How data integrity powers AI for patent searching

To realize an exciting breakthrough or new innovation, you need to see if a certain substance has been patented. Patent information is dispersed across many sources, and time is of the essence. You know the question you want to ask: “Which catalysts across the literature have achieved >90% selectivity for CO₂‑to-methanol conversion?” Should you enter it into a general purpose large language model (LLM) and see what you get?

‍

It may seem counterintuitive, but the answer is no. Standard LLMs can scour large amounts of publications quickly, but searching for patent information and prior art involves unique complexities that AI-powered tools aren’t always equipped to handle. This isn’t a simple “garbage in, garbage out” issue, either. For patent searches to be successful, researchers must use indexed, structured data with AI-enabled solutions that can navigate the intricacies of patent literature and reveal actionable insights.

Why patent search is different

Patents are written for a different purpose than journal articles or other scientific publications: to protect intellectual property (IP). Rather than aiming to publicize their content and be discovered and cited, as in journal publications, authors of patent applications seek to give only the amount of data that is absolutely necessary to successfully receive a patent. These applications may maintain a degree of secrecy about or even obfuscation of novel breakthroughs.

Patent documents may also contain a lot of “noise” in their early sections, such as long descriptions or lists of related materials or compounds that have no impact on the new innovation itself. These extensive lists are usually included to create fallback positions or articulate minor variations to prevent minor optimizations of the invention from being subsequently claimed as inventive. However, this can make a document appear to be relevant when it only superficially describes a substance or topic, confounding search processes.  

On top of this, patent literature lacks the standardization of scientific publishing. Documents are organized differently across jurisdictions, applicants, and classification systems, and their terminology often reflects organizational jargon or historical quirks rather than any shared convention. For example, in peer-reviewed scientific literature, the word “copolymer” most frequently means any polymer comprising two or more monomers. In patent literature, it often means a polymer with exactly two monomers, while the term “interpolymer” is used more broadly to refer to polymers with multiple monomers.

Patent information is therefore dilute, i.e., spread across a diffuse set of documents, and researchers must use “dilute” questions to search across that broad set with inconsistent terminology and generally vague language (think platform patents that attempt to broadly cover methods and technologies that are a foundation for further solutions). To return to our earlier example with the question about catalysts, let’s look at that in comparison to a “concentrated” question:

• Concentrated: What is the boiling point of acetonitrile?

• Dilute: Which catalysts across the literature have achieved >90% selectivity for CO₂‑to‑methanol conversion?

‍

How AI facilitates patent search — and how it doesn’t

It may sound like patent searches are the exact use case for generative AI and LLMs, since these tools can review huge sets of disparate documents across any database or even the entire internet. However, the complexity of patent literature makes it difficult for standard LLMs to return quality results. These platforms look for patterns, but as we discussed, patents don’t often adhere to patterns or even use the same language to describe the same thing.  

LLMs often struggle with dilute questions. They may return a robust answer, but researchers can’t be fully confident that the LLM understood all the types of sources to check or that it didn’t get tripped up by semantic differences. To verify patentability, researchers must evaluate patent filings and scientific literature together. To do this, there must be data harmonization so that AI tools, which can traverse more source material faster than a human possibly could, can review all relevant information.  

The key is having indexed information that is structured and consistent for the AI-powered solution to use. Without indexing of the material that an AI tool is searching, that tool is akin to someone being dropped in a library that has no Dewey decimal system and expecting them to find books on a specific topic.

Human expertise plus technological efficiency

Why is indexed data critical for AI-powered patent searches? Structured, indexed data allows LLMs and AI search tools to more quickly distinguish between reactants, catalysts, and other key data points without wasting time or energy. Data integrity also facilitates searching for relevant information across disparate sources.

For example, CAS indexes the CAS Content Collection^TM, the largest human-curated repository of scientific information, which creates a unified data layer for patent search. This unified layer enables comparisons between patents and scientific literature by normalizing different terminologies. Non-patent literature often contains workarounds and non-infringing methods that researchers must know to compare their innovation against existing patents. A unified data layer can also surface cross-industry connections and invalidating materials, which are all crucial for patent search but can get lost in different source materials or their semantics.

Ontologies are key to this unified data layer because they define canonical terms, synonyms, domain relationships, and contextual meaning. Without ontologies, every domain and even every company or organization would speak its own scientific dialect. AI requires the consistency that ontologies provide to avoid hallucinations and incorrect pattern recognition. However, it takes humans to develop ontologies and ensure that data is clean, structured, and consistent so AI tools can leverage that data properly.

This is what separates CAS data from other databases or search engines: applying human expertise to ensure data integrity allows powerful technological tools to uncover accurate insights faster. For example, CAS IP Finder™ delivers unmatched precision across highly structured content. Additionally, CAS Newton℠ brings conversational AI to patent and IP search, letting you ask questions in plain language and pulling answers from more than 100 patent and non-patent literature sources, with next-step, smart suggestions.

The future of patent search with AI

The ability of AI-powered tools and LLMs to search massive amounts of data quickly makes them standard, even necessary, for patent searching today. Yet, as we saw with the dilute nature of questions in patent search and the patent documentation itself, AI can struggle to find all relevant information and overcome differences in terminology. Without data integrity through ontologies and other forms of curation, even the most powerful AI tools can miss critical prior art.

AI will continue to play a role in all parts of the patent lifecycle, from exploratory searching and determining freedom to operate to performing infringement analysis and identifying oppositions. For researchers looking to advance their breakthroughs now, structured data that blends human expertise with the speed of cutting-edge technology remains the catalyst they need to drive innovation.