gaps when their training data is incomplete, which can lead to hallucinations. AI needs to access high-quality, peer-reviewed content, and we want to make sure that if our material is being used, it's done in a structured, responsible way.
Rockefeller University Press is relatively small. How does that shape your approach to AI?
We are very small. We publish five journals. We're not like the big commercial houses, and we're never going to dominate the landscape. At the same time, our size is a strength because we can provide a high level of author service and maintain connections to our authors and reviewers.
Because of our size, we've always recognized that we must ensure that the research we publish is highly discoverable. I often think of the research literature as a huge database, in which you have to make sure the individual units (articles) are well connected.
We therefore richly tag our content with persistent identifiers for authors, institutions, and funders. We clearly tag corrections and, when they occur, retractions, so that AI systems can parse this information and handle it appropriately. Increasingly, we're aware that our content will be "read" first by machines rather than humans, so we're asking: What do machines need? What are they looking for? Right now, we see structuring and tagging content well as becoming even more important to help AI tools do good work.
Scientific publishing depends on trust. How is AI changing trust between authors, reviewers, and readers?
For us, it's about accountability and transparency. We know people are using these tools. Some journals initially tried to say researchers couldn't use AI at all. That's not the stance we took. It's both too late and very hard to police every instance of AI use.
There are instances of AI use that we would not want to prevent. We recognize that for authors whose first language isn't English, AI can be a great help for language polishing. Even for native speakers, AI can help refine writing. We don't consider that inherently problematic.
What we do ask is that authors disclose their use of AI, so we can learn how it's being used, though we've changed our approach there a bit. Initially, we said: disclose all AI use, and then we started seeing many statements like "I used AI for language editing." We quickly realized that use made a lot of sense and is increasingly ubiquitous. So we revised our policy: if you're using AI purely for language editing, you don't need to disclose it. For more substantive uses, disclosure remains important.
For reviewers, the focus is confidentiality. Authors trust that their manuscripts, which aren't yet public, will be handled with care. A survey of 1,600 academics found that more than half have used artificial-intelligence tools while peer reviewing manuscripts, as reported in Nature at the end of last year. We ask reviewers to use systems that avoid putting confidential manuscripts into the world. And we do not allow AI systems to be listed as authors; they aren't accountable.
In your experience, is generative AI changing the nature of scientific misconduct, or amplifying existing issues?
Going back to our earlier example, it's certainly easier now to fake images than it used to be. But we haven't seen a material increase in problematic images in submissions to our journals. That may be an advantage of being small and having strong community ties-we're not being hit with paper mills at the scale some large commercial publishers report.
While it's true that AI can make it easier to commit various kinds of scientific misconduct, I generally tend to believe that many of the problems we see are shaped by long-standing reward systems in science. Problematic systems emphasize having many publications over having a smaller number of solid, careful pieces of work. That environment can push some people toward cutting corners, with or without AI, though AI does make it easier.
So yes, I think AI could be amplifying issues. It's an opportunity to revisit what we value in scientific publications. Because the underlying incentives matter a great deal.
Looking ahead five years, which parts of publishing will be more AI-driven, and which will remain deeply human?
A very obvious place that AI can play a much larger role is in manuscript submission. Right now, authors have to manually key in a lot of information, which is time-consuming and frankly a waste of their energy. AI can extract metadata like institutions, funders, authors from a manuscript and present it back to the author with a "Have we got this right?" check. That's critical for consistent tagging and attribution, which in turn can help foster discoverability down the line. The technology isn't quite perfect yet, but it's getting close. I expect that process to become much more AI-driven.
There's also a lot of interest in AI's role in supporting peer review. Some groups are experimenting with tools that can, for instance, check whether the right statistical tests have been applied. I can see real value there.
But I strongly believe we will still want humans involved in the intellectual assessment: Is this work novel? Does it push boundaries? Will it have an impact in the field? AI is inherently trained on what is already known; so much of science is about pushing into the unknown. That's where scientists are essential.
We've seen AI do amazing things in areas like protein structure prediction and coding. But when it comes to judgment about novelty, significance, and those more intangible dimensions of scholarship, that's a human quality.
