Introduction

If your invention involves software algorithms, artificial intelligence, or business methods, software and AI patentability search requires significant adaptation of traditional patentability search methodologies. Inventions in this category face a dual challenge:

1. Subject Matter Eligibility Risk — Your invention may fail at the very first hurdle of “whether it qualifies as patentable subject matter,” never even reaching the discussion of Novelty and Inventive Step.
2. The “Hidden” Nature of Prior Art — In the software and AI domain, a vast amount of prior art exists outside patent databases: arXiv papers, open-source code on GitHub, technical blogs, API documentation…

Software and AI patentability search is not a “simplified version” of traditional search — it is a distinct branch with its own unique challenges and methods. This article systematically outlines the key considerations.

The Dual Challenge: Subject Matter Eligibility + The Traditional Triad

Challenge One: Patentable Subject Matter (Subject Matter Eligibility)

Under Article 52(2)(c) and (3) of the European Patent Convention (EPC Art. 52(2)(c) and (3)), schemes, rules and methods for performing mental acts, playing games or doing business, and programs for computers “as such” are not regarded as inventions. In examination practice, pure algorithms, pure business methods, and pure mathematical methods stripped of concrete technical implementation are frequently classified under this exclusion.

However, EPC Art. 52(3) also makes clear: the above exclusion applies only to the extent that the application or patent “as such” relates to those subject matters or activities. According to EPO case law, if a claim contains at least one technical feature and, taken as a whole, possesses a “Technical Character,” it does not fall within the excluded subject matter.

The Gold Standard for Assessment:

Does the solution employ technical means to solve a technical problem and achieve a technical effect?

Software inventions likely to face subject matter rejection:

• “A user profiling method for improving advertisement click-through rates” → Concerns improvement of a business method, not a technical problem
• “A document classification method based on keyword matching” → The classification logic itself is a mental act or rule

Software inventions likely to pass subject matter examination:

• “A neural-network-based image denoising method, comprising processing raw data acquired by a sensor to…” → Involves specific technical processing of sensor data
• “A memory management method for reducing power consumption of a mobile device, which reduces IO operations by dynamically adjusting caching strategy” → Solves the technical problem of power consumption

Challenge Two: The “Hidden” Nature of Prior Art

The prior art landscape in software and AI differs fundamentally from that of traditional manufacturing/mechanical fields:

Search Strategy for Software and AI Patentability Search

Step One: Expand Search Sources (Do Not Search Patent Databases Alone)

Mandatory Sources:

Step Two: Semantic Search as a Supplement

Technical solutions in the software/AI domain are often difficult to capture comprehensively with a few precise keywords. Leveraging the semantic search capabilities of Google Patents, its “similar patents” recommendation feature, or AI-powered semantic search in commercial databases such as PatSnap Analytics can help uncover prior art that keyword-based search struggles to reach.

Step Three: Pay Attention to the arXiv Timeline

The upload date of an arXiv paper sometimes precedes the formal publication date of the corresponding conference paper. During patentability search:

• Take note of the upload date of an arXiv paper’s v1 (initial version)
• Treat the arXiv upload date as the effective public disclosure date of the Prior Art Document (rather than the later journal publication date)

Subject Matter Eligibility Analysis Framework for Software and AI Patentability Search

In a patentability search report, it is advisable to include a dedicated section on “Subject Matter Patentability Analysis,” rather than assuming that the invention will certainly pass subject matter examination. The analytical framework is as follows:

Analysis Steps

Step 1: Abstract the Core Innovation
Strip the invention down to its most essential “what-it-is”:

• If the core innovation is “a decision rule / classification method / recommendation logic” (pure algorithm) → High subject matter risk
• If the core innovation is “a specific approach to processing data from particular hardware/sensors, solving a concrete technical problem” → Low subject matter risk

Step 2: Confirm the Technical Nature of the Problem
Ask three questions:

1. Does the invention involve the acquisition of data from the physical world? (sensors, images, etc.)
2. Does the invention alter control over the physical world? (machinery, displays, etc.)
3. Does the invention improve the internal operation of a technical system? (reducing power consumption, increasing processing speed, reducing storage footprint, etc.)

At least one answer is “yes” → Subject matter risk is manageable.

Step 3: Examine the Expression at the Claim Level
The same technical solution, described in different language, can yield entirely different outcomes in subject matter examination. It is advisable to provide, within the patentability search report, strategies for circumventing subject matter objections at the claim drafting level.

Differences Across Jurisdictions (Software/AI Special Edition)

Building upon the “Global Patentability Search” article in the previous chapter, the following table focuses on the particularities of software and AI:

Real-World Cases: Pleasant Surprises and Unwelcome Shocks in AI Patentability Search

Case One: A Pleasant Surprise

A company developed a “Transformer-based method for predicting faults in industrial equipment,” which analyzes time-series data from equipment vibration sensors to provide early fault warnings.

Search Path:

• Subject matter analysis: Involves specific processing of physical sensor data, solving the technical problem of industrial safety → Low subject matter risk
• Patent search: Discovered numerous patents on fault prediction based on traditional ML, but none employing the Transformer architecture
• NPL search: Found several papers on arXiv about using Transformers for time-series analysis, but all targeting NLP or general-purpose time-series forecasting
• GitHub search: Found open-source implementations of Transformer time-series models, but no specific application to industrial fault prediction was identified

Conclusion: Novelty is established. The key question for Inventive Step is — would applying Transformers to fault prediction be regarded as an “obvious application” by a Person Skilled in the Art? Given that the data characteristics of industrial fault prediction (high noise, non-stationarity) differ significantly from general-purpose time-series forecasting, and the invention demonstrated markedly improved results on real-world data, Inventive Step could be established.

Actual Outcome: The patent was ultimately granted.

Case Two: An Unwelcome Shock

A startup developed a “dynamic pricing algorithm based on reinforcement learning” for e-commerce platforms, designed to automatically adjust product prices based on real-time supply and demand.

Search Path:

• Subject matter analysis: The core innovation is the “pricing strategy” itself — this is a business method, not a technical solution. Unless the algorithm involves specific technical improvements (such as faster model inference or optimization of computational resources), the subject matter risk is extremely high.
• Patent search: Found a large number of similar “dynamic pricing” patents, mostly in the United States (which is relatively more friendly to this type of invention)
• NPL search: In arXiv and academic journals, dynamic pricing using reinforcement learning turned out to be a hot topic at the intersection of economics and computer science — a substantial body of literature had existed for at least five years

Conclusion: Not only is the subject matter risk high, but Inventive Step is also doubtful.

Actual Outcome: The client accepted the judgment of the patentability search report and adjusted the patent application strategy to focus more on “how the system optimizes the allocation of computational resources on the recommendation server, reducing pricing update latency to within 100 ms” — framing the application around this technical improvement angle in order to carve out room for both subject matter eligibility and Inventive Step.

Software and AI Patentability Search Checklist

• [ ] Has NPL search been given priority equal to or higher than patent search?
• [ ] Have non-patent sources such as arXiv, Google Scholar, and GitHub been covered?
• [ ] Has a dedicated subject matter patentability risk analysis been conducted?
• [ ] Have the differing attitudes toward software/AI patents across jurisdictions such as CN, US, EP, and JP been examined?
• [ ] If subject matter risk is high, have “circumvention rewrite” recommendations been provided in the report?
• [ ] Has attention been paid to arguing the “technical nature” of the technical solution?

Key Takeaway: Software and AI patentability search faces the dual challenge of subject matter examination and the “hidden” nature of prior art. The search must cover NPL sources beyond patent databases (arXiv, GitHub, and conference papers are of paramount importance). The search report should include a “Subject Matter Patentability Analysis,” assessing whether the invention employs technical means to solve a technical problem and produces a technical effect. Subject matter standards vary across jurisdictions, requiring differentiated strategies when filing in multiple territories. PatSnap Analytics can support broader patent and technology evidence review for software and AI inventions.