How AI Prior Art Search Works, Where It Misses, and What the Law Counts as Prior Art

AI prior art search is the use of machine-learning retrieval models to find documents that may disclose a claimed invention before its effective filing date. This explainer covers semantic search, which ranks documents by closeness of meaning rather than by shared keywords, and sets it beside Boolean keyword and classification searching. Whether a returned document is prior art at all is a separate question, answered by 35 U.S.C. § 102 and Federal Circuit case law. More is on the Tech & AI hub and the Explainers hub.

How does semantic search work?

Semantic search turns text into numbers and compares the numbers. A neural network called an encoder maps each passage of each document to a vector, a fixed-length list of numbers; the query is encoded the same way, and the system returns the passages whose vectors lie closest to the query’s. Because passages are encoded in advance, a search is a lookup in a pre-built index rather than a fresh reading of every document.

That is the design in the 2020 paper introducing Dense Passage Retrieval: encoders produce 768-number vectors, similarity is the dot product of query and passage vectors, and every passage is indexed offline (pp. 2–3). Its premise is that “synonyms or paraphrases that consist of completely different tokens may still be mapped to vectors close to each other,” such as “bad guy” and “villain” (p. 1). The paper evaluated open-domain question answering, not patents.

How is that different from Boolean keyword and classification searching?

Keyword searching matches the words themselves. A review of patent-retrieval research by Walid Shalaby and Wlodek Zadrozny, published in Knowledge and Information Systems, describes keyword methods as looking for exact matches between query terms and documents, so that they “fail to retrieve relevant documents which use different vocabulary but have similar meaning” (pp. 3, 13). A Boolean query adds operators such as AND, OR and NOT to set which terms must appear together or be absent.

The two approaches miss different things. The Dense Passage Retrieval authors found that term-matching methods “are sensitive to highly selective keywords and phrases,” while their dense model “might lack sufficient capacity to represent salient phrases which appear rarely” (pp. 7, 12). They reported that combining the two improved results in some cases (p. 5).

Classification searching works from codes rather than words. The official site of the Cooperative Patent Classification (CPC) describes a common, internationally compatible classification system “for technical documents, in particular patent publications.” The Shalaby and Zadrozny review explains that such schemes are hierarchical, that codes are assigned by the technical features of the invention, and that this metadata is language independent (pp. 4, 21).

What do recall and precision mean in a prior-art search?

Recall is the share of all relevant documents that a search retrieves; precision is the share of retrieved documents that are relevant. That is how the TREC evaluation appendix published by the National Institute of Standards and Technology computes them, describing recall as the ability “to present all relevant items” (p. 1).

Prior-art searching sits at the recall end. Shalaby and Zadrozny call it “a total recall task” and add in a footnote: “It is required to achieve 100% recall at acceptable precision” (p. 9). Research systems are scored at cut-offs, such as recall within the top 10, 100 or 1,000 results, and the review describes a measure, PRES, proposed for recall-oriented tasks that weighs recall against review effort (p. 12). Recall at a cut-off counts only what appears above it.

Why can a semantic search miss a relevant reference?

Vocabulary is the gap semantic search was built to narrow, not one it removes. The review observes that inventors “tend to use different vocabulary to express same or similar ideas” (p. 9). In its own error analysis, the Dense Passage Retrieval paper gives an example where keyword ranking did better because the dense model could not capture the salient phrase “Thoros of Myr” (p. 12).

A model that encodes text has nothing to match where a disclosure exists only in a drawing. Patent documents “contain multiple data types (e.g., text, images, flow charts, formulae...etc)” (review, pp. 3–4), and the CLEF-IP 2011 evaluation ran image-based retrieval as a separate task (p. 2). The law counts what a drawing shows. In In re Mraz, 455 F.2d 1069 (C.C.P.A. 1972), the reference’s specification “says nothing about the angle,” but its Fig. 3 showed a groove angle within the claimed range, and the court repeated that “Description for the purposes of anticipation can be by drawings alone as well as by words” (at 1071–72).

A document outside a tool’s index cannot be ranked at all. The review lists among the challenges of prior-art search covering “all previously published material (patent or non-patent literature) in all forms (electronic or printed) which is infeasible,” and searching “materials written in different languages” (p. 9). Section 102(a)(1) refers to a “printed publication” with no language qualifier, and the cases below concern a doctoral thesis at a German university, a conference poster and a university technical report.

Date filters can cut the wrong way. Section 102(a)(1) reaches disclosures made “before the effective filing date of the claimed invention.” Section 102(a)(2) separately reaches patents and published applications naming another inventor that were “effectively filed” earlier, and § 102(d) dates them from their actual filing date or, where they claim priority or benefit, from the earliest application describing the subject matter. A filter on publication date set at the filing date can therefore exclude an application published later that is prior art under § 102(a)(2). In the other direction, § 102(b)(1) excepts certain disclosures made one year or less before the effective filing date, including the inventor’s own.

What counts as a “printed publication” under § 102(a)(1)?

A document that was publicly accessible before the relevant date. In In re Hall, 781 F.2d 897 (Fed. Cir. 1986), the court wrote that because “there are many ways in which a reference may be disseminated to the interested public, ‘public accessibility’ has been called the touchstone” (at 898–99). A reference is publicly accessible if it was “disseminated or otherwise made available to the extent that persons interested and ordinarily skilled in the subject matter or art, exercising reasonable diligence, can locate it” (Acceleration Bay, slip op. 11, quoting Jazz Pharmaceuticals v. Amneal Pharmaceuticals).

In Hall, a doctoral dissertation reached the Freiburg University library in November 1977. The court rejected the argument that “a single cataloged thesis in one university library” was not sufficiently accessible (at 898–900).

Indexing is not required. In In re Klopfenstein, 380 F.3d 1345 (Fed. Cir. 2004), a 14-slide poster was displayed for about three days in all at a scientific meeting and a university; no copies were distributed, and it “was never catalogued or indexed in any library or database” (at 1347). The court held that “the key inquiry is whether or not a reference has been made ‘publicly accessible’” (at 1348), weighed the length of display, the audience’s expertise, any reasonable expectation against copying and the ease of copying (at 1350), and held the poster a printed publication (at 1352).

Hall, Klopfenstein and Acceleration Bay applied the version of § 102 in force before the America Invents Act (Acceleration Bay, slip op. 11 n.6); the current § 102(a)(1) keeps the words “printed publication.”

Can a document a search tool finds today fail to qualify as prior art?

Yes. In Acceleration Bay, LLC v. Activision Blizzard Inc., Nos. 2017-2084 et al. (Fed. Cir. Nov. 6, 2018), Judge Moore, for a panel with Chief Judge Prost and Judge Reyna, affirmed the ruling under review that a 1999 technical report was not a printed publication. It had been uploaded to a university department’s online library by 23 November 1999, but the site listed reports “indexed only by author or year,” and the Board found, on evidence the court held sufficient, that its advanced search was not reliable (slip op. 11–13).

The court agreed with the Board that “public accessibility” requires more than technical accessibility: “The test for public accessibility is not ‘has the reference been indexed?’” The question is whether an interested artisan exercising reasonable diligence would have found the document (slip op. 12, 14).

Read together, the cases separate two questions: whether a search tool can surface a document now, and whether persons skilled in the art could have located it before the relevant date. A poster that no database ever indexed was prior art in Klopfenstein; a report sitting on a website was not in Acceleration Bay.

What does checking a returned reference against the document involve?

The cases point to three things a ranked list does not settle. The first is what the document discloses, read from the document itself, figures included. Mraz relied on a drawing, but it also repeated that “Patent drawings are not working drawings,” and distinguished reliance on “a greatly enlarged section of a small drawing obviously never intended to show the dimensions of anything” (455 F.2d at 1072).

The second is when the document became publicly accessible, a question of fact (Acceleration Bay, slip op. 11). In Hall, evidence of the library’s routine practice “may be relied upon to establish an approximate time when a thesis became accessible” (781 F.2d at 899); in Acceleration Bay an upload date did not establish public accessibility.

The third is which provision applies and from what date: § 102(a)(1) for publications, or § 102(a)(2) and (d) for earlier-filed patents and applications, subject to the exceptions in § 102(b). The guide to reading a patent covers the parts of a patent document, and the explainer on patentability covers novelty and inventive step.