I left Frankfurt Book Fair 2026 with more questions about how we use artificial intelligence (AI) than about what it can do. The conversations that stayed with me were about permissions, responsibility, and whether trusted content will actually be found and used in the workflows taking shape around it.
My work in AI and product strategy & development inevitably shaped which discussions caught my attention. These are reflections from that perspective, rather than a comprehensive account of the fair. Across publishing, research, and education, I kept returning to one question: how do we turn trusted content into useful knowledge — and help people act on it?
Shared priorities do not yet mean shared readiness — or trust
Research Information’s Scholarly Publishing: Trends for Constructive Growth report offered a useful counterweight to the product discussions. Respondents increasingly need greater understanding of AI, yet academic institutions are widely seen as underprepared. Research integrity is a shared priority, but industry respondents perceive more improvement in support than their academic counterparts. Relationships remain predominantly transactional or strained, with affordability, licensing, and reuse rights contributing to the tension. Financial constraints and misaligned incentives limit collaboration.
What struck me was the gap between agreeing on the priorities and agreeing on what progress looks like. For those of us building products, that is a reminder: investment and technical capability do not automatically translate into benefits that researchers and institutions recognize. We need clearer expectations and a better understanding of whose problems we are solving.
Content is becoming a foundation for connected, reusable knowledge
One question I kept returning to was what we lose when AI breaks a publication into smaller pieces. Making a claim easier to retrieve is useful — but only if its evidence, limitations, and context remain attached.
Uthereal’s “Knowledge Objects”, for example, organize source-linked concepts, claims, procedures, and standards for reuse. ElevenLabs and Cadmore address a different aspect of accessibility through audio creation and audiovisual content management. Together, these examples prompted me to think beyond the publication as a self-contained document.
In The Anchor and the Drift: Academic Publishing in the AI Age, Cashmere and Johns Hopkins University Press described a shift toward consuming concepts and data snippets rather than whole documents. My interest is in the connections around those fragments: what supports a claim, what contradicts it, and where it can be applied.
That also brings us back to submission. Publishers should rethink which metadata and supporting information are mandatory — not only for editorial assessment, but for downstream reuse, practical application, reproducibility, and replication. The aim should be to preserve useful research context, not simply collect more fields.
AI discovery needs a permission and value model — not just a technical connection
KnowledgeWorks Global’s (KGL’s) AI Gateway, Elsevier’s connectors, and Molecular Connections’ Tessera illustrate different approaches to governed AI access to scholarly content. The question for me is less about comparing their features than understanding what makes that access useful and sustainable.
In his licensing presentation, Roy Kaufman of Copyright Clearance Center (CCC) highlighted legal uncertainty and inconsistent approaches across jurisdictions, presenting licensing as a practical way to agree on permitted uses. That points beyond connectivity to permissions, attribution, reporting, and payment.
Cashmere contrasted one-off training arrangements with inference licensing’s potential for recurring revenue and controlled access when content is used to generate answers. Hopkins’ principles favored defined uses, attribution, and ongoing payments. I took this as an argument for deliberate contract design, not a universal rule about which licensing model is best.
At the same time, I wonder how durable a clean separation between training and inference will be. Research into test-time training (TTT), which enables AI systems to learn from inputs during use, suggests that the technical boundary may be less straightforward than these labels imply. That does not make the two activities equivalent, but it raises a question for future licensing frameworks: should we focus more on what a system actually does with content — including whether it uses that content to adapt, and whether that adaptation persists — rather than relying on the label “training” or “inference”?
A working connector does not guarantee that content gets selected. Whether publishers build their own connections or use shared gateways, they need to consider how users and their agents discover, authorize, and choose them. Model Context Protocol (MCP) infrastructure is only part of that challenge. How clearly are capabilities described, and how does a connector reach relevant workflows? The competitive question is not just “Can AI reach our content?” but “When and why will it choose to use it?”
This also shapes my view of “zero-click” discovery. Lost traffic and attribution matter, but I’m not convinced that zero click must mean zero value. As agents help compare evidence, analyze problems, and complete tasks, trusted content could contribute beyond the initial search. That might create new licensed usage even without a website visit. It is an opportunity, not a guarantee — and it requires us to measure contribution beyond clicks.
Trust must be built into both the product and the infrastructure
“Trust” was one of the words I heard most often at Frankfurt. Beyond peer review and integrity, the question was whether people could confidently work with AI-powered tools.
Charlie Rapple and Kudos’ Taming the Crocodile findings make the concern tangible. Roughly a third of respondents begin searches through AI tools or search engines featuring AI answers. Forty percent mistakenly believe search-engine AI overviews are human-created or curated; librarians report requests for nonexistent articles. That is a literacy problem as well as a reliability problem.
TRACE — Trusted Retrieval & Attribution for Content Ecosystems — is worth watching here. It connects Ithaka S+R’s work on agent discovery and access, the National Information Standards Organization’s provenance and attribution pilot, and COUNTER’s AI usage reporting. More consistent source connections and off-platform reporting could help, but emerging infrastructure is not a guarantee of accurate answers. Participation, adoption, and timely implementation will determine its impact.
Roy also reminded us that much open access publishing rests on licenses whose terms matter, and that licensing can encourage use of the Version of Record. For me, this connects permission with provenance.
His discussion prompted a broader question: can organizations confidently take responsibility for AI-enabled work? Humans make mistakes too; reliable performance must be accompanied by clear responsibilities, oversight, and ways to investigate and remedy failures. I suspect uncertainty about these arrangements can slow adoption. The challenge is not to make AI legally accountable in its own right, but to make its use accountable through the people and organizations behind it.
Connected publishing should reduce friction and move quality checks earlier
The discussions around KGL’s Smart Publish, MPS, and Silverchair reflected interest in more connected publishing workflows. Their portfolios differ, but the practical ambition I heard was similar: cost saving, less fragmentation, rework, and complexity.
I would start by moving useful quality checks earlier. Preventing avoidable problems from traveling downstream should benefit authors, editors, reviewers, and production teams — not just make one department faster. Earlier checks should help authors get things right, rather than simply give them more forms to complete.
Moving checks earlier also makes me question what we are checking. If AI-assisted fabrication becomes more convincing and those producing fraudulent papers adapt to familiar checks, could the absence of obvious warning signs give us false reassurance? AI-assisted writing is not itself evidence of fraud, just as polished prose and accurate references do not establish that the underlying research is genuine. That leaves me wondering what the next generation/level of screening should look for — and how it can help us assess the evidence behind a paper, rather than simply detect suspicious features. For me, “no issues detected” should never be confused with “research verified.”
Success, then, should mean both less friction and better-supported editorial and integrity decisions. Fewer incomplete or unnecessarily returned submissions, less repeated correction, and shorter decision and publication times still matter — but not at the expense of meaningful scrutiny. Product design should help authors, editors, and reviewers do their work more effectively, rather than optimize only publishers’ internal processes.
A connected workflow need not mean one supplier for everything. It should ensure that supporting evidence, check results, and unresolved concerns travel with the manuscript, with clear responsibilities and reliable handovers between systems and teams.Those are the outcomes I would look for beneath an “end-to-end” label.
Research products are moving from finding information towards supporting decisions
The research tools raised a related question: are we helping people find more information, or helping them decide what to do with it?
Elsevier describes LeapSpace as an AI workspace for academic and corporate researchers. Wiley’s Applied Research Intelligence, introduced in beta, emphasizes evidence-based decisions for corporate research and development (R&D) and data teams. These examples suggest a shift from “Which papers match my query?” toward “What does the evidence mean for this problem?”
For that second question, I would want more than titles, keywords, and abstracts. Concepts, methods, experimental conditions, limitations, and conflicting findings can determine whether research is applicable. Full-text access and traceable evidence matter because a useful answer must preserve more than the headline result.
Corporate R&D interests me particularly where reducing uncertainty or avoiding unproductive work could create substantial value. But the proposition needs to be demonstrated through better decisions and research outcomes — not assumed from enterprise budgets or the size of a content collection.
AI agents require new ways of working, not just new capabilities
One question I left Frankfurt with was what vendors meant by “agentic.” From so many vendors presentations and materials, I could not always tell whether their systems are real agentic system which orchestrate tools, run self-reflection, and adapt their next steps, or just integrate AI/LLM in their existing workflow. A simpler workflow may be exactly right for the job. What matters is clarity about the autonomy offered, its safeguards, and the outcomes it improves.
Straive’s discussion of moving beyond isolated pilots and my own presentation brought these operational questions into focus.
I would start with a specific problem and define success, failure, and exceptions before increasing autonomy. What may an agent do? What evidence must it use? Which decisions remain with people? When must it stop or escalate?
Verification, visibility into actions, and careful management of what agents remember are essential. So are redesigned roles, training, and evidence of value that helps colleagues develop informed confidence.
Counting deployed agents tells us little. Quality, exception rates, human review effort, and the cost of successfully completing a task seem much better measures of progress.
Education is moving towards support for the learning journey
The education discussions brought me back to the person using the content. My notes span augmented and virtual reality, peer communities, personalized pathways, bite-sized modules, and integration across platforms and media. The common question is how these approaches help someone learn.
EdTool connects interactive content with teaching and assessment. EBCONT’s work with Verlag SKV links structured publisher content to multimedia learning and the NAVE AI learning coach, which is designed to draw on verified educational material. The latter interested me because it connects content quality with personalized support rather than treating them as separate concerns.
Generating more material or adding a chatbot is not, by itself, the outcome. I would look for a stronger connection between explanation, practice, feedback, and human support. Whether an experience is conversational, immersive, or community-based, does it improve understanding and help learners apply what they have learned?
The thread connecting these developments
For me, Frankfurt’s most important message was about the connections around knowledge: between a claim and its evidence, a use and its permission, and an automated action and the people responsible for it. I left encouraged by the possibilities, but more interested in how we will judge their value.
Author’s note: Generative AI was used as an editorial aid in summarization, reviewing drafts and copyediting. The ideas, arguments, source selection, and final decisions are entirely the author’s own.
