In the opening moments of the HighWire Best Practice Webinar on Protecting Publisher Content in the Age of AI, one theme emerged repeatedly from the questions attendees were asking even before the first speaker began: publishers know they need to respond to the rapid rise of generative AI, but few feel they have a clear roadmap. What exactly should scholarly publishers be doing to protect their content, their authors, and the integrity of the scientific record when large language models (LLMs) can absorb and remix vast bodies of literature with no visibility, no attribution, and no guarantees of accuracy?
Pascal Hetzscholdt, Senior Director of AI Strategy & Content Integrity at Wiley addressed this question directly in his presentation. Pascal’s approach was concrete, detailed, and grounded both in Wiley’s institutional experience and in the voices of thousands of researchers whose needs Wiley has actively studied. He offered not just a conceptual model for “Responsible AI,” but a practical operating framework that positions publishers as stewards of ethics, accuracy, and trust in a fast-moving technological landscape.
In an AI-powered era where scientific content can be distorted, misrepresented, or quietly decontextualized at scale, Pascal argues that publishers must define the standards for responsible use before someone else defines them for us.
A Disruption Arrives Faster Than Anyone Expected
Pascal began by noting the speed of change that has blindsided even organizations long involved in machine learning. Wiley, he emphasized, had years of experience with ML technologies. But when ChatGPT launched in late 2022, the effect was unlike any prior innovation. “The sudden introduction of the chatbot,” he said, “being used in a matter of weeks by hundreds of millions of users, has indeed been very disruptive for everybody active in what we call ‘critical sectors.’”
Critical sectors, healthcare, legal, finance, and scientific research, are environments where accuracy is not optional. In these domains, an AI-generated conclusion is not entertainment; it is actionable guidance, sometimes with life-changing consequences. The question of how to safeguard the fidelity of information therefore becomes an urgent one.
Yet the scale of enthusiasm among researchers is undeniable. Wiley’s latest researcher survey shows widespread adoption of AI tools for tasks such as summarization, translation, and literature exploration. Researchers want help navigating the overwhelming volume of new publications. They want assistance writing in a second language. They want support synthesizing complex topics or identifying contradictory findings. They want speed, but they also want reliability and they are increasingly uneasy about whether AI provides it. The survey revealed deep concern about the accuracy of outputs, especially when models cannot reveal the provenance of the information they generate.
The Heart of the Matter: Accuracy Depends on the Integrity of Training Data
One of the most striking moments of Pascal’s talk came when he revealed a fact that many in the industry had suspected, but few had quantified:
Major AI models have been trained on – at minimum – tens of thousands of retracted scientific papers. Retracted articles, by definition, contain flawed, falsified, plagiarized, or otherwise invalid research. When LLMs ingest these papers indiscriminately, they learn from corrupted data. The downstream effect is that the models may confidently generate incorrect claims, incomplete interpretations, or misleading conclusions that contradict the scientific consensus.
Even worse, Pascal noted that some models have been trained on data scraped from the dark web, including breaches of personal information and other ethically problematic datasets. When foundational models draw on such material, they inherit its distortions. The public cannot see those influences, their outputs merely appear as fluent, authoritative prose.
This is where Responsible AI must begin: with ethical, transparent training data. Without clarity about what a model was trained on, stakeholders cannot evaluate its reliability.
Wiley’s licensing strategy follows this principle closely. The company licenses its content to AI developers only under agreements that ensure scholarly material is used responsibly, legally, and with proper attribution and traceability. Pascal made it clear: “We want to make sure that AI models can train on reliable, verifiable, and updated information.” Without such agreements, publishers have no visibility into how their content is used and no way to correct misrepresentations.
The Case for Explainability: AI Must Not Be a Black Box
Even if the training data is sound, AI outputs can still go wrong. Models may conflate studies, misinterpret scientific nuance, or generate inaccurate citations. Pascal explained that these problems often stem not only from training data but from the system prompts that shape how models behave. These prompts, written by AI developers, instruct models to follow certain conventions or behave in particular ways. But they are also written in natural language, meaning they can conflict with the user’s prompts or with the model’s own reasoning.
For example, a model may be simultaneously instructed to “always answer confidently” and “cite references,” even when relevant references are missing or inadequate. The model resolves such tensions by improvising, resulting in fabricated citations or overly confident misinformation.
Explainability is essential for evaluating whether a model is fit for use in scientific contexts. Wiley requires model cards and documentation that describe:
- what data a model was trained on
- how it was fine-tuned
- what known failure modes it exhibits
- what biases or limitations may affect outputs
- what steps were taken to ensure appropriate use
Pascal stressed that researchers, peer reviewers, and editors need to contest, verify, and understand AI outputs. Without this transparency, models become black boxes that masquerade as authoritative voices, undermining scholarly communication.
Human Oversight: Keeping Humans in the Loop
Pascal emphasized that human expertise must remain central. Wiley’s policies reflect this in multiple ways:
- Generative AI tools cannot be listed as authors.
- Authors must disclose how AI was used in their manuscripts.
- Peer reviewers and editors must verify claims rather than accept AI outputs at face value.
- Attribution, fact-checking, and bias mitigation remain human responsibilities.
This human-guided paradigm is encapsulated in Pascal’s concept of the Triad of Trust, replacing the older “dyad” of expert and individual researcher. The new triad consists of: individual ↔ AI assistant ↔ human expert.
In this triad, AI plays a supporting role, not a determinative one. Experts are not displaced, they are repositioned. Instead of simply delivering authoritative judgments, they increasingly help users interpret AI-generated insights. They act as validators, guides, and guardians of context. They help people distinguish between evidence-based conclusions and AI-generated speculation.
Pascal described a shift already underway in clinical care as an illustrative warning. Patients now come prepared with ChatGPT-generated explanations of their symptoms, and many challenge their doctors using AI-generated arguments. These conversations, he said, “will get more tough in the future” because patients often trust AI more than human experts.
The Triad of Trust aims to restore balance by empowering experts to engage with AI outputs rather than ceding authority to them.
Protecting Scholarly Content: Licensing, Rights, and Guardrails
One of Wiley’s strategies is the Wiley AI Gateway, a platform that allows developers such as Perplexity, Claude, Mistral, and AWS to retrieve scholarly content directly from Wiley through licensed, ethical channels. Users can prompt these AI tools and receive verified research results sourced from Wiley’s publications, with clear indication of where the information originated.
This stands in direct contrast to the unlicensed scraping that has characterized much of the AI training ecosystem to date. Pascal emphasized that publishers must both protect their intellectual property and ensure that models treat scholarly content responsibly. Unauthorized ingestion risks producing outputs that are inaccurate, incomplete, or misleading. Licensed ingestion, by contrast, allows publishers to update content, correct errors, enforce attribution, and monitor usage.
He also stressed the importance of auditing AI models. Just as models must be re-evaluated following corrections or content updates, AI outputs must be monitored for accuracy, security risks, and privacy issues. This includes evaluating misuse scenarios such as impersonation and disinformation, areas where generative models have already demonstrated troubling vulnerabilities. For publishers, this means developing internal expertise in AI oversight and building partnerships with developers who demonstrate ethical practices.
Sustainability: The Hidden Cost of AI
A dimension of AI risk not explicitly related to scholarly publishing is its environmental footprint. Training and operating large models consume vast quantities of electricity and water, so much so that some regions have reported spikes in utility costs when new AI datacenters come online.
Pascal urged publishers, especially those committed to open science and global equity, to factor sustainability into their evaluation of AI tools. Responsible AI should include measuring and mitigating carbon emissions, investing in greener computing infrastructure, and supporting standards that promote energy-efficient model design.
This is not incidental to academic values. Scientific publishing is part of an ecosystem concerned with human wellbeing, global development, and evidence-based policymaking. AI’s environmental burden cannot be ignored simply because its outputs seem abstract or digital.
A Vision for the Future: Responsible AI as Competitive Advantage
In closing, Pascal offered both a warning and a promise. AI, he said, will not replace publishers. But publishers who embrace Responsible AI, and who insist on ethical, explainable, traceable use of their content, will have a competitive advantage over those who do not. They will help democratize access to scientific information without compromising rights or accuracy. They will keep human insight at the center of research communication. And they will contribute to shaping the regulatory frameworks that govern AI’s use in critical sectors.
Responsible AI, then, is not merely a philosophy. It is a practical necessity for ensuring that scientific knowledge remains reliable, interpretable, and anchored in the values that underpin scholarly publishing. Wiley’s framework, ethical training data, explainability, and human oversight, offers a model for the industry at large. Publishers must shape the future of AI, or they will be shaped by it.
– By Tony Alves
