Machine Learning Patents

Yes — machine learning models, training methods, inference systems, and AI-powered applications can be patented in the United States, Europe, China, Japan, and most other major patent jurisdictions. But they face specific eligibility challenges that other types of inventions do not. In the United States, the central obstacle is 35 U.S.C. § 101 patent-eligibility under the Alice/Mayo framework: courts and the USPTO examine whether the claims are directed to an abstract idea (which includes mathematical concepts, mental processes, and certain methods of organizing human activity) and, if so, whether the claims include an 'inventive concept' that goes beyond applying the abstract idea using a generic computer. A claim to 'train a neural network to classify inputs' is essentially a claim to a mathematical procedure — an abstract idea — and is not patentable as stated. But a claim to 'a system for detecting anomalous sensor readings in a wind turbine, comprising a recurrent neural network trained on 10 Hz vibration data, that outputs a maintenance alert when a learned threshold is exceeded' is far more concrete — it is directed to a specific technical problem (predictive maintenance), with specific technical means (recurrent neural network, specific input data, specific output), in a specific application domain (turbines). The line between 'abstract ML math' and 'patentable ML application' runs through the specificity and technicality of the claim. The USPTO's 2019 Revised Guidance on Subject Matter Eligibility provides a structured framework, and AI/ML applications have had improving (though still challenging) success rates since then.

Under the USPTO's 2019 Revised Guidance (implementing Alice/Mayo), a claim survives § 101 if: (Step 1) It is in a statutory category — process, machine, manufacture, or composition of matter — which ML system claims easily satisfy. (Step 2A, Prong 1) If it recites a 'judicial exception' (abstract idea, natural phenomenon, law of nature). Most ML claims recite some mathematical concept — neural networks are math, loss functions are math, optimization is math — so they often reach Prong 1 analysis. (Step 2A, Prong 2) Whether the claim 'integrates the judicial exception into a practical application' — this is where the key analysis happens. Indicators of integration into a practical application include: the claim applies the mathematical concept to 'effect a particular treatment or prophylaxis for a disease or medical condition' (medical diagnostics is a strong area); reflects an 'improvement in the functioning of a computer or other technology' (a more efficient neural network training algorithm that reduces computation); uses the mathematical concept 'with a particular machine' (a neural network combined with specific sensors, hardware, or actuators); or 'transforms or reduces a particular article to a different state or thing.' (Step 2B, fallback) If Step 2A Prong 2 fails, whether there is an inventive concept beyond well-understood, routine, conventional activity. Practically: the claims most likely to survive describe (1) a specific technical improvement (faster inference, lower memory, better accuracy on a defined task), (2) in a specific application domain (medical imaging, industrial control, network security), (3) with specific technical means (model architecture details, training data characteristics, hardware integration), and (4) producing a specific, concrete technical result. Boilerplate 'apply it to a computer' drafting fails; genuine technical specificity survives.

This is an evolving and unsettled area. A trained neural network model — the artifact consisting of a specific architecture with learned weight values — is arguably a 'machine' or 'manufacture' under § 101: it is a physical configuration of data stored in memory that causes a computer to perform specific functions. Several lines of argument support its patentability: (1) A trained model with specific weights is a machine in a particular configuration — analogous to a circuit with specific component values, not a generic computer. (2) The model encodes a specific, non-obvious learned function (the weights are not derivable by routine math — they emerge from training on specific data). (3) Case law on software patents (Enfish, McRO, Berkheimer) supports eligibility when the claims are directed to a specific, concrete technical improvement rather than an abstract result. However: (4) Competitors can potentially design around or reverse-engineer a claimed model architecture; weights are often proprietary as trade secrets, not patents. (5) The § 101 analysis for claims reciting 'a model comprising weights' may still hit abstract-idea objections if framed too broadly. Best practice: claim the trained model as part of a system — 'a memory storing weights of a neural network trained to [specific task] and a processor configured to execute the network' — rather than as a standalone data artifact. Include method claims directed to the training process with specific training data characteristics, and system claims anchored to specific hardware. This multi-claim strategy protects different aspects of the ML invention and increases the likelihood that at least some claims survive examination.

The patents-versus-trade-secrets question is more acute for AI/ML than for most technologies, because the core of an ML model — the trained weights — is typically undetectable in the output. A model's weights cannot be reverse-engineered from its predictions in most practical settings, making trade secret protection naturally strong for the model itself. The strategic calculus: Patent when (1) the inventive method is visible in the output or in the deployed system — i.e., a competitor using the same approach can be detected, which is necessary to enforce a patent. (2) You want the legal right to exclude competitors who independently develop the same approach, not just protect your specific implementation. (3) The invention is in the training data pipeline, hardware integration, or application domain in ways that are observable and enforceable. (4) You need 'patent pending' status for licensing leverage, investor valuation, or competitive signaling. Protect as trade secret when (1) the model weights themselves are the invention and they are not inferable from output. (2) The training data, data preprocessing steps, and feature engineering are novel but not observable externally. (3) The competitive advantage derives from the specific trained model, not from the architectural choices. (4) The marginal cost of independent invention by a competitor is high — if it took your team 3 years of specialized work to develop, trade secret protection plus a head start may be more valuable than a patent that discloses your approach. Many AI companies do both: patent the system architecture, application methods, and training pipelines (observable and enforceable), while protecting trained model weights, proprietary datasets, and data preprocessing as trade secrets.

As of 2024, the USPTO requires that every patent application name a human inventor — an AI system cannot be listed as an inventor, and a patent naming only an AI as inventor would be invalid. This rule was established through a series of USPTO decisions rejecting applications filed by Stephen Thaler naming his DABUS AI system as inventor (2020–2022), affirmed by the Federal Circuit in Thaler v. Vidal (Fed. Cir. 2022), and confirmed by the Supreme Court declining to review the case in 2023. The legal status of AI-generated inventions where no human contributed to the conception is therefore: currently unpatentable in the U.S. under existing statute (35 U.S.C. § 100(f): 'inventor means the individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention'). The USPTO issued guidance in February 2024 (following Executive Order 14110 on AI) clarifying that AI-assisted inventions — where humans use AI as a tool but humans contribute to the conception — can be patented, with the human contributors listed as inventors. The inventorship determination focuses on who contributed to the conception of the claimed subject matter; using AI to generate, screen, or refine candidates does not automatically make the AI an inventor if humans directed the process and evaluated the results. Companies using AI tools in R&D should document human decision-making in the inventive process — which AI outputs were selected, how they were modified, which problems the humans defined — to support inventorship claims. This is an evolving area: the USPTO has launched rulemaking on AI and inventorship, and Congressional action is possible.