Note: this piece initially appeared as Helen King’s interview with Tim on PubTech Radar

As AI systems become serious consumers of research, scholarly publishing must rethink its formats, business models, and quality controls — because machine readers don’t want PDFs, they want structured, reliable science.

Large language models (LLMs) have already been trained on huge volumes of the written word; to an order of magnitude, everything humans have ever published online. As these systems mature, we will need to find new forms of input. If we want these tools to contribute meaningfully to scientific discovery, we need to rethink how research content is structured for their consumption and – necessarily – how it’s paid for.

The core limitation of today’s LLMs is that they operate by averaging. Their predictions are generated by assimilating vast numbers of tokens from existing texts and inferring what likely comes next. This approach may be useful in many domains, but it is ill-suited to scientific inquiry, where progress often stems from outliers – individual studies that take a new approach and reshape the consensus.

Scientific literature, moreover, is not of uniform quality. Some articles are foundational; many are derivative, flawed, or outright misleading. Yet because high-impact research is typically paywalled, and less rigorous work is often openly accessible, LLMs are disproportionately trained on lower-quality material. We’re asking AI systems to model the state of scientific knowledge by digesting the most available – not the most accurate – content.

To fix this problem we need to provide AI with the underlying raw materials of science, but what are the raw materials of science? It has to be an individual claim or conjecture: a set of conditions, a hypothesis that’s tested, and then a result in terms of both a dataset and a conclusion drawn from that dataset.

For the majority of research articles, the only accessible information comes as tables and figures that summarize the results, plus the text describing the authors’ interpretation of their results. Authors can – and often do – overstate the robustness of their conclusions and the significance of their work. Unless you’re very skilled and able to filter out these bloviations, it’s hard to determine the value and validity of the work from the words alone.

To enable AI systems to assist in research, we need to provide them with structured, reliable input. Not just the prose of published articles, but the underlying claims, data, and reasoning. A truly AI-ready version of a research paper might look very different from the human-readable PDF. It could be decomposed into modular units: individual hypotheses, linked datasets, methodological summaries, and direct mappings between evidence and conclusions. This kind of formatting would allow AI to assess the strength of specific claims and build a more nuanced understanding of the field.

The current publishing infrastructure is optimized for human readers. Research articles are delivered in a format that presumes a shared understanding (between humans!) of how to communicate findings; most commonly, via a PDF. This static medium persists because it enables a predictable exchange between sender and receiver. But AI readers operate under entirely different assumptions. They aren’t concerned with narrative flow or rhetorical nuance; they want structured input that enables reasoning.

This divergence opens the door to a new business model: offering AI-ready versions of research articles via a subscription. Human-readable content could remain open access, while publishers monetize machine-optimized content streams for institutional clients, particularly those running proprietary, in-house AI systems.

To give a concrete example of how this would work, imagine a pharmaceutical company subscribing to an AI feed that continuously ingests structured research about Parkinson’s disease, helping it identify promising leads in real time. In this scenario, access to high-quality, up-to-date content becomes essential, not optional.

The AI subscription model forces a renewed focus on quality control because the value of your journal depends on your ability to keep the fraudulent work out – nobody wants to train their AI on results that are fabricated or misleading. 

The missing piece here is the data that underpins these articles. How can we prompt authors to provide their datasets, and make them sufficiently well curated that they can be understood by an AI? It comes back to the business model behind open data – and in this case, the AI subscriptions really do provide a financial incentive for publishers to work with authors to ensure that their datasets are shared alongside the manuscript. Moreover, if publishers are going to be earning subscription revenue, then maybe the author should see some part of that in return for the extra curation work on their datasets. Perhaps they could get a reduced APC for the Open Access (= human) version of the article, or there’s some revenue sharing.

AI models are only going to get better and better. The new generation of reasoning models is already so impressive, but what they need to keep improving is knowledge about how the world works, and not just more information about the most likely next word in a sentence. The AI’s are not going to get that from the ‘words only’ version of human-readable articles that academic publishing currently makes, and that means there’s a huge opportunity for a new way to disseminate research.