Traffic Stop

Traffic Stop

1

With all apologies to The Brief reader who wrote in last month to thank us for “an AI-free issue,” we can’t help but devote much of May to the enormous AI-driven changes on the increasingly close horizon. For the short term, much of the focus for scholarly publishers has been around rights and licensing of materials for AI training sets. Taylor & Francis’ parent company Informa announced a $10 million AI deal with Microsoft, providing not only for the use of AI within the company, but also for the use of specialist data to train the Microsoft AI. We’ve heard from our society clients that other publishers are socializing the idea of content licensing for AI, and last month Times Higher Education offered a call from Oxford University Press’s David Clark to “engage in good faith.” From a legal perspective, offering licensing opportunities makes sense, at least according to Copyright Clearance Center’s Roy Kaufman, who notes that the availability of a license makes it much easier to enforce infringement claims. It also makes business sense to look at revenue from licensing for AI training to make up for declining revenue from other streams, at least in the short term. However, the complex nature of these licensing deals for enormous swathes of content will, like everything else in our current market, further encourage consolidation. A major publisher with thousands of journals is more likely to get the attention of a technology behemoth than a small, self-publishing society. 
 
In the long term, two major issues suggest this may not be as remunerative a strategy as is hoped. First, publishers, often at the behest of funder mandates, have been issuing more and more CC BY licenses for scholarly materials, which would allow the enormously wealthy technology corporations free rein to absorb the literature into their AI products without any compensation to those who published it. Although (in our experience) support for CC BY borders on religious dogma, perhaps rethinking why it is essential to achieving open access (OA) goals is worth considering. 
 
But our larger concern is the lack of imagination among publishers for the potentially game-changing nature of AI. While some continue to doubt the rapid advance of AI (and we must admit we harbor some skepticism as well), it is worth considering potential scenarios. Google, this month, announced its new “Search Generative Experience” (SGE), in which the search engine will begin directly answering queries with “multi-paragraph replies that push links to other websites further down the page, where they’re less likely to be seen.” The placement of links on the page is not the primary problem here – it is that Google will be increasingly answering search queries directly, obviating the need to click on any links and visit the source web pages from which Google draws those answers. As Casey Newton observes, “Google extended itself into so many different parts of the web that it became synonymous with it. And now that [large language models] promise to let users understand all that the web contains in real time, Google at last has what it needs to finish the job: replacing the web, in so many of the ways that matter, with itself.” Gartner predicts that traffic to websites from search engines will fall by 25% by 2026, and others predict even larger drops.
 
Losing 25% of your traffic seems like a big problem, particularly if that usage is factoring into purchase decisions, but it may be just the beginning. Platform providers are perhaps out in front of the rest of the scholarly community in thinking about such things, with, for example, Silverchair offering a session titled “What’s your strategy if there is no platform?” at its next Platform Strategies meeting. What does it mean for a journal if machines are its primary readers, and the research community instead gets its information from AI summaries of the literature?
 
With a few exceptions, such as clinical medicine or drug discovery, much of academia is at best a niche audience, and for most AI companies, the literature may not be particularly valuable because it is not pertinent to the sorts of questions their customers are asking (e.g., which Olivia Rodrigo song rocks the hardest? The obvious answer is “Bad Idea Right?” but that is not something AI would determine based on scanning the contents of Cell). 
 
A few alternative scenarios could play out here. We could see the rise of “academic AIs” trained on the scholarly literature. This would be like the evolution of Google Scholar or perhaps a database such as Ovid. We could also see more of a “pull” model along the lines of RAG (Retrieval Augmented Generation), where a generalist AI ingests content on-the-fly based on the question being asked and the access constraints of the person asking (i.e., which content does their university subscribe to?). In other words, the institutional site license (and not the AI copyright license) could continue to be the primary revenue stream for most publishers, even if most reading is not occurring on their platforms.
 
For many, the perceived gold rush toward AI licensing may turn out, in retrospect, to feel a lot like the days when we were all promised that ads would pay for everything on the internet.

Negativity

2

Most journals welcome negative results, and many explicitly state that this is the case, provided that the studies live up to the journal’s standards for experimental rigor and significance. This does not seem to have stopped the near constant complaints about publication bias that have been with us since the 1970s. Two articles this month, however, offer a new angle on negative results – their value in balancing out the bias in AI training sets based on the research literature.

Steve Cranford, Editor-in-Chief of Matter, offers a tour-de-force look at both the value of negative results and the reasons why they don’t get published. Obvious reasons include: no one wants to look like a failure; once something doesn’t work few want to invest the time and effort required to convincingly prove it and write it up; and, in the end, “we can never quite know why a study generated negative results.” Aside from these, Cranford makes an important point that we at The Brief had not seen previously expressed: negative results are part of any positive experiment (and thus are incorporated into the paper), because they are a “necessary byproduct” of the experimental process:

Often, the results aren’t useless – they are used to evolve a hypothesis, tweaking aspects until the observation supports the prediction. The null and negative results, as they guide and evolve the hypothesis in a self-correcting manner, are commonly a key aspect of the story.

Cranford also explores the issues that could arise should publication of negative results become more widespread, particularly salami slicing (given there are an infinite number of ways to be wrong, how easy would it be to fill a CV with variations of the same failure?) and the resulting metric manipulation that researchers might use to avoid being “ratioed,” a careful balance of publishing some, but not too many, negative papers.

But Cranford, like Rachel Brazil writing in Nature, recognizes the value of negative results for machine learning. The more robust a training set, the better the outputs, and therefore having only the positive results from experiments that worked may bias scientific AI tools, making them a poor substitute for the actual experimental process where failure is more common than success. The question, though, is whether the published scientific paper is the right place for making those results available. Does it make sense for a researcher to invest the same level of time and effort needed to publish a new discovery into something that didn’t quite work out? Should we be thinking more about open data and repositories capturing all outputs (positive and negative) from a project rather than the research literature?

Transfer

3

Elsevier and the American College of Cardiology (ACC) have announced a manuscript transfer collaboration between The Lancet and JACC. While the announcement does not say so explicitly, we assume that relevant cardiovascular papers will be transferred from The Lancet to JACC. As Elsevier does not publish a Lancet title in cardiology, this fills a hole in the Lancet portfolio for Elsevier. At the same time, it provides JACC a way to better compete with JAMA Cardiology and provides an advantage as compared with specialty competitors Circulation and European Heart Journal, neither of which has a transfer agreement with a top multidisciplinary journal. 
 
The agreement between The Lancet and JACC is noteworthy. There are only two multidisciplinary medical journals in the world that are comparable to The Lancet – New England Journal of Medicine and JAMA – and there are few specialty titles with the caliber of JACC. There are no other examples we can think of where two such journals are published by the same publisher, which is not a strict limitation but it makes both financial incentives and editorial workflows easier to manage. Perhaps more noteworthy than the specifics of this agreement, is the doorway it opens for other titles. 
 
Societies have been asking publishers for help in growing their submissions. Including society journals in transfer strategies would help both society partners and publishers by keeping papers in the publisher’s portfolio. The details matter greatly, however, placing natural limits on what publishers can offer. For example, the journal receiving the transfers needs to have an overlap in scope with the transferring journal. The journals must have comparable formats for papers so that they do not have to be rewritten. And the Journal Impact Factors (JIFs) must be relatively close, or comparable with alternatives – JACC’s JIF, for example, is some distance from The Lancet but close to Circulation,European Heart Journal, and JAMA Cardiology – such that authors find the transfer palatable. In other words, transfer options need to be thoughtfully constructed, and therefore limited. 
 
All that said, a publisher able to offer a society a compelling transfer strategy will have an advantage both in wooing a society to their portfolio and in retaining that society in future contract cycles. It would be difficult to leave a publisher if the society is also leaving behind a successful transfer source.
 
Publishers also have other calculations to consider. If a publisher transfers a paper to a society’s Gold OA journal, the publisher will have to split the article processing charge (APC) with the society. If the publisher transfers a paper to their own, proprietary Gold OA journal, they will keep 100% of the APC. Moreover, by helping to build a society Gold OA title, publishers are investing in titles that may, despite transfer programs, depart for greener pastures. However, it may make sense for publishers to take the “APC dilution” in certain circumstances – such as if the publisher has a gap in their portfolio, a society title is a key account, or the publisher makes revenue on the society portfolio in other ways.   
 
Of course, the transfer may work in the other direction, with society journals transferring to publisher-owned titles. Journals with high rejection rates may find themselves in a position to create some revenue related to the papers they reject as we are hearing more discussion about “finder’s fees” (a portion of the eventual APC) paid to journals for articles transferred. That said, as with publishers, societies will need to weigh bolstering the growth of other journals against growing those in their own portfolios.

Briefly Noted

4

This AI-heavy issue of The Brief would feel incomplete without a roundup of AI-based products hitting the market this month, which included Elsevier’s SciBite Chat (“a transformative AI-powered semantic search tool for life sciences R&D”), Clarivate’s Web of Science Research Intelligence platform (“an AI-native platform that embodies a vision centered on three pillars: unification, innovation and impact”), and Silverchair’s AI Playground (“a space for Silverchair clients to easily experiment with a variety of AI models, applications, & use cases”).
 
Meanwhile, OUP has surveyed more than 2,000 researchers to find out how they are reacting to and using AI in their work. Unsurprisingly, like most things in academia, researchers seem to trust their own responsible use of AI but are concerned that others might not be as skilled or ethical. More than three-quarters of those surveyed have used AI in their work, and two-thirds feel it has benefitted them. But there are also concerns about AI reducing both critical thinking skills and the quality of research results. And just like publishers (and Scarlett Johansson), there seems little faith among researchers that AI companies will respect their intellectual property.
 
Does the release of a new computational tool that doesn’t offer access to its source code or the ability for the community to evaluate its results warrant publication in a journal? Controversy erupted this month over Nature’s publication of Google DeepMind’s protein-folding prediction program AlphaFold 3, particularly because, unlike the release of the previous version, no source code was made available to reviewers or to the community of users. Furthermore, access to the tool is limited to 10 predictions per day, preventing evaluation of its results at scale. In this case – a high-profile, likely to be highly cited paper – Nature chose to exclude the work from its requirements for making code available to reviewers and alongside publication. A letter of protest from members of the research community has been submitted to the journal.
 
The fallout from Wiley’s purchase of Hindawi just keeps coming, with 19 more journals shuttered in the clean-up effort following the widespread discovery of papermill fraud.
 
Springer Nature released their 2023 benchmarking report Editor Diversity at Springer Nature Journals. The report is filled with fascinating data, but we at The Brief can’t help being surprised at the ongoing lack of gender balance among editors. While some of this may be due to the makeup of specific fields, overall women make up only 29% of editors and 22% of editors-in-chief. Kudos to Springer Nature for their transparency here, and their ongoing efforts to better balance their editorial boards, although the example given for Scientific Reports, where women are only 75% as likely as men to accept an invitation to join the board shows the significant efforts that are still needed. C4DISC (Coalition for Diversity & Inclusion in Scholarly Communications) also provides data with the results of the 2023 Workplace Equity in Scholarly Communications survey.
 
In the era of preprints, the speed advantages offered by the F1000Research model seem no longer as compelling. In what appears to be an attempt to improve the journal’s author experience, F1000 will no longer require that authors identify their own peer reviewers, and will instead rely on editorial staff to lead the process. It should not surprise anyone that F1000 has found that a professionally led editorial process is faster than relying on authors to take care of things for themselves. This seems like a smart move, given that, other than authors particularly smitten with F1000’s open peer review model, they are competing against similarly priced megajournals that spare their authors this extra work burden.
 
The Library Journal’s Periodicals Price Survey projects a 5.5%–6% price increase for journals in 2025, “well above the overall inflation rate but in line with what the market has historically tolerated in non-recession years.”
 
In perhaps a harbinger of researcher reaction when the policies resulting from the Nelson Memo are fully implemented, Brian McGill suggests that, from the researcher point of view, “open access has been a disaster,” and that unfunded researchers, and those outside of Plan S funding, will struggle to find affordable outlets for their work.
 
Bloomberg reports that the research society Optica failed to disclose that Huawei, a company “blacklisted by the US,” was the sole financial backer of the society’s “Optica Foundation Challenge” competition. Through this competition, Optica has distributed millions of dollars to researchers at American universities that have banned their researchers from accepting such funds directly from Huawei.
 
Harvard University’s library announced the launch of the Harvard Open Journals Program, meant to provide guidance, support, and seed money to Harvard researchers looking to launch new Diamond OA journals or to flip existing journals to the model. While experimentation in the market should always be applauded, we note that, as with similar such programs, no mention is made of the scale necessary for such journals to become a significant part of the publishing landscape, or of long-term support for the journals once the seed money is gone.
 
A Letter published in the journal Quantitative Science Studies asking “Are open access fees a good use of taxpayers’ money?” seems to contain more caveats and disclaimers about the poor quality of data presented than actual data or meaningful conclusions. The author, Graham Kendall, suggests that taxpayers paying for the publication of research is an “unintended consequence” of funder policies, rather than a reasonable assumption that making results public is an essential and costly part of the research process. There is a failure to recognize that not all the papers included in the study are funded, let alone taxpayer funded. Even with all these problems, the estimated $600 million per year spent on APCs is 0.024% of global research funding, which, to answer the question asked in the article’s title, seems a very effective use of taxpayer money.
 
In another heavily caveated study, “On the peer review reports: does size matter?,” authors Maddi and Miotti conclude that “beginning from 947 words, the length of reviewer reports is significantly associated with an increase in citations.” While they suggest this is because longer reports contain requested improvements that enhance the quality of the submitted manuscripts, a few confounders are difficult to parse out. The sample of peer review reports is taken from postings on Publons, and it is impossible to know how well the reports that users choose to post represent the overall population of peer review responses. Furthermore, the reviews are only studied for accepted articles, and there are no indications whether each sample is from the first or subsequent rounds of review after revision. A more effective study would look within a set of journals and compare the length of reviews of papers published within a specific journal to their citation performance, rather than the mixed bag here that may be heavily influenced by editorial policies at higher citation journals. These issues aside, the study does suggest value in a well-run and thorough peer review process.
 
That process is, however, not without costs, the bulk of which come from the time put in by staff. In the case of the Nature journals, those staff are looking for better compensation for their efforts, complicating Springer Nature’s long and winding road to an IPO.
 
Aside from long peer reviews, another route to improving citation performance appears to be posting a preprint or sharing data, but apparently sharing code doesn’t make a difference. In the preprint “An analysis of the effects of sharing research data, code, and preprints on citations,” Colavizza et al., find a correlation between preprinting and data sharing and higher citations, but no measurable impact from code sharing. The authors are careful to point out, however, that correlation does not equal causation, as a reasonable read of the paper suggests that either authors may selectively engage in open science practices for the papers in which they have more confidence, or simply that better scientists doing better science practice better transparency. As has long been the case for the alleged OA citation advantage, once again it seems there are no easy ways to game the citation system.
 
So, if longer peer reviews, preprints, and open data (but not code) correlate (but don’t cause) higher citations, what about open peer review? “Open peer review correlates with altmetrics but not with citations: Evidence from Nature Communications and PLoS One,” published in the Journal of Informetrics, suggests that there is no correlation between making peer review reports available and higher citation levels. In fact, for Nature Communications, articles that used traditional, non-open peer review averaged more citations than those that were open. In both Nature Communications and PLOS ONE, openly peer reviewed papers saw higher Altmetric scores, which the authors suggest mean they are more trusted by the press and the public, but we suspect may be more a reflection of the online behavior of authors, with those favoring open peer review potentially being more active on social media.
 
A WIPO study concludes that programs such as Research4Life are tremendously successful in driving research output in low- and middle-income countries.
 
Is the best business model for OA books, gasp, subscriptions?
 
Melinda French Gates has resigned as co-Chair of the Bill & Melinda Gates Foundation. Her exit, we were glad to learn, was not due to our unfavorable review of the Foundation’s new OA policy last month. 
 
Penny Ladkin-Brand was named the Chief Executive of Taylor & Francis. With a background in consumer-facing digital media (she was formerly Chief Financial and Strategy Officer at Future plc), Ladkin-Brand has the potential to bring new ideas to our industry, although faces a steep learning curve in adapting to the very specialized scholarly publishing market.
 
Wolters Kluwer posted strong first-quarter financial results, with 6% organic growth in their Health division.
 
Continuing its efforts to improve research integrity, the Chinese government passed a new law aimed at preventing misconduct. The law, which goes into effect January 1, 2025, standardizes degree requirements and specifies the responsibilities of institutions, supervisors, and students. Degrees can be revoked, or not conferred, when plagiarism, identity theft, ghostwriting, or fraudulently obtained qualifications are discovered.
 
Now in its 11th year, the Declaration on Research Assessment (DORA) has not yet achieved its goal of eliminating the use of journal-level metrics in research assessment. Although perhaps a sign of capitulation to the research community’s unwillingness to give up quantitative metrics, DORA has released a tremendously useful explainer of the shortcomings of common metrics and how those issues should influence what one takes from any given score.
 
Learned Publishing is going fully Gold OA in 2025. While this is a boon to readers and to ALPSP (Association of Learned and Professional Society Publishers) member authors, whose APCs will be covered by the organization, one has to wonder what it means for non-member authors, as most of the journal’s publications are not funded. Dimensions (an inter-linked research information system provided by Digital Science https://www.dimensions.ai) shows that only 32 of the journal’s 75 articles (43%) in 2023 list funding, which leaves 57% of paper authors on the hook. Furthermore, the move lessens the value of membership to the organization, as there are more readers than authors among its membership who can now choose to access the journal without joining (not to mention the negative impact on Society for Scholarly Publishing members for whom access was also a member benefit). This aligns with the struggles many research societies are facing in a transition to OA – they will have to make up for the loss of journal access as a valued benefit provided by membership in the society.
 
An investigation into suspicious phrases in peer review reports found that they were often repeated, suggesting that plagiarism in peer review is now a thing. It is hard to understand why this is the case – these are not papermill outputs receiving fake reviews to deceive editors, rather, they are legitimate papers seeing peer review reports that were “vague and lacked substance.” Nature suggests it may be due to reviewers trying to save time or not being confident in their writing ability, but if either is accurate, then why bother volunteering for peer review at all? There’s no direct financial or career reward offered for peer reviewing a paper, so the motivation for faking it seems like it should not exist. But with expectations for “the problem to get worse in the coming years,” it looks like running plagiarism checks on peer review reports may get added to the long list of time consuming and expensive things editorial offices must do to counter unethical behavior from researchers.
 
Bleak news for both book publishers and authors – Galley Beggar Press offers a breakdown of how much book production costs have increased in recent years and how much profit margins have plummeted. And Elle Griffin looks at the lessons learned about the publishing industry from the legal case when Penguin Random House tried to buy Simon & Schuster. Two categories, celebrity books and repeat bestsellers from the backlist, “make up the entirety of the publishing industry” from a revenue perspective – not only do most books not make money, “90 percent of them sold fewer than 2,000 copies and 50 percent sold less than a dozen copies.”
 
No surprise here – a study suggests authors greatly overestimate the quality of their own papers.
 
The research project everyone at The Brief longs to be a part of: trying to pet all 200 breeds at the Westminster dog show.

***
AI cannot simulate futures. If a machine can’t think ahead, it’s not an intelligent machine, just really good at running spreadsheets. – Warren Ellis (Orbital Operations)