Highlights include the major OpenReview breach that leaked thousands of ICLR submissions (and revealed how much AI is already being used in peer review), a wave of new launches from VeriMe’s identity verification service to Elsevier’s LeapSpace workspace, thoughtful pieces on search evolution, AI transparency policies, competing visions of AI in education, and the ongoing debate about how AI should be used in peer review.

🆕News

• 🦹 A bug in OpenReview allowed author, submissions, reviewer details, and their reviews to be leaked, and a dataset of 10,000 ICLR submissions circulated. The Glass Wall Shatters: A Professor’s Reflection on the ICLR 2026 Breach is a good summary of the implications and the fragility of the peer review system. I’ve only just twigged that this is where the data for Pangram Labs’ “21% of the 75,800 peer reviews for ICLR 2026 were fully generated by AI” came from.

• 📈 This stat from Richard Wynne caught my eye “...during the next three years scholarly journals will send out an additional 150 million invitations to volunteer peer reviewers OVER AND ABOVE the current baseline”. I think you can quibble with the calculations behind this stat, but it’s still a thought-provoking number.

• 🙄 Springer Nature flags paper with fabricated reference to article (not) written by Retraction Watch’s cofounder Ivan Oransky. Retraction Watch says “The paper with the nonexistent reference, published November 13 in DARU Journal of Pharmaceutical Sciences, criticizes platforms for post-publication peer review — and PubPeer specifically — as being vulnerable to “misuse” and “hyper-skepticism.” Five of the paper’s 17 references do not appear to exist, three others have incorrect DOIs or links, and one has been retracted. ”

• 📅 UX in Publishing is a new cross-industry community of practice for anyone interested in user experience in academic publishing. Next meeting 8th December, 2:30 pm GMT (online) on Demonstrating impact and Linking UX to business goals.
  

🚀 Launches

• VeriMe is a new global digital identity verification service for researchers run as a limited cooperative association based in Colorado, US. You can take their survey to help them better understand experiences with identity management and privacy.

• openRxiv announced is piloting a reviewing tool from q.e.d Science to generate rapid AI-generated feedback on biomedical manuscripts. See also: AI reviewers are here — we are not ready

• Elsevier has launched LeapSpace, the “...integrated workspace will combine the broadest collection of trusted scientific content with responsible AI to help researchers move from curiosity to discovery faster” [Press release].

• Google Scholar has launched Scholar Labs, an AI-powered Scholar search that is designed to help you answer detailed research questions. [Google’s blog post][Aaron Tay’s review]

• A British start-up, Books By People, is introducing an ‘Organic Literature’ certification to help readers identify books written by humans rather than machines.

• 🎧The Informed Frontier. A new podcast “examining innovation at the intersection of science, technology, and media. Season 1 explores AI as a ‘reader’ of content, and how this fundamental shift is changing the way we access, interpret, and share information. Hosted by Jessica Miles, each episode looks at the philosophical considerations and practical consequences of this disruption, focusing on scientific and scholarly content.”

• 🎧 Deanta’s Trends in Academic Publishing podcast features Patrick Shafe, talking to Jennifer Regala (Wolters Kluwer Health), Tim Williams (Edward Elgar Publishing), Nicola Ramsey (Edinburgh University Press), Gillian Shasby (Journal of Neurosurgery) and Darren Ryan (Deanta).

• Scholarly Futures is a new weekly Substack from seasoned publishing voices, offering shared pondering on AI, integrity, policy shifts, and the unknowable roads ahead.

🤖 AI

• Ann Michael‍ from AIP Publishing on What We Learned from the Peer Review Assistant Trial. “Citation checking, metadata validation, and ethics compliance stood out as areas where automation added genuine value. … Not every feature landed perfectly. The manuscript digest didn’t always capture a paper’s nuance, and there were some latency issues - the full report sometimes took longer than ideal to generate. Addressing the latency issue is a top priority to allow a more effective fit into editorial workflows.”

• 📈 Nick Morely, from Veracity, asks: What is the cost of citation checking? “Est 75s (1.25 mins) average per ref (or 41 clicks) x 30 refs = 37.5 mins per paper (1230 clicks)”.

• According to Amazon, “less than 5% of titles” on its site are available in multiple languages, so it’s launched Kindle Translate.

• “Gross, disrespectful, and infuriating! The Association for the Advancement of Artificial Intelligence (AAAI)’s move to include “AI”-produced reviews into their selection process is UNACCEPTABLE.” says Dagmar Monett on the introduction of AI reviews. You can’t always please everyone, but you can, like AAAI, be transparent about the process and how AI is being used, as the FT article below explores.

• I think Tricky Trade-Offs on a Transparency Spectrum: How the Financial Times Approaches Transparency about AI Use in News is a helpful read for anyone thinking about policies around the use of AI and what to consider.

• Notes from Nancy Roberts about Deanta’s roundtable about The Future of AI and Its Implications for Academic Publishing. I think this point does need a lot more thought: “Google Zero – the idea that Google’s AI summaries will soon eliminate the need for anyone to click through to a website – is a concern for any of us who publish high value, peer reviewed content. There are no easy answers here but it seems to me that we need to think about the implications of this as an industry, and in conjunction with the wider ecosystem; how do we measure impact of academic research without the traditional metrics of downloads and citations? What are the implications for our publishing platforms? For academic libraries? For our authors?”

• 📈 Steve Cramer in This Liaison Life highlights this stat and beautifully diplomatic reporting language from the UX Team in Design and Discovery at the University of Michigan Libraries survey: “Undergraduate students led overall AI usage at 72.1%, graduate students followed at 62%, while faculty trailed at just 39.4%. Librarian participants reported the lowest usage of AI at just 34%. This inverse technology adoption pattern means librarians have opportunities to deeply understand student AI usage and to integrate much needed guidance for more ethical and effective use of AI tools within library instruction and consultations.”

• The Journalism Benchmark Cookbook is creating a community-oriented AI benchmark for journalism. Is anyone creating community benchmarks for LLMs in journal publishing? Seems like a good idea?
  

🔎 Search

• I enjoyed Michael Upshall’s summary of the Search Solutions Conference, especially the closing point: ““Is AI making search better?”. One respondent put it nicely: “Search has fundamentally changed because of AI, but I am confused about what search is now”. ”

• “We’re Good at Search”… Just Not the Kind That the AI era Demands - a Provocation. I think this is a must-read about the uncomfortable reality that traditional search expertise doesn’t help you evaluate AI-powered tools - and librarians are looking at the wrong metrics entirely.

• Andrew White on how The Edison API can be used to convert PowerPoint presentations to scientific claims and then diligence each claim to see if they are contradicted in literature/patents/clinical trials is a nice demo of how scientific agents can be used.

📚 Longer reads

• Short but challenging read about trust and the internet by Jimmy Wales [Behind a paywall]: “The challenge of our time is not that information is scarce but that authenticity is. Important aspects of the early internet succeeded because people could trace what they read to another human being, even if the other human being was operating behind a pseudonym.
  The new internet must restore that chain of custody. We are entering an era when machines can mimic any voice and invent any image. If we want truth to survive that onslaught, we must embed transparency, independence and empathy into the digital architecture itself. The early days of the web showed it could be done. The question is whether we still have the will to do it again.“ 🎧Or listen to HBR’s interview with Jimmy Wales on the same topic.

• Book Publishing & Technology: The ONIX 3 Debacle by Thad McIlroy, “Will the management of publishing companies figure out ways to harness the power of AI, while mitigating its drawbacks? There are some positive early indicators. But publishing’s historic discomfort with technology seems likely to block the industry from robustly integrating AI into current workflows. The implications are concerning.”

• 🎧TWiST podcast interviews Medium CEO Tony Stubblebine on the company’s new Really Simple Licensing (or RSL) initiative, and how it’s helping to compensate writers for their content and Human Native CEO Dr. James Smith on AI licensing, why a writer’s content isn’t just DATA to them, and why the company is pivoting away from its old marketplace model.

• Deepak Varuvel Dennison, a PhD student at Cornell University, explores the knowledge that is missing from LLMs and the complexity of including that knowledge: Generative AI has access to a small slice of human knowledge. “The marginalisation of local and Indigenous knowledge has long been driven by entrenched power structures. GenAI simply puts this process on steroids.”

• CUP’s Publishing Futures report is an insight into how the community is thinking about the future. 📈 I’m not sure that the future of traditional journal publishing is all that hopeful if only “...54% agreed that the peer review system was still effective in ensuring the quality of published research”.

• Two recent white papers show how differently people see AI in Education. Google’s AI and the Future of Learning is pure Silicon Valley optimism. The authors see AI solving education’s resource problems through personalisation at scale. Their vision: AI tutors providing “high-dosage” support while freeing teachers for “meaningful human interaction.” They acknowledge problems like hallucinations and cheating, but mostly as bugs to fix. For Google, education’s messy human problems just need better tech. Tom Chatfield’s AI and the Future of Pedagogy comes from a completely different world. Citing a survey showing that 92% of British undergraduates already use AI tools, Chatfield worries about what he calls “distant writing,” humans designing prompts while LLMs generate text. A way of looking productive without actually doing the thinking. Where Google celebrates AI handling the “cognitive load”, Chatfield worries about “metacognitive laziness,” students losing the ability to think about their own thinking. Learning, he argues, cannot be automated. “You cannot outsource the construction of knowledge any more than you can outsource physical exercise.”

🔗 Extra links

• Extra links from my backlog that didn’t make it into the main issue. More reading for the truly committed

And finally…

How to Spot a Poison Book: I had no idea that some Victorian publishers used arsenic-based green dyes in their book covers. Suddenly, going to the library seems like a much more dangerous and exciting activity - at least AI-generated content can’t actually poison you!

End Notes

• If you found this useful, you can always buy me a coffee.

• If you’re interested in AI, you can subscribe for free to my other Substack, GenAI for Curious People, to see entries for my new book.

• If you need consulting help navigating any of this, find me at Maverick.

• Looking for my AI-powered use cases for Academic Publishing? Find the list here
  

🆕Recentish news/opinion

• Open Intelligence - by Adam Hyde on Why it is time to see open movements as one shared project

• The Great Turning Point: a Return to Publishing’s Roots - by Ravi Venkataramani, CEO of Kriyadocs

• Doug Way summarises Clarivate’s Q3 earnings call from a librarian’s perspective. “The second was an emphasis on agentic AI tools and the development of native AI tools, like Alma Specto. Specto seems (to this librarian) to be a competitor to JSTOR’s Digital Stewardship Services. When asked specifically about it, leadership sidestepped details and instead discussed their general approach to new products and a couple of other initiatives in more depth. This could mean absolutely nothing—as many companies don’t always answer questions directly, sometimes for business reasons and sometimes because extemporaneous speaking is hard. Just something to file under “interesting.””

• Innovation in scientific publishing and its implications for Crossref DOI registration practices - Request for input

• Is Digital-first Publishing Finally a Reality? An Interview with Liz Ferguson of Wiley

🤖 AI

• Ian Mulvany, Some thoughts on AI Agents.

• AI tools combat paper mill fraud in scientific publishing as peer review system struggles by Nina Notman.

• Ethan Mollick “If the the predictions that AI will create minor scientific discoveries next year and major ones a couple years later come true, it is worth noting that we have no real mechanism in academia for accommodating, reviewing, processing, and disseminating a sudden increase in science. Who is going to read thousands of new papers that marginally advance science? Who is going to read thousands of papers that represent genuine breakthroughs in narrow fields? Who will integrate the knowledge? Who is going to build on them to transfer them into practical products?”.

• When AI Feeds on Poisoned Knowledge — Lessons from Library Genesis to Llama 3.

• Matt Rogerson commenting on Alex Reisner’s article in The Atlantic about Common Crawl.

• We’re not taking the fact-checking powers of AI seriously enough. It’s past time to start. By Mike Caulfield.

• AI Scientist by Andrew White on the history of the term, and gives some ideas on what might define an AI Scientist.

• Beyond the Artifact: The Brutal Economics of Liquid Content by Shuwei Fang. “The cold reality is that when content becomes infinitely replicable and reformattable at near-zero marginal cost, the economic value of any individual piece approaches zero. I believe this will be the major forcing function that drives structural change in the existing news media industry. In an AI-mediated information ecosystem, few existing publications will be able survive on current business models. New upstarts are already emerging, building businesses around capabilities and infrastructure, changing the competitive environment. News media organizations face a stark choice: radically innovate or risk extinction.“

• BadScientist: Can a Research Agent Write Convincing but Unsound Papers that Fool LLM Reviewers? [See also Chris Leonard’s/Claude’s review of this paper ]

• Fauzia Burke on Here’s what authors and publishers need to know about AI-powered book suggestions.

• Large Language Model Relevance Assessors Agree With One Another More Than With Human Assessors.

• 🎧 A conversation with John Frechette, Ben Kaube, and Aaron Tay. (ATG Podcast) “This conversation explores how AI is reshaping the research process - from how papers are discovered to how findings are analyzed, shared, and rewarded. John talks with Ben and Aaron about the new realities of AI-driven search, publishing, and research workflows. Together, they discuss where automation helps or harms, how publishers and librarians are adapting, and what the rise of AI means for research quality and integrity.”

• 12 lessons from news outlets on the cutting edge of AI - key points, ideas and tips from the first day of the JournalismAI Festival in London.

This is the second post in Big Ideas in Publishing, exploring concepts that could reshape research communication.

In July, I spoke with Tim Vines about AI subscriptions and how publishers might create AI-optimized content as new revenue streams. This time, I’m interviewing Jonathan Woahn, whose startup, Cashmere, has built infrastructure to give publishers control over how their content works with AI systems. As Jonathan puts it, they provide tools to manage entitlement, security, access, reporting, and turn-key Retrieval-Augmented Generation infrastructure from ingestion to Model Context Protocol.

What led you to start Cashmere?

Back in 2022, I was building an AI tool for corporate training based on leadership books. It was fun until it wasn’t. The AI couldn’t stay consistent. Even a program built around The 7 Habits of Highly Effective People would come back with four habits one time and ten the next.

That’s when I realised there was no proper infrastructure for publishers to control how their content gets used in AI systems. This broke down into two core issues. First, long-form content doesn’t work great with AI—there’s too much context. There needed to be a better way to help the AI only get access to the content it needs, when it needs it. That led me down the path to inventing Omnipub, which is the foundational way we model data at Cashmere and how we present it to AI.

The second issue was from the publisher's perspective. They didn’t have any tools to control their data with AI. Licensing was long and complicated, the terms were not standardized, and the use cases varied. For example, using content to train an AI model is very different from when that AI quotes or references the content in its responses to users - but publishers lacked clear frameworks to handle these different scenarios. So we realized premium content publishers needed rails for their content with AI—without it, they were going to lose their position.

The obvious question is who should build this infrastructure. Why couldn’t big tech or publishers build this themselves?

If big tech builds it, why would publishers trust it? They’ve already used premium content at scale without permission. Amazon eroded control from many publishers over distribution and strong-armed pricing across print, audio and digital. AI is a new distribution channel—if big tech builds it, it feels like yet another moment when publishers lose control. This is publishers’ moment to gain it back by controlling how their content is used with AI. So if not big tech, then the publishers?

Individual publishers may build their own solutions, and if they have the resources, it could work. But it’s going to be a full-time effort, and not a cheap one. Building with AI comes at a price and requires skillsets that are more in demand now than ever. So now publishers are competing with big tech for talent. Let’s just say publishers aren’t known for building bleeding-edge technology.

How does Cashmere solve this problem?

We’re the infrastructure layer between publishers and AI companies. Publishers set the rules - what content can be used, how, and at what price. AI companies integrate once and get legitimate access to all that licensed content. We handle the authentication, tracking, and reporting automatically.

We don’t necessarily handle the financial settlement part - we can, but we also allow our partners to handle that on their own. The key is that publishers keep control while AI companies get the content they need without negotiating hundreds of separate deals. In short, we make premium content AI-ready and make sure everyone in the chain gets credit and compensation.

What’s making this urgent for publishers?

AI is replacing the need for users to access publisher websites or purchase content. If the AI has been trained on it, then the users have access to it for free. Publishers are already feeling it. Some academic houses we’ve spoken with report double-digit textbook declines as students shift to AI-based learning. This is the brief window before legal precedents are set around AI content licensing. Publishers can either help shape how this works or watch it happen to them.

Is there anything publishers can do?

Napster allowed users to get access to basically any music they wanted for free. The music industry had to evolve. I call it a Great Reallocation. The demand didn’t disappear, it shifted - from full albums to tracks, from media stores to iTunes, then streaming. Once music moved into legitimate channels, the industry recovered. Today it’s bigger than ever, driven by streaming.

This same thing will happen for content. The market is already shifting from Google-based discovery to access via chatbot conversations. But we need more granular licensing mechanisms. The legal environment is crystallising - using purchased content to train AI models might be fine, but when AI quotes or references that content in live responses, that will likely require specific licensing. For example, publishers need control over how their content gets used in each case - and that’s exactly the infrastructure Cashmere provides.

You’re already working with Wiley and Perplexity. How does that work in practice?

The partnerships prove this infrastructure works. With Wiley, we support two models. The first is Bring Your Own License: if your university (or anyone who has a license) already pays for Wiley content, that access travels with you into Perplexity, like adding Spotify on your iPhone—you already have the Spotify account, and now you’re provisioning new channels to access the content. The second is Metered Token Consumption. Every time Perplexity uses Wiley content to answer a query, the usage is metered in tokens, and Wiley is paid proportionally.

Both models show that legitimate participation is possible and profitable.

If you were advising a publisher today, what would you tell them?

Start with strategy, not technology. Decide where you want to play, training, inference, or both, and how much visibility you need. Then make sure your content can move with full control through those channels. The key mindset shift is to see AI as a new distribution channel, not a disruption.

Cashmere’s partnerships prove these licensing models work. The Wiley deal shows publishers can extend existing subscriptions into AI systems while keeping control over how their content gets used. I asked Jonathan where he sees this heading. What will an AI-native publisher look like five years from now?

What will an AI-native publisher look like five years from now?

I think in 5 years, AI agents are going to act heavily on our behalf. They’re going to know how to find the content we want and need. They’re going to be our tutors. They’re going to be able to dig into any topic at the level of depth of our understanding and take us to the very edge of human knowledge.

And in all of this, the agents are going to respect the creator control of content. They’re going to automatically transact on our behalf in the background. As a user, you’ll have a license for the content you care most about (i.e., Times subscription, academic journals, industry specialized textbooks, trade novels, etc.) that you either purchase directly as a patron or your agent discovers and procures on your behalf (with authorization).

Your agent will use that data to tailor the content to your needs. It will provide sources for everything, so you can know it’s legitimate. You may choose to click on sources because you’re curious, or perhaps you don’t. For the publisher/creator, it won’t matter. They get the attribution, their content has been legitimately procured, and the AI is acting fully within the bounds the consumer and publisher have authorized it to do.

The biggest challenge to solve in all of this will be about discovery. That will be one of the beauties of having an autonomous agent working on your behalf that has the context of your interests, activities, etc., because it will be able to continually monitor the market on your behalf, tirelessly scouring for information that is relevant to you and your interests, and will surface it when it knows you want/need it.

You buy a print book, and it automatically unlocks AI access. Libraries let you explore their collections conversationally. Publishers know exactly where their content travels, readers own their licences, and AI platforms become just another interface.

To see how Cashmere is building this infrastructure, visit cashmere.io.

Welcome new subscribers! This newsletter covers publishing technology with my usual idiosyncratic take - entries reflect what caught my eye, nothing’s sponsored. This issue is 100% AI-focused to tackle a backlog after a bit of a delay in publishing. Next issue returns to the usual mix of launches, news, and analysis. It shouldn’t be quite so long either.

🆕 News/New to me

• Holtzbrinck opens AI Hub in San Francisco to bridge its European publishing empire with Silicon Valley’s AI ecosystem. The German media giant will test AI innovations in science, education, and media while scouting investment opportunities

• Russell Michalak at Goldey-Beacom College Library shares a nice story about how JSTOR’s Seeklight helped students describe 172 historical photographs in 4 hours rather than 250 per year. The change, from authoring every word to reviewing and governing at scale, hints at how AI is going to change the work many of us do.

• Meet the industry’s top innovators and trailblazers as they present their latest breakthroughs at STM Innovator Fair 2025 on December 9, in London. This year’s theme is “Building Tomorrow’s Research Integrity Framework”.

🚀 Launches/newish products

• Scopus AI launches Deep Research, an agentic AI feature that performs comprehensive literature reviews automatically. When users ask research questions, the system develops detailed research plans, conducts dozens of searches across Scopus literature, reviews findings, and refines its strategy as insights emerge. The tool produces downloadable reports that Adrian Raudaschl (Principal Product Manager at Elsevier) says can “save days of preliminary research.”

• Refine is a research-focused AI tool that aims to offer deep, substantive feedback for academic and technical authors. Unlike grammar tools or compliance checkers, and distinct from broader peer-review simulators like Reviewer3 or ReviewerZero, Refine targets the logic, internal structure, and consistency of a manuscript. (H/T: Aaron Tay)

• Nicia.ai has launched an AI assistant for peer review. They claim 95% time savings and position it as “agentic AI” that can make autonomous decisions within defined parameters, continuously learning from each review while always deferring to human judgment on matters of scientific merit and impact.

• Paper2Agent converts computational research papers into chat agents you can interrogate with questions like “apply your method to my data”. Only works with well-structured, public code repositories, takes 30 minutes to 3+ hours per paper, and costs ~$15 per conversion. Successfully demonstrated on three bioinformatics papers, reproducing original results and answering new scientific queries. (This is an early proof of concept, but an interesting direction) [GitHub | Medium article | Preprint | Comment]

• Daniel E. Acuña’s team developed an AI classifier that flagged over 1,000 questionable journals from 15,191 open-access titles, achieving 76% precision while generating 345 false positives. The system, described in Science Advances, analyzes website design and bibliometric features using DOAJ criteria. ReviewerZero, Acuña’s startup, plans to commercialize the tool as ‘Journal Monitor’, part of a suite targeting research integrity specialists.

• Wiley AI Gateway has launched. More on this in a future issue when I’ve had a bit more time to explore.

👎Problematic AI

• AI chatbots are citing retracted scientific papers without warning users, according to MIT Technology Review. Researchers found ChatGPT-4o referenced retracted papers in five of 21 test cases about medical imaging, advising caution in only three instances. A separate study showed ChatGPT-4o mini evaluated 217 retracted papers without mentioning retractions or quality concerns in the responses.

• Iris van Rooij’s “AI slop and the destruction of knowledge“ looks at AI-generated misinformation about cognitive science concepts on ScienceDirect.

• Kent Anderson describes how he created a fake 15,000-word research paper in 20 minutes using Science42: DORA from insilico.com. His demo study on “music and mental health” contained detailed methodology, SPSS analysis, and 119 references, all from a three-word prompt to show how easy it is to create a fake paper. (The same technologies that accelerate legitimate research are, of course, going to be used to flood journals with sophisticated fraudulent papers).

• 🎥 Dr. Kimberly Becker’s YouTube short “The Dark Side of AI Research Tools“ examines how AI can misleadingly amplify scientific evidence through decontextualized citation stacking, creating a “compound interest” effect where misrepresented studies get incorrectly built upon.

🤖 AI and Peer Review

• 🚨New acronym alert: ASPR (Automated Scholarly Paper Review) from a paper examining AI-powered peer review.

• Research output has increased 212% since 2004, but reviewer capacity hasn’t kept pace. At the September AI Publishing Collective meetup, Chris Leonard, Wiley’s Sam Parker and Rebecca Windless, and Digital Science’s Leslie McIntosh explored how AI might address this challenge. See also Laura Harvey’s excellent write-ups here and here.

• If you need some brightness in your life, check out Oleg Ruchayskiy’s LinkedIn post on how academic publishing will die when LLM-based peer review takes over 🙂.

• Roohi Ghosh interviews Christopher Leonard and me in The Scholarly Kitchen on “Peer Review in Transition”, exploring how AI will reshape peer review.

• The AI Peer project is assessing whether LLMs can score academic journal articles for research quality. Initial findings show LLM scores align positively with expert judgements across almost all fields, with stronger correlation than citation rates achieve. However, the researchers caution that LLMs are “guessing” at scores rather than truly evaluating, likely by matching abstract claims to quality definitions - evidenced by similar results whether using full texts or just titles and abstracts. [LinkedIn post about initial findings]

🔮 Future thinking

• Knowledge as a commodity: If scientific publishing is shifting from being the endpoint of research to just one step in a much bigger process, then the entire role of the publisher is fundamentally at stake, according to David Worlock. Worlock asks as scholarly communications migrate from content to facilitation, which organizations will researchers trust to support their work? (I wouldn’t bet on publishers winning an arms race against big tech players. I might bet on a society brand beating a big tech player for niche but essential tools tailored to their specific community.)

• Jon Treadway’s post on AI-first publishing is excellent, and I agree with much of what he’s written, but I think what AI-first research looks like is more interesting than AI-first publishing. In my quickly put together vision, researchers use AI tools to explore literature, build personal RAG (or whatever comes next) models, let AI handle study design and paperwork, deploy voice agents for interviews, have AI analyze responses, generate papers in their style, and automatically produce reports. This is technically possible today - the key question is who controls the underlying technology, see David Worlock’s article above.

• Phil Jones offers a measured three-year retrospective on ChatGPT’s impact on scholarly publishing, tracking the steady evolution from early hype to today’s practical applications in writing tools, discovery platforms, and research integrity.

📚 Longer reads/listens

• 📰 Lee Konstantinou in The Chronicle of Higher Education describes peer-reviewing a humanities paper where every single quote from the novel being analyzed was completely fabricated. While fraud discussions usually focus on the sciences, AI-generated fakery is hitting literary scholarship too. His account captures the messy human reality of being the reviewer who has to check every citation and wrestle with sympathy for what’s likely a desperate contingent academic.

• 🎥 Wiley launches “Chats” video series on AI in academic publishing, featuring BMJ’s CTO Ian Mulvany and Wiley’s Ray Abruzzi navigating practical AI challenges for publishers and researchers. The conversation covers prompt engineering tips, publisher partnerships, and AI hallucination misconceptions.

• 🎧 O’Reilly’s “Generative AI in the Real World” features Ben Lorica interviewing Faye Zhang on how foundation models are transforming recommendation systems. Zhang discusses the shift from traditional next-click prediction to models that can forecast user behavior over 30-day windows, moving beyond simple content matching to understand long-term preferences. The conversation covers how companies like Pinterest are leveraging these advances in their recommendation engines

• 🎥 Emily Bender, one of AI’s most prominent critics, argues “We do not have to accept AI (much less AGI) as inevitable in education“ in a recent UNESCO talk.

• 📰 Mark Piesing explores how AI writing tools are subtly changing authorship in The Bookseller. While some writers embrace AI assistance, others strongly resist, highlighting the concern that AI assistants designed to help writers may actually be training writers to conform to algorithmic preferences rather than developing their own voice.

“I’ve never knowingly used any AI tools while researching my books,” says Alex Larman (Spectator books editor), “because I am arrogant – or experienced – enough to believe that I will always have a better connection with my subject than an algorithm can. I have no doubt that AI will change writers’ lives beyond recognition, but am unable to see this as a good thing; instead, I fear that it will steal the soul from this most human of professions.”

• 📰 Tomasz Tunguz explores “PageRank in the Age of AI”, arguing that content distribution will soon mirror online advertising architecture. Instead of advertisers bidding to show ads, publishers will compete in real-time auctions to have their content included in AI responses. AI systems would evaluate submissions based on quality metrics - relevance, accuracy, freshness, and authority - rather than payment, creating “PageRank for real-time AI responses” operating in milliseconds.

• 📰 Michael Upshall asks, “Have LLMs overtaken humans for knowledge organization?“ following the ISKO-UK conference. Most speakers emphasized AI limitations, except Tony Russell-Rose’s presentation on 2dsearch, which demonstrated how GenAI can automatically handle complex term selection with better results than manually curated resources.

And finally…

I adore this, LibraryThing’s @librarycat.org is battling AI scrapers, so they built a cat- and literary-themed captcha.

End Notes:

1. If you found this useful, you can always buy me a coffee.

2. If you’re interested in AI, you can subscribe for free to my other Substack, GenAI for Curious People, to see entries for my new book.

3. If you would like to employ me, you can find me at Maverick.

4. Looking for my AI-powered use cases for Academic Publishing? Find the list here.

📰 News

• Wiley has announced that 1,000 scholarly journals have successfully transitioned to its Research Exchange platform, representing more than 50% of the company's journal portfolio. [I’ve always been a fan of REX and the thinking behind the service. It is no easy feat to move journals to a new peer review platform]

• This LinkedIn thread and post by Chris Dicker, CEO of CANDR Media, explains how Trusted Reviews was scraped 1.6 million times in a single day, 18.5 times a second, allegedly by OpenAI, despite having a robots.txt file in place. Out of 1.6 million scrapes, we saw just 603 users arrive on site, a CTR of 0.037%, Chris wants to know who is going to pay for this? [I have an interview with Jonathan Woahn from cashmere.io on one approach to solving this problem coming shortly.]

• Open infrastructure is crumbling under political pressure and vanishing funds. IOI’s 2025 report sounds the alarm and calls for urgent, collective action

• A call for the US federal government to adopt persistent digital identifiers (DOIs for research awards and ORCIDs for researchers) across all agencies. The article outlines the problems of the current reporting burden, explains how digital identifiers would simplify everything, and then lays out steps to make it happen.

• Sticking with the Identifiers theme, the main Crossref systems have [finally] moved to the cloud 🎉!

• Effective 8 August 2025, the Center for Open Science (COS) has suspended new submissions to its generalist OSF Preprints server because of an increasing number of suspicious or low-quality submissions and other reasons. See also Academia has a new preprints problem by Mark Hahnel.
  

🚀 Launches

• Near Missives a newsletter from Chris Reid & Laura Harvey “It is often said that there are two ways to learn about a topic: from deep inside and from the outside. As much as we should all be listening to and learning from our peers and colleagues, we wanted to create a way to bring ideas, concepts and trends from other industries to a scholarly comms audience. With that thought, Near Missives was born.”

• The Thomas Kuhn Foundation has developed the KGX3 engine, a tool that can predict the significance of scientific research before it is even published. Instead of waiting decades to see whether a paper is “paradigm shifting” or simply solid, normal science, KGX3 analyzes research in real time to reveal whether it confirms, stresses, or breaks existing ways of thinking. Using proven signal-detection methods similar to those behind Bloomberg, Dataminr, and Brandwatch, KGX3 wants to transform the scientific record into a dynamic map of emerging ideas, helping researchers and decision-makers spot the structural tensions that precede major scientific breakthroughs.

• posters.science, a new open-source and free platform where researchers can easily share, find, and reuse posters.

• Copilot Mode in Edge includes multi-tab RAG. You can use Copilot to analyze your open tabs, like Satya Nadella does here with papers his company has published in Nature journals over the last year.

• AlphaXiv transforms dense arXiv papers into blog‑style summaries in seconds. Just swap ‘arxiv’ with ‘alphaxiv’.

• Clear Skies is making genAI detection with Pangram available to subscribers in Oversight. Adam claims it’s “GenAI detection that actually works”.

• Veracity from GroundedAI is now available as a standalone, self-service tool, which includes Citation Stacking Detection and Citation Quality Assessment.
  

🤖 AI

• Editors using the FT's internal CMS now “get a suggested description automatically – short, accurate, and context-aware – generated using a mix of machine vision and agency captions. Every suggestion is flagged for fact-checking and reviewed before publication to ensure accuracy.” [LinkedIn Post]

• This Science Advances study uses a novel ‘excess word’ analysis to track how AI, especially ChatGPT, has reshaped the language of biomedical research. It uncovers a measurable shift in vocabulary across over 15 million PubMed abstracts, revealing that at least 13.5% in 2024 were likely LLM-assisted.

• Researchers have tried to create tools that check whether a specific piece of text (like a paragraph) was included in a model’s training data. These tools are called membership inference attacks (MIAs), but they can be unreliable. Rather than trying to prove whether one paragraph was in the training set, dataset inference focuses on whether the model was trained on a larger dataset, like a whole book, better matching real copyright concerns, since authors usually care about their whole work being used, not just one page.

• I don’t know this group, so DYOR, but I thought it looked interesting: workshop for Artificial Intelligence for Scientific Publications (WASP) 2025 in December.

• Michael Upshall explores Thomas Krichel’s publication alert system, which is in use both for economics (where it is entitled NEP, short for New Economics Papers), based on the content in RePEc, and for life sciences (where it is called Biomed News, or BIMs), using content from PubMed.

• Rishabh Lohia contrasts two models of digital publishing: one where users must search and sift through static content, and another where content actively adapts and responds to the user’s needs.

• I saw this and thought perhaps publishers don’t need to abandon their cherished spreadsheets: Why Paradigm built a spreadsheet with an AI agent in every cell

• Great thread by Aaron Tay on the development of Academic Search on Bluesky. Aaron has been publishing some excellent articles on Deep Research, AI Search, Deep Search, and how they work recently - subscribe to his Substack.

• An academic, Liudmila Zavolokina, discovered her name cited in fake papers published by a shady journal filled with AI-generated articles, often authored by ghost writers (or no one at all), filled with hallucinated references.

• Video showing how the Wiley/Anthropic content deal works:

📚Longer reads

• 📰 The em dash has had enough. The em dash speaks out in its own defence, claiming centuries of usage and lamenting that it’s now unfairly viewed as a marker of soulless AI writing. It cites usage by literary giants—Mary Shelley, Emily Dickinson, David Foster Wallace—and scolds the accusers for not reading enough.

• 🎥 A 2017 The Great Automatic Grammatizator: writing, labour, computers by Martin Paul Eve led me to Roald Dahl's excellent 1953 ‘Great Automatic Grammatizator’, a story about an automatic writing machine in which human writers are replaced by the machine. [YouTube reading of story]

• 📰 Subscribe-to-Open Is Doomed. Here’s Why by Rick Anderson. This post reminded me that around ten years ago, I explained Gold Open Access to a tech audience by comparing it - loosely and imperfectly - to a pyramid scheme. Publishing carried two kinds of costs: the one-off expense of producing a journal with a carefully managed number of papers, and the recurring expense of keeping the archive and related services online. In the subscription model, this is straightforward; stability is provided by many subscribers, and ongoing subscriber revenue covers both categories of cost + profit. In the Gold OA unlimited papers model, each new paper must carry its own production costs, help offset the costs of rejected papers, and also contribute to maintaining the growing archive of everything published before. I have always wondered how sustainable this model is over the very long term. Storage costs may decline, economies of scale kick in, but labour, compliance, infrastructure, and the rising expenses linked to AI, from detection to blocking unwanted scraping, are all increasing recurring costs. All of these models have problems.

• 📰 Stocking the Librarian’s Publishing Integrity Toolkit via Fabienne Michaud

• 🎥 NPTV Unscripted: In conversation with Martin Delahunty about Publishing, AI, and what the future holds.

• 📰 The PDF may no longer be the untouchable container it once was. But with foresight and action, scholarly publishers can still shape what comes next by Pascal Hetzscholdt and ChatGPT-4o. Interesting because of the changes to PDF but also because most Western publishers haven’t relied on PDFs for decades, and I think you can really see the limits of AI writing in this article. [I know this sounds daft, but I think the differences between HTML and PDF format online are really poorly understood by many users… I remember troubleshooting one librarian’s problem and realising that they didn’t realise that they were looking at an HTML page rather than the PDF… sadly, this wasn’t that long ago...]

• 🎧 Midnight at the Casablanca interviews Timo Hannay. There’s a section in here where Paul is asking about business models for some of the early work, and Timo’s response is a bit vague. It’s hard to appreciate now just how experimental web tech was then, how easy it was to put something together quickly, and how much Annette Thomas supported the team’s freedom to experiment.

Almost finally…

Thought-provoking from Hannah Shelley, Google Scholar Is Doomed. I’m inclined to agree, not sure about the timeline, but a couple of years after Anurag, the founder, leaves? Mind you, 25 years is an awfully good innings for a technology product.

End Notes:

1. If you’re interested in AI, you can subscribe for free to my other Substack GenAI for Curious People, to see entries for my new book.

2. If you found this useful, you can always buy me a coffee

3. If you would like to employ me, you can find me at Maverick

4. If you’re in London on Monday, 22 September, join us for the next AI in Publishing Meetup. Registration form coming soon.