Executive Summary
- Structured search is bounded by the question. Curated databases return authoritative, precise results, but the query sets the limit. A researcher finds what they already knew to look for, and the adjacent fields where rich connections may be hiding stay out of reach.
- Language models describe scientific relationships without encoding them. Prose is fluent about how entities relate, yet it carries no verified record of those links. A model working from text alone also has no reliable way to resolve the many names a single substance travels under.
- Some discoveries are already sitting in the published record. Don Swanson’s 1986 link between dietary fish oils and Raynaud’s syndrome came from reading two bodies of literature together that no single paper had joined. He called it undiscovered public knowledge, and neither keyword search nor a language model reliably surfaces that kind of connection.
- Verified structure is what makes the reasoning checkable. Substances, reactions, properties, biological targets, and patent claims connected as a graph give a language model something to reason over rather than something to describe. Every answer traces back to the peer-reviewed literature behind it.
There is a difference, of course, between doing something and writing about doing something. You can be a lousy cook but a good food critic or vice versa.
In science, writing is part of the work. Without publishing, discoveries do not enter the shared record. Without reading, researchers cannot build on what came before. Scientific progress is, in an important sense, a conversation carried out through published literature.
These conversations have become more complex as the volume of knowledge has grown. A researcher in a pharmaceutical company is trying to answer a question that sits at the edge of two fields:
Does anything in the published record on membrane transport proteins suggest a mechanism that might explain the anomalous solubility behavior in their lead compound series?
The question is real. The relevant science exists, distributed across journals in biochemistry, physical chemistry, and pharmacology. Every available search tool will return something, but none will provide the most useful result. This is not because the answer is missing, but because traversing the terrain between those fields reliably has been challenging and time-consuming.
That gap is what our team at CAS has spent years working to close. Understanding why it exists requires looking at how scientific search has evolved, and where two dominant approaches today each fall short.
When the science is there, but the question limits what you find.
Traditionally, the most powerful way to search scientific knowledge has been through structured, curated records. The data is authoritative. The results are precise. The limitation is the question. To a great degree, that means the results depend on how well a user understands the field they are querying and how it intersects with other fields.
The paradox there is that the more time you spend gaining deep knowledge in your own field, the less time you have to understand other adjacent areas where rich connections may hide.
This is not a new dilemma. Scientists have always produced more knowledge than any individual can absorb. From the early journals that circulated among scholarly societies in the seventeenth century onward, the central challenge of scientific research has been making discoveries accessible to the people who need them.
By the early twentieth century, the sheer volume of chemistry literature had outpaced any individual researcher's ability to keep track of it. CAS was founded in 1907 to meet that challenge through systematic abstraction and indexing: summarizing, categorizing, and organizing the world's scientific output so researchers could find what was relevant to them. The introduction of the CAS Registry Number® in 1965 solved a related problem: the same chemical substance referred to by dozens of different names across papers written in different languages and research traditions. A single, unambiguous identifier for every known substance made searches more precise and results more trustworthy. CAS ONLINE launched in 1980, enabling researchers to search the CAS REGISTRY® remotely for the first time, followed by CAS SciFinder® in 1995.
However, these advances, for all their power, shared a common structural limitation. A researcher forms a question, the system returns relevant documents, and the researcher evaluates and refines the question.
A traditional database system can be extraordinarily good at surfacing what is known. What it cannot do is surface connections that have never been made explicit. For example:
- Financial systems can provide access to more market data than any human could process and return exactly what you ask for with extraordinary reliability. What they will not do is volunteer that an unusual pattern in shipping lane delays correlates with credit default swap movements in a sector the analyst had not thought to examine.
- Real estate software can access every listing in every market with detailed specifications but can’t tell you that a factor driving appreciation in one market is starting to appear in another.
- Legal databases can search by topic, jurisdiction, and precedent with precision, but can’t tell you that two different rulings in tax law had similar environmental implications unless you already know to query both domains.
The data is rich. The constraint is the interface: a mirror that reflects your question back at you with precision will be bound by what you are able to articulate.
Multi-step research workflows run autonomously. Vast reference sets become clear, actionable direction. As findings emerge, hidden relationships and trends surface across the CAS Content Collection™ without requiring you to manually piece results together.
The paradox
The more time a researcher spends gaining deep knowledge in one field, the less time remains to understand the adjacent areas where rich connections may be hiding.
When the interface is intuitive, but the science is uncertain.
The arrival of large language models (LLMs) seemed to change this paradigm. Suddenly, a researcher could ask a question in plain language, the way they would ask a colleague, and receive a fluent, synthesized response that drew on enormous bodies of text across many different fields.
This is a real advance for many uses. However, problems quickly emerged in areas where precision was required.
LLMs can read prose fluently, but that prose describes relationships rather than encode them. A model trained on millions of papers learns how experts talk. What it does not have is a verified record of how known, defined entities are linked. In research contexts where precision is non-negotiable, that distinction matters.
AI models will often confidently state mistakes or fabrications in ways that make it difficult for users to spot. For example, in April 2026, a law firm had to apologize publicly for “inaccurate citations and other errors generated by artificial intelligence.” While the firm in question said it has policies and training practices in place to govern the use of AI tools, those policies weren’t followed, nor did a review identify the AI-generated inaccuracies. While these models can review the related literature, they can also write fluently about it without referring back to actual facts or evidence.
There is also a structural problem specific to science. The same compounds and substances are referenced under different names: systematic IUPAC nomenclature, common names, brand names, abbreviations, and transliterations from other languages. A large language model working from text alone has no reliable mechanism to resolve those synonyms into a single entity. Two patents, three journal articles, and a toxicology dataset may all be describing the same substance. The model may treat them as distinct, or collapse them inconsistently, depending on what patterns it encountered in training. Furthermore, a result that sounds plausible but is not grounded in verified scientific relationships is unhelpful at best. In drug discovery or materials research, it can be misleading and even harmful.
The conversational experience is better than structured query languages in many circumstances for most users. The limitation is what sits underneath it: unstructured text, from which relationships can be inferred but not verified. Language models are a powerful layer for conversation and reasoning. What they need is something verified to reason over.
Spanning over 150 years of scientific discoveries from around the world published in journals, patents, and a wealth of other sources, the CAS Content Collection provides a comprehensive foundation to inform confident decisions throughout your innovation process.
Serendipity is not a strategy
In 1986, researcher Don Swanson published a paper demonstrating something that remains striking. He showed that dietary fish oils and Raynaud's syndrome (a condition causing restricted blood flow to the extremities) were likely related, despite no paper having ever linked them.
His insight came from reading across two bodies of literature that hadn’t earlier been “part of the same conversation.” Fish oils, he found, reduce blood viscosity and inhibit platelet aggregation. Reduced blood viscosity and platelet aggregation, in turn, are associated with improved outcomes in Raynaud's patients. The connection was hiding in plain sight, distributed across papers that had never been read together.
This happens again throughout history. Many significant scientific breakthroughs arise from recognizing that concept A, well-established in one area, is connected to concept C, well-established in another, through some intermediate concept B that both bodies of literature discuss. The discovery is the connection itself, latent in what is already known, waiting to be surfaced. Swanson called this "undiscovered public knowledge."
This is a problem that keyword search, by design, cannot solve. A search for fish oils returns papers about fish oils. A search for Raynaud's syndrome returns papers about Raynaud's syndrome. No query surfaces the implicit relationship between them unless a researcher already suspects it exists. Neither type of search tool handles this well. Structured database search cannot surface a connection that the researcher did not already suspect, because the query bounds the result. Language model search can generate plausible text about fish oils and Raynaud's syndrome but cannot verify whether the causal links it describes are real, because those links are not encoded in the prose.
What Swanson did once, reading across two literatures, forming connections using his own scientific judgment, is what an AI system built on structured scientific knowledge can do at scale.
Key takeaway
Many significant breakthroughs arise from recognizing that concept A, well established in one area, connects to concept C in another through an intermediate concept B that both bodies of literature discuss. The discovery is the connection itself.
150 years of the chemistry conversation
This is where history becomes relevant. This year, the American Chemical Society marks its 150th anniversary. ACS was founded in 1876 to give chemists a place to share their work, and by 1907 that conversation had grown so large that ACS launched Chemical Abstracts, the start of CAS, to help researchers keep up with it.
Over more than a century, CAS scientists have indexed papers and encoded the science inside them. True discovery and its reliability depend on access to the relationships within and between documents.
Every substance, reaction, property, biological pathway, and patent claim in the CAS Content Collection has been tagged, linked, and connected as nodes in a structured graph of scientific relationships. A compound’s record connects it to the reactions that produce it, the properties that characterize it, the biological targets it interacts with, and the patents that claim it.
The CAS Registry Number, which has uniquely identified every known chemical substance for over sixty years, is a foundation for exactly this kind of reasoning. When the same compound appears in a pharmacology paper, a toxicology database, and a patent filing under three different names, CAS REGISTRY ensures they are recognized as the same entity. Without that, cross-domain discovery is impossible. With it, a system can follow a conceptual thread across the entire published record of chemistry and life science.
Expert CAS scientists read, classify, and connect the literature, turning raw text into a structured record of substances, reactions, properties, and relationships that have been verified rather than inferred. The expertise built into this collection over decades is what makes precision searching possible. Each advance in how CAS organized and structured the record has deepened the same asset: a verified record that can be searched and reasoned over.
Signal, not noise. Curated by the team behind the world’s largest human-curated scientific data collection.
Where conversation meets science
CAS Newton is built to navigate this terrain. It pairs language models with the curated structure of the CAS Content Collection, deploying multiple agents that operate the way scientists work together: combining existing knowledge, testing connections, and looking for patterns that no individual could see.
Rather than requiring researchers to formulate queries aimed at narrowing relevant documents, CAS Newton turns questions into answers grounded in the structure of science itself. The output is a reasoned answer, drawn from multiple lines of evidence, traceable to the peer-reviewed literature that CAS and ACS have been building for 150 years, and capable of exploring questions no previous tool could readily answer.
The practical difference is significant.
- A medicinal chemist without deep intellectual property training can ask a freedom-to-operate question in plain language and receive an answer grounded in verified patent claims, with citations traceable to the underlying record.
- A researcher pursuing a target in a neglected disease can ask a multi-part question that crosses chemistry and biology, and CAS Newton decomposes it, traverses the relevant graph, and returns a reasoned answer with evidence.
- A materials scientist optimizing a polymer formulation can ask how properties achieved in one class of materials might transfer to a different application space, surfacing synthesis routes and property combinations that exist in the literature but have never been connected.
In a 2026 survey, 83% of users trust the results CAS Newton provides more than those from general AI tools. That figure reflects something structural: when an AI system reasons from verified relationships rather than from text descriptions of those relationships, the answers it produces can be checked. Every output is traceable to the peer-reviewed literature behind it.
Today, CAS Newton combines existing research in new ways and at speeds no human researcher could match. That means you can surface potential answers that are latent in the content, but no previous tool could reach. The CAS scientists who curate and maintain the collection remain central to what makes this possible. AI extends the reach of their expertise.
Or as Glaucius Oliva, a senior professor at the Universidade de São Paulo working on malaria and arboviral infections, observed, “CAS Newton becomes a sophisticated medicinal chemistry partner, enabling researchers without deep chemistry training to navigate complex design challenges with confidence.”
The question you haven’t even thought of asking? The one sitting at the intersection of what you know well and what is encoded in the content CAS has curated? Now you can ask it. When you can ask questions like that and trust the answers, latent connections in the published record become new, active scientific knowledge you can build on.
The foundation of curated science that ACS and CAS have built over 150 years now goes far beyond search based on what you already know. With CAS Newton, you can have a conversation with the data, ask broad questions, and find unexpected connections, all of which are grounded in precise, curated, scientific knowledge.
By the numbers
