Software and AI Patentability Search: A Practical Guide
A practical framework for technical-character screening, patent and non-patent literature searching, online-date verification, and jurisdiction-specific review.
A software and AI patentability search must address two connected questions: what earlier technical teaching may affect novelty or inventive step, and whether the proposed claim fits the patent-eligibility framework of each target jurisdiction. Patent databases remain essential, but papers, preprints, public code, standards, manuals, and other online materials may reveal relevant teaching earlier or in different language.
The method therefore combines feature-based searching with an evidence review. It also keeps eligibility analysis separate from novelty and inventive step. A promising search result does not decide eligibility, and a technically framed invention is not automatically novel or inventive.
The dual challenge in software and AI patentability search
Challenge one: patentable subject matter
Under EPC Article 52, programs for computers are listed among excluded subject matter, but only to the extent that a European application relates to that subject matter “as such.”1 The EPO guidelines explain that a computer-implemented method, a device, or a computer-readable storage medium has technical character as a whole and is not excluded merely because it involves software. However, only features that contribute to the invention’s technical character can support the inventive-step analysis.2
For AI and machine learning, an abstract model or algorithm is not enough by itself. A contribution may arise when the claim applies the model to solve a technical problem in a technical field or reflects a specific technical implementation. The EPO gives examples such as processing low-level image or audio characteristics, while classifying text solely by its textual content is not, by itself, a technical purpose.3
The following examples are screening illustrations, not legal conclusions:
| Illustrative concept | Initial eligibility question |
|---|---|
| User profiling that selects advertisements to improve click-through rate | Does the claimed contribution go beyond a commercial or cognitive objective and produce a verifiable technical effect? |
| Keyword-based document classification | Is the classification based only on linguistic meaning, or does it serve a specific technical purpose? |
| A neural network that removes sensor noise from captured images | Are the model, signal characteristics, and resulting image-processing effect technically defined and supported? |
| Memory management that reduces processor load or device power consumption | Does the claim identify the implementation and causal link to the internal computer improvement? |
Challenge two: prior art may be hidden outside patent databases
Software ideas often circulate through research, open-source development, standards work, product documentation, or developer communities. Patent applications may become public later than some of those materials, and authors may use vocabulary that differs from patent claims. A search limited to patent keywords can therefore miss relevant technical teaching.
| Source category | Why it may matter | Evidence to preserve |
|---|---|---|
| Patent literature | Claims, embodiments, families, classifications, citations, and prosecution context | Publication number, family member, priority data, relevant passages, and figures |
| Papers and preprints | Early algorithms, benchmarks, architectures, experimental methods, and field terminology | Version, identifier, authors, dated record, PDF, and relevant technical content |
| Public code repositories | Implementation details, commit history, documentation, releases, and issue discussions | Repository URL, access status, commit or release, files reviewed, timestamps, and archive evidence |
| Standards and product materials | Protocol behavior, interfaces, implementations, manuals, and technical specifications | Document version, publication or access date, issuing body, and exact passages |
For internet disclosures, the EPO treats public posting as potentially relevant but separates the reliability of the stated date from whether the particular content was publicly available on that date.5 The evidence is assessed as a whole. Accordingly, neither an arXiv version label nor a GitHub timestamp should be treated as an automatic legal conclusion without checking the content, access, history, and applicable law.6
Search strategy for software and AI inventions
Step one: expand the source set
Begin with the invention’s technical features, inputs, processing steps, outputs, constraints, and claimed effects. Then select sources that fit the field. The table is a planning aid, not a ranking or guarantee of completeness.
| Source | What it can reveal | Search or evidence note |
|---|---|---|
| Espacenet and other patent-office databases | Patent publications, classifications, families, citations, and legal bibliographic data | Combine keywords, classifications, applicants, inventors, citations, and family review |
| arXiv and Google Scholar | Preprints, papers, theses, and terminology used by researchers | Record versions and verify the content and public date relevant to the analysis |
| GitHub | Code, readme files, releases, issues, and implementation history | Preserve the exact files and history; do not infer public availability from a timestamp alone |
| IEEE Xplore and ACM Digital Library | Peer-reviewed computing, electronics, and engineering literature | Search technical synonyms, authors, affiliations, references, and citing papers |
| Papers with Code | Connections among papers, tasks, datasets, benchmarks, and implementations | Verify important facts against the underlying paper and repository |
| Technical blogs, manuals, and standards repositories | Product behavior, interfaces, deployment details, and industry terminology | Check authorship, version, access conditions, publication history, and archived copies |
Step two: use semantic search to supplement keywords
Software vocabulary changes quickly. The same function may be described through a mathematical term, an engineering outcome, an industry acronym, or an implementation detail. Build concept groups for each essential feature, then test combinations that express the technical relationship rather than only the product label.
Semantic search can surface conceptually related documents that do not share obvious keywords. It should supplement, not replace, classification searching, citation tracing, assignee and inventor review, family analysis, and manual examination of the disclosure. Record each search path so another reviewer can reproduce the reasoning.
Step three: verify the online timeline
For a preprint, preserve the identifier, version history, PDF, displayed submission time, and any corroborating records. For public code, preserve the repository URL, relevant files, commit or release identifiers, branch history, access status, and an archived capture when available. Note the time zone if it could affect the relevant date.
The EPO recommends recording the URL, the relevant publication date, the date on which the page was retrieved, and how the date was obtained.7 This discipline is useful even when another jurisdiction’s law governs, because it separates the technical content from the evidence supporting when and how the public could access it.
Subject matter eligibility analysis framework
Step one: abstract the core contribution
Describe the invention without marketing language. Identify the technical system, the data source, the processing steps, the output, and the asserted effect. Then ask whether the contribution is directed mainly to a business rule, presentation of information, mathematical method, or mental activity, or instead to a technical process or technical improvement.
This is an issue-spotting exercise, not a binary result. A claim can contain technical components yet still fail another eligibility step. Conversely, an invention that uses an algorithm may qualify when the algorithm forms part of a supported technical solution.
Step two: test the asserted technical effect
- Does the invention process physical measurements, images, signals, network traffic, or machine-state data for a technical purpose?
- Does it improve the operation of a computer, device, communications system, manufacturing process, or other technical system?
- Does the disclosure explain how the claimed features cause the asserted effect, rather than merely stating a desired result?
A “yes” answer is useful evidence, but it is not dispositive. The claim, specification, comparative evidence, and applicable legal test must support the asserted contribution. Under the EPO’s AI guidance, a technical effect may be established through explanation, mathematical proof, or experimental data where appropriate; a bare assertion is insufficient.3
Step three: review claim-level expression and disclosure support
Map each claim feature to the technical contribution and to supporting disclosure. Avoid relying on generic computing language that does not explain the implementation. For AI inventions, document the model components, data relationships, training or inference steps, constraints, and evaluation evidence that matter to the claimed effect.
Eureka IP Drafting’s Invention Disclosure workflow can turn rough technical notes into a structured disclosure and highlight missing details or overlooked inventive points.8 This can help teams prepare for a claim and search discussion. It does not guarantee patentability, determine claim scope, or replace drafting and review by qualified professionals.
Differences across jurisdictions
There is no reliable one-dimensional ranking of which jurisdiction is “strictest” for every software or AI invention. The answer depends on the claim, specification, technical field, procedural posture, and current guidance. The safer comparison is to identify the official framework and adapt the search report accordingly.
| Jurisdiction | Current official framework | Search and review implication |
|---|---|---|
| European Patent Office | EPC Article 52 and the computer-implemented-invention guidance distinguish excluded subject matter as such from claims with technical character; inventive step focuses on features contributing to technical character.2 | Separate eligibility screening from novelty, then organize closest-art comparisons around the technical contribution and objective technical problem. |
| United States | USPTO examiners apply the Alice/Mayo framework through MPEP Section 2106, including whether a claim recites a judicial exception and whether it integrates that exception into a practical application or adds significantly more.9 | Search both the claimed technical implementation and the alleged technological improvement; preserve support for how the claim operates as a whole. |
| Japan | The JPO applies its current Examination Guidelines and Handbook, including computer-software-related material and published AI-related case examples.10 | Use the current examples to frame the review, but assess eligibility, enablement, novelty, and inventive step on the actual claim and disclosure. |
| China | CNIPA evaluates algorithmic or business-rule features together with technical features in the claim as a whole. Current AI-related amendments to the Patent Examination Guidelines took effect on January 1, 2026 and include added disclosure guidance for model and application details.11 | Capture the technical means, technical problem, technical effect, feature interactions, and the AI implementation details needed to support the contribution. |
A multi-jurisdiction report should therefore avoid converting legal frameworks into star scores. Instead, it should state the searched feature set, evidence limitations, eligibility issues for professional review, and any jurisdiction-specific terminology or disclosure gap.
Illustrative software and AI search cases
Illustrative case one: sensor-based fault prediction
Assume an industrial monitoring system uses a transformer-based model to predict equipment faults from multivariate sensor streams. A useful search starts with the physical context, sensor types, preprocessing, temporal architecture, fault output, deployment constraints, and claimed effect. Patent classifications and citations can be combined with papers on time-series transformers, condition monitoring, and edge inference.
The eligibility review should not stop at the word “transformer.” It should test whether the claim defines a technical application, such as controlling or monitoring equipment, and whether the specification supports the causal relationship between model features and the asserted technical result. The novelty conclusion then depends on what each dated reference actually discloses, feature by feature.
Illustrative case two: reinforcement-learning pricing
Assume a platform uses reinforcement learning to set prices dynamically. Searching only “AI pricing” will likely produce noise. The reviewer should decompose the concept into state representation, reward function, action constraints, feedback, update method, system architecture, and any claimed technical resource or latency improvement.
If the contribution is limited to optimizing a commercial result, eligibility risk may remain. If the invention also claims a supported improvement to computing, communications, or resource allocation, that implementation should be searched and evaluated on its own facts. Reframing a business objective in technical vocabulary is not enough; the disclosure and claim must establish the technical relationship.
Software and AI patentability search checklist
- Separate patent-eligibility questions from novelty and inventive-step questions.
- Translate the invention into essential technical features, relationships, constraints, and asserted effects.
- Give non-patent literature a search priority appropriate to the technology and disclosure practices of the field.
- Search patents, papers, preprints, code, standards, manuals, and other relevant public sources.
- Use keywords, classifications, semantic methods, citations, families, inventors, authors, and organizations as complementary paths.
- For online material, preserve the exact content, version, access status, URL, retrieval date, and evidence supporting the publication date.
- Do not treat a preprint label, commit timestamp, repository name, or search snippet as a complete public-availability analysis.
- Review the current official framework for every target jurisdiction rather than relying on comparative strictness labels.
- Document claim and disclosure review options that remain consistent with the supported technical contribution.
- State search limitations and route legal conclusions to qualified professionals.
Sources and verification
- European Patent Convention, Article 52 — Patentable inventions. Verified July 2026.
- EPO Guidelines 2026, G-II 3.6 — Programs for computers. Verified July 2026.
- EPO Guidelines 2026, G-II 3.3.1 — Artificial intelligence and machine learning. Verified July 2026.
- Patsnap Eureka IP Search. Verified July 2026.
- EPO Guidelines 2026, G-IV 7.5 — Internet disclosures. Verified July 2026.
- EPO Guidelines 2026, G-IV 7.5.2 — Standard and burden of proof. Verified July 2026.
- EPO Guidelines 2026, G-IV 7.5.6 — Technical details and general remarks. Verified July 2026.
- Patsnap Eureka IP Drafting. Verified July 2026.
- USPTO MPEP, Section 2106 — Patent Subject Matter Eligibility. Current guidance verified July 2026.
- Japan Patent Office, AI-related technologies: examination case examples. Updated May 2026; verified July 2026.
- CNIPA, Interpretation of the 2025 Patent Examination Guidelines amendments. Amendments effective January 1, 2026; verified July 2026.
This article provides general information, not legal advice. Patent eligibility, prior-art rules, grace periods, evidence standards, and examination practice differ by jurisdiction and matter. Qualified professionals should review the applicable claims, facts, and current law.
Build a reviewable software and AI search record
Structure the technical features, test multiple search paths, and compare source-linked evidence before qualified professionals reach a patentability conclusion.
Explore Eureka IP Search