AI & Patent Law
Can You Patent an AI Invention?
Yes — with caveats. AI-related inventions are patentable. The main hurdle is § 101 (Alice), which bars abstract ideas on computers. Here is what crosses the line, what doesn't, who can be the inventor, and how to draft claims that hold up.
Educational guide. For your specific AI invention, consult a registered patent attorney.
The fundamental rule: specific technical improvement, not abstract outcome
The Alice framework (Alice Corp. v. CLS Bank, 2014) asks two questions: (1) does the claim recite an abstract idea? (2) if yes, does it add something 'significantly more' — a specific technical improvement, a particular inventive concept? Most AI claims fail at step 1 (they recite an abstract mathematical process) and succeed or fail at step 2 based on how specifically they describe the implementation.
A claim that says 'using a neural network to classify images' is abstract. A claim that says 'a method of classifying images using a convolutional neural network with a specific architecture, trained on a specific dataset type, achieving a specific metric improvement, using a specific loss function modification' may survive Alice — because it describes a specific technical solution to a specific technical problem.
The key insight: patent the HOW, not the WHAT. Describe specifically how your AI system achieves its result, not just what result it achieves.
What IS patentable in AI
Novel neural network architectures
A specific arrangement of layers, connections, attention mechanisms, or activation functions that produces measurable improvements over the prior art. The Transformer architecture (attention is all you need) is the canonical example of something patent-eligible — a specific novel architectural approach with demonstrated advantages.
Specific training methodologies
Novel approaches to training AI models: a new loss function formulation, a curriculum learning approach, a data augmentation technique, a regularization method. Must be described specifically enough that skilled practitioners can reproduce it.
AI applied to specific technical problems
An AI system that solves a concrete technical problem in a novel way: AI-driven chip layout optimization, AI-based protein structure prediction (with a specific claimed method), AI-based network routing optimization with a specific algorithm. The more specific and technical the problem, the more defensible the patent.
Compression and efficiency innovations
Model compression techniques (specific pruning methods, quantization approaches, knowledge distillation architectures), inference optimization systems, hardware-software co-designs for AI. These are 'technical improvements to computing' and survive Alice more readily.
AI training systems and infrastructure
Systems and methods for distributed training, gradient communication protocols, hardware-aware model parallelism, memory-efficient training. These are concrete technical systems with measurable improvements.
What is NOT patentable in AI
Abstract mathematical methods: 'using a mathematical model to predict X' without specifying the concrete technical implementation is not patentable. The abstract idea of using statistics or probability to make predictions is prior art.
Business applications of generic AI: 'using AI to identify fraud' or 'using machine learning to recommend products' without specifying a technical implementation are not patentable. These apply abstract ideas to business problems using generic AI components.
Desired outcomes without specific methods: 'a system that generates realistic images' without specifying the architecture, training procedure, or technical approach is not patentable. You need to claim the specific way you achieve the outcome.
Pure data manipulation without physical transformation or technical improvement: sorting, filtering, or classifying data using standard ML techniques applied to standard data types without showing a specific technical improvement is problematic under Alice.
Who can be named as inventor: humans only
Under current US law, only natural persons (human beings) can be named as inventors. The USPTO and federal courts have been unambiguous: AI systems cannot be inventors.
The DABUS cases settled this definitively. Stephen Thaler filed patent applications naming DABUS (an AI system he developed) as the inventor. The USPTO rejected the applications. The Federal Circuit affirmed in Thaler v. Vidal (2022): the Patent Act's term 'individual' in the inventorship provisions means a natural person. The Supreme Court declined to hear the case.
The practical question for AI-assisted inventions: if an AI contributed to the conception of an invention, who is the human inventor? USPTO guidance (February 2024) states that the named inventor must be a natural person who made a significant contribution to the conception of at least one claim. Using an AI as a tool does not make you an inventor; using AI-generated ideas that you then significantly modified, refined, or reduced to practice may make you an inventor.
Best practice: document your team's contributions carefully. Which humans conceived which aspects of the claimed invention? Which aspects came from AI-generated suggestions that humans substantially modified? This documentation may matter if inventorship is ever challenged.
USPTO AI guidance (2024)
In February 2024, the USPTO issued updated guidance on AI-assisted inventions clarifying the inventorship standard for the modern era. Key points: (1) AI cannot be an inventor; (2) a human inventor must make a significant contribution to conceiving at least one claim; (3) merely recognizing a problem and prompting an AI to solve it is not sufficient for inventorship; (4) a human who significantly contributes to reducing an AI output to a specific claimed invention can be an inventor.
The 2024 guidance also addressed patentable subject matter for AI inventions, confirming the Alice framework applies and providing examples of AI claims that do and do not clear the § 101 hurdle. The guidance made clear that the novelty and non-obviousness requirements also require careful attention — LLM-generated prior art searches may be incomplete.
Additional USPTO guidance on AI use in patent practice (2024) addressed disclosure obligations when AI is used to draft patent applications: while there is no duty to disclose AI use to the USPTO, inventors and practitioners retain their obligations of candor and good faith.
Drafting AI patent claims that hold up
Structure claims around the technical problem. Start the specification by identifying a specific technical problem: 'conventional image classifiers require [X] memory and achieve [Y] accuracy on [specific dataset type]; the present invention provides a method achieving [Z] accuracy with [W] memory.' This framing helps the Alice step 2 analysis.
Claim the specific architecture, not the general approach. Instead of 'a neural network for image classification' claim 'a convolutional neural network comprising: a first convolutional layer with [specific parameters]; an attention mechanism configured to [specific operation]; a training process comprising [specific steps].' The specificity is protection.
Include performance claims in the specification. Measurable improvements over prior art help demonstrate that the invention adds 'something more' than abstract idea implementation. 'Achieves 15% reduction in inference latency while maintaining 99.2% accuracy on the ImageNet benchmark' is concrete evidence of technical improvement.
File early for competitive AI inventions. AI is moving fast. An improvement that's novel today may be anticipated by a competitor's filing or a public research paper within months. Track One prioritized examination is especially valuable for AI patents.