AI and Machine Learning Patents

Patent eligibility for AI/ML inventions is governed by the Alice/Mayo two-step framework, but the USPTO's 2019 Revised Guidance has clarified how this applies to ML inventions: THE ALICE PROBLEM FOR AI/ML: Alice Corp. v. CLS Bank (S.Ct. 2014) established a two-step test: Step 1: is the claim directed to a patent-ineligible concept (abstract idea; law of nature; natural phenomenon)? Step 2: does the claim add significantly more — an inventive concept beyond the abstract idea?; MACHINE LEARNING AND ABSTRACT IDEAS: ML algorithms are mathematical concepts (mathematical relationships; mathematical formula; mathematical calculation) that qualify as abstract ideas; a neural network is a mathematical construct; a statistical model is a mathematical relationship; simply applying math to a domain — even medical imaging, natural language processing, or drug discovery — may not be enough; THE 2019 REVISED GUIDANCE — THE KEY FRAMEWORK: the USPTO restructured the Alice analysis in 2019: PRONG 1 (Prong 2A, Prong One): is the claim directed to a judicial exception? For AI/ML: identify the mathematical concept or abstract idea in the claim; PRONG 2 (Prong 2A, Prong Two): if yes, does the additional claim language integrate the exception into a practical application? INTEGRATION INTO PRACTICAL APPLICATION: the claim is eligible if it uses the mathematical concept as part of a process that goes beyond the exception itself; examples that pass: a specific ML system that improves computer performance by reducing memory access (technical improvement to the computer); an AI model trained on specific sensor data to classify specific physical defects in a manufacturing process (specific real-world application with defined technical inputs/outputs); a neural network architecture specifically designed for processing sparse medical imaging data with reduced computational requirements (technical improvement); examples that fail: a claim to 'training a neural network to predict disease outcomes' without specifying technical implementation details; a generic ML model configured to improve user recommendations (just applying math to a commercial objective); DOMAIN-SPECIFIC ELIGIBILITY PATTERNS: MEDICAL IMAGING AI: typically eligible if claim specifies technical image processing steps; specific segmentation or detection task; improvement in accuracy or speed over prior art methods; NATURAL LANGUAGE PROCESSING: eligible if claim specifies specific syntactic or semantic processing steps; hardware-accelerated transformer architecture; specific technical application; DRUG DISCOVERY AI: harder if claim merely correlates molecular features with activity (Mayo analysis may apply); eligible if claim involves specific computational chemistry steps; specific structural analysis method.

Effective AI/ML patent claim drafting requires careful attention to the structure and specific technical details that differentiate eligible from ineligible claims: SYSTEM CLAIMS (often the strongest form for AI/ML): a system claim that includes specific hardware components, a specifically trained or configured ML model, and specific technical functionality is typically more defensible than pure method claims; EFFECTIVE SYSTEM CLAIM ELEMENTS: one or more processors (can be generic if the ML improvement is software-defined); specific ML model architecture (attention mechanism; convolutional layer configuration; specific node structure); specific type of training data or training process (domain-specific training data; specific preprocessing steps); specific technical output beyond abstract predictions (control signal; modified data structure; hardware command); EXAMPLE ELIGIBLE SYSTEM CLAIM STRUCTURE: 'A system comprising: one or more processors; and non-transitory computer-readable memory storing instructions that, when executed, cause the processors to: receive [specific technical input]; process the input using a [specific ML model architecture] trained on [specific technical domain] data; and generate [specific technical output that causes a technical effect]'; METHOD CLAIMS: method claims should include specific algorithmic steps beyond the high-level description; ELIGIBLE METHOD CLAIM ELEMENTS: specific preprocessing steps (not generic); specific model architecture details; specific loss function or optimization approach if novel; specific technical post-processing steps; specific hardware interaction; CRM CLAIMS: non-transitory computer-readable medium claims are system-equivalent and avoid method claim infringement proof issues (no divided infringement under Akamai if all steps are stored on a single medium); INVENTORSHIP OF AI-ASSISTED INVENTIONS (CRITICAL): under Thaler v. Vidal (Fed. Cir. 2022), AI cannot be named as an inventor on a US patent; a human must make a significant contribution to the conception of each claimed invention; USPTO's February 2024 inventorship guidance: if AI generates a specific design and a human merely asks AI to solve a problem, the human has NOT made a sufficient inventive contribution to be an inventor; human inventors must identify a specific embodiment; understand how to apply it; appreciate how it solves a problem; this guidance has significant implications for AI-assisted drug discovery and materials science where AI models generate specific candidates; SPECIFICATION REQUIREMENTS: detailed description of the neural network architecture; specific training process description; mathematical description of the loss function; code-level pseudocode or algorithmic description; experimental data showing technical improvement; comparison to prior art methods on objective technical metrics.

The AI patent landscape is highly concentrated among major technology companies and a growing number of AI-specialized companies, with significant activity across multiple technology domains: MAJOR AI PATENT HOLDERS (by volume): IBM: historically the largest holder of US AI-related patents; extensive patent portfolio across NLP, computer vision, knowledge representation; IBM regularly monetizes through licensing and assertion; Google/Alphabet: Google Brain and DeepMind patents; transformer architecture; attention mechanism; AlphaFold; neural architecture search (NAS); extensive claims on search ranking, recommendation systems, language models; Microsoft: natural language processing; GitHub Copilot-related claims; Azure AI services; LinkedIn AI; OpenAI partnership-related; Microsoft Research contributions; Qualcomm: edge AI inference; neural processing unit (NPU) architecture; efficient ML for mobile; MediaTek and ARM: neural network acceleration hardware; NPU design patents; Baidu/Alibaba/Tencent: large Chinese AI portfolios; strong in NLP; autonomous driving; recommendation systems; IMPORTANT FOUNDATIONAL PATENTS: TRANSFORMER ARCHITECTURE: Google's 'Attention Is All You Need' paper (Vaswani et al., 2017) describes the transformer; several patents on attention mechanism variants; NOTE: the transformer concept is widely implemented, and many foundational architecture claims have been filed by multiple parties; DEEP LEARNING FOUNDATIONS: LeCun's CNN patents (assigned to AT&T; later to Meta); Hinton's deep belief network patents; these foundational patents have largely expired or are near expiration; GENERATIVE AI: multiple competing patent families on GAN (Generative Adversarial Network) architecture; diffusion model patents (Sohl-Dickstein et al. → Ho et al. → Song et al.); large language model training method patents (Microsoft/OpenAI collaboration); DEFENSIVE PATENT AGGREGATORS: the Open Invention Network (OIN) has a royalty-free defensive pool that includes some AI patents; LOT Network: companies commit to royalty-free licenses to pool members if patents are sold to non-practicing entities (NPEs); PATENT ASSERTION IN AI: NPEs are acquiring AI patents; AI/ML companies should consider building defensive portfolios and joining defensive aggregators; the risk of AI patent assertion is growing as the technology becomes more commercially valuable and foundational patents become more accessible to NPEs.

The emerging questions around AI inventorship, AI-assisted invention, and who owns AI-generated inventions are reshaping patent strategy for companies using AI in their R&D processes: THE AI INVENTORSHIP QUESTION — THALER v. VIDAL: Stephen Thaler attempted to name an AI system he calls DABUS as the inventor on two patent applications; the Federal Circuit held in Thaler v. Vidal (Fed. Cir. 2022) that US patent law requires inventors to be natural persons; the Patent Act refers to 'individuals,' which the court interpreted as natural persons; DABUS and any AI system cannot be named as a US patent inventor; this ruling is consistent with similar decisions in the UK (Thaler v. Comptroller General, UK Supreme Court 2023), the EU, and most major patent jurisdictions; WHAT IS NOT SETTLED — AI-ASSISTED INVENTIONS: while pure AI-generated inventions are clearly not patentable without a human inventor, the harder question is: how much AI assistance is permitted before the human contribution is insufficient for inventorship?; THE USPTO 2024 GUIDANCE: the USPTO issued inventorship guidance for AI-assisted inventions in February 2024: a claim must have at least one human inventor; for each claim, a human inventor must have made a significant contribution to the conception of the SUBJECT MATTER of THAT claim; merely providing the AI with a problem to solve and accepting its output is NOT a significant contribution to the conception; WHAT QUALIFIES AS SUFFICIENT HUMAN CONTRIBUTION: designing and training the specific ML model used; selecting the specific training data and understanding why it was relevant; identifying that the AI output represents a specific, non-obvious solution to a technical problem; appreciating the significance of a specific AI-generated candidate and understanding how it achieves the desired result; translating the AI output into a specific implementable embodiment; PRACTICAL IMPLICATIONS FOR AI DRUG DISCOVERY: large pharmaceutical companies use AI to generate drug candidate molecules; the question is whether a human scientist who interprets the AI output, selects specific candidates, and performs confirmatory experiments has made a sufficient inventive contribution; the USPTO guidance suggests: selection alone may not be enough; appreciation of HOW the candidate works and WHY it solves the problem adds to inventive contribution; identifying specific structural features that explain the AI-generated candidate's properties strengthens inventorship claims; OWNERSHIP CONSIDERATIONS: patents on AI-related inventions are typically owned by the employer under work-for-hire doctrine (35 U.S.C. § 262); assignment agreements with employees should specifically address AI-assisted inventions; companies using third-party AI platforms should clarify IP ownership rights in their service agreements.