How to Patent AI Inventions

The Thaler/DABUS cases established definitively that AI systems cannot be named as inventors on patent applications in the US, UK, EPO, Australia, or most major jurisdictions. Stephen Thaler filed patent applications in multiple countries naming 'DABUS' (Device for the Autonomous Bootstrapping of Unified Sentience) as the sole inventor. The result: USPTO rejected the applications (CAFC affirmed 2022); UK IPO rejected (UK Supreme Court affirmed 2023); EPO rejected; Federal Court of Australia (first instance Allsop J. 2021 allowed AI inventor, but Full Federal Court reversed 2022). The rationale: patent statutes require inventorship by a 'natural person' or 'individual' (35 U.S.C. § 100(f): 'the term inventor means the individual or... individuals who invented or discovered the subject matter'). Implication for AI companies: a human being must be named as inventor — the person who conceived the claimed invention. If AI tools assisted in the process, the humans who directed, designed, and applied the AI are the inventors. USPTO issued guidance (February 2024) confirming: AI-assisted inventions are patentable if a human made a significant contribution to the claimed invention; the human inventor need not have worked independently of AI tools.

AI and machine learning patents face a particular risk under Alice Corp. v. CLS Bank Int'l (2014) and the § 101 abstract idea doctrine. The two-step Alice test: Step 1 — is the claim directed to an abstract idea (including mathematical concepts, methods of organizing human activity, mental processes)? Step 2 — does the claim contain an inventive concept that transforms the abstract idea into patent-eligible subject matter? The problem for AI patents: a machine learning algorithm is inherently mathematical; a training process is a mathematical optimization; a neural network inference is a mathematical computation. Examiners frequently reject AI claims under Step 1 as directed to mathematical concepts or mental processes (a human could theoretically perform the computation given enough time and paper). How to survive Alice in AI patent prosecution: (1) Claim the specific technical application: instead of claiming 'a method for classifying images using a neural network,' claim 'a method for detecting early-stage diabetic retinopathy in retinal fundus images by processing at least one layer of a convolutional neural network trained on a labeled dataset of N > 5M fundus images to generate a probabilistic map of microaneurysm locations with sensitivity > 90% at 80% specificity' — the specific technical application (automated medical diagnosis, defined performance thresholds, specific image type) makes the claim directed to a practical application; (2) Claim the specific architecture improvements: if the invention is a new neural network architecture that is computationally more efficient, claim the architectural improvement itself (specific attention mechanism, pruning strategy, quantization method) with structural specificity, not just the result; (3) Claim hardware-software integration: claims that recite specific hardware configurations (custom silicon, memory access patterns, specific processor pipeline modifications for ML workloads) are stronger under Step 2 than pure software claims; (4) Claim training data pipeline innovations: data preprocessing steps with specific technical character — novel data augmentation methods, specific labeling quality metrics, novel synthetic data generation procedures — can survive Alice because they are directed to technical improvements in a data pipeline, not just a mathematical model.

The EPO's approach to computer-implemented inventions (CII) applies equally to AI/ML patents. Under EPO practice (Guidelines for Examination Part G Chapter II Section 3.3, as updated for AI/ML), an AI or ML innovation can be patentable if it has a technical character. The technical character requirement can be met in two ways: (1) The AI/ML system is itself technical: a specific neural network architecture implemented on hardware with defined technical constraints; AI-assisted medical image analysis system with concrete technical performance parameters; AI system for controlling a technical process (autonomous vehicle steering, industrial robot path planning, network packet routing); (2) The AI/ML method has a technical purpose: classification of physical objects for a technical application (e.g., classifying EEG signals for seizure detection, classifying histological images for cancer staging, classifying radar returns for weather prediction) — the technical nature of the input/output data gives the method technical character even if the underlying ML algorithm is mathematical; training a neural network to improve compression of specific types of images (JPEG artifacts reduction) — the compression process is technical. EPO examination practice: the EPO has published CII/AI Guidelines that give numerous examples of AI inventions with and without technical character. Key EPO decisions on AI: T 0702/20 — classification of images for medical diagnosis is technical even if ML algorithm is abstract; T 0161/18 — training a neural network is generally mathematical, but training specifically to improve a technical application can be patentable. Practical difference from US Alice: the EPO's technical character test is generally considered more permissive than US Alice for AI patents. An AI system that fails Alice Step 2 in the US may pass EPO technical character because its application domain (medical, industrial, communications) provides technical character.

The strategic decision between patent and trade secret for AI IP is more nuanced than for most other technologies: (1) Patent the application architecture: specific system designs that integrate AI into particular products — how an autonomous vehicle combines camera inputs, lidar scans, and HD map data using specific AI inference pipelines — are patentable applications that should be patented because competitors can study the deployed product and reverse-engineer the general approach; (2) Trade secret the weights and training data: trained model weights (the billions of parameters that encode an AI model's learned behavior) are not patentable (they are a collection of numbers, not an invention) but ARE protectable as trade secrets. OpenAI's GPT-4 weights are trade secrets. Meta chose to open-source LLaMA weights (forgoing trade secret protection) as a strategic decision to build an ecosystem. Training datasets are also trade secret-protectable: the combination of data sources, data cleaning procedures, labeling methodology, and curation criteria that went into building a training corpus represents significant economic value and can be protected as a trade secret if kept confidential; (3) Patent the training innovations: novel training procedures (new loss functions, new regularization techniques, new curriculum learning strategies, new fine-tuning approaches) that achieve meaningfully better performance are patentable as process inventions, but only if the claim is specific and technical enough to survive Alice; (4) Patent evaluation and safety methods: AI evaluation methodologies (specific benchmark designs, red-teaming procedures, adversarial robustness testing protocols) are patentable where they involve novel technical methods; constitutional AI approaches, RLHF (Reinforcement Learning from Human Feedback) specific pipeline designs; (5) Trade secret the scaling laws and hardware efficiency secrets: empirical findings about how model performance scales with compute, data, and parameters ('scaling laws') are often kept as internal trade secrets rather than published or patented; (6) Copyright the outputs (where applicable): AI-generated content is generally NOT copyrightable (US Copyright Office has consistently held that human authorship is required); however, the prompts, fine-tuning, and human-curated outputs that are substantially the product of human creative choices may retain some copyright protection — this is still evolving.

The world's largest AI patent holder. Google has filed thousands of AI patents at USPTO covering: TensorFlow framework implementation optimizations (though TensorFlow itself is open-sourced, specific implementation improvements are patented); Transformer architecture: the original Transformer architecture ('Attention Is All You Need', Vaswani et al. 2017) was published as a research paper rather than patented — Google chose publication (academic precedent) over patent protection for the foundational attention mechanism; TPU (Tensor Processing Unit) hardware patents: ASIC design for matrix multiplication (systolic array architecture), high-bandwidth memory interface, chip-to-chip interconnect for TPU pods; DeepMind AlphaFold protein structure prediction pipeline patents; Google Search AI ranking algorithm implementations; Waymo autonomous driving perception and planning system patents (hundreds of patents on LiDAR, camera fusion, HD mapping); Google Brain/DeepMind biomedical AI (medical imaging, EHR analysis). Google's strategy: file broad application patents on the most commercially important AI use cases; publish breakthrough research as papers (academic priority without patent); maintain training infrastructure details as trade secrets.

Microsoft holds extensive AI patents across: natural language processing system patents (pre-LLM era conversational AI, post-LLM Copilot integration system patents); GitHub Copilot code generation system patents (specific system architecture claims); Azure AI services infrastructure patents; OpenAI integration patents (Microsoft invested ~$13B in OpenAI; Microsoft holds exclusive commercialization rights to some OpenAI technologies; however, OpenAI's core model weights remain with OpenAI); Xbox AI game content generation; Microsoft Research has published thousands of AI papers without filing patents — a common academic-adjacent trade secret protection approach for foundational research.

OpenAI files relatively few patents compared to its research output. OpenAI's IP strategy is primarily: (1) trade secret protection for GPT-4 and subsequent model weights, training data composition, and fine-tuning procedures; (2) publication of research papers (establishing academic precedent, not patent priority); (3) the GPT architecture itself (decoder-only Transformer with causal attention mask) was not patented; (4) RLHF (Reinforcement Learning from Human Feedback) procedure for aligning language models to human preferences was published in InstructGPT paper (Ouyang et al. 2022) rather than patented; OpenAI does hold patents on specific system designs for its deployed products (ChatGPT interface features, API rate limiting methods) but the foundational LLM technology is deliberately not patent-protected to prevent the defensive patent accumulation that characterized the smartphone patent wars. OpenAI's lack of patent aggression is a strategic choice reflecting its unusual mission structure.

NVIDIA holds foundational GPU architecture patents that underpin essentially all modern AI training: CUDA (Compute Unified Device Architecture) parallel computing architecture patents; Tensor Core specialized matrix multiplication unit patents (introduced Volta architecture 2017); NVLink and NVSwitch GPU interconnect patents (enabling multi-GPU training at scale); cuDNN, cuBLAS, NCCL library algorithm patents (though most library details are trade secrets or openly published); Blackwell B200 and Grace Blackwell GB200 architecture patents. NVIDIA's patent strategy in AI is hardware-centric: NVIDIA files patents aggressively on chip architecture innovations but benefits from an open ecosystem for software — CUDA's openness attracts the developer base that makes NVIDIA's hardware indispensable for AI training.

IBM has filed the most AI patent applications globally by volume for many years (IBM's patent count has consistently led the USPTO filing league tables). IBM AI patent domains: Watson-era NLP system patents (question-answering, information extraction); quantum computing algorithm patents (quantum AI, quantum machine learning); IBM Research AI in materials science (molecular design AI); enterprise AI governance and explainable AI patents; IBM's large patent portfolio is partly a cross-licensing tool for its business relationships rather than purely an offensive enforcement portfolio.

No — AI cannot be named as an inventor in the US, UK, EPO, or most major patent jurisdictions. The Thaler/DABUS cases resolved this question definitively: (1) USPTO (US): the Court of Appeals for the Federal Circuit (CAFC) affirmed in Thaler v. Vidal (2022) that 35 U.S.C. § 100(f) requires the inventor to be an individual (natural person); AI systems lack legal standing as inventors; USPTO's February 2024 guidance confirms that AI-assisted inventions are patentable when a human made a significant contribution to the conception of the claimed invention; (2) UK: the UK Supreme Court ruled in Thaler v. Comptroller-General of Patents (2023) that DABUS could not be named as inventor under the UK Patents Act 1977 — only a natural person can be an inventor. The UK Supreme Court noted this was a policy question for Parliament if it wanted to extend inventorship to AI; (3) EPO: the EPO rejected the DABUS applications and the Technical Board of Appeal upheld the rejection — AI systems cannot be inventors under the EPC; (4) Australia: the Full Federal Court of Australia reversed the earlier first-instance ruling that had allowed AI inventorship (the only court to initially allow it) — AI cannot be named as inventor in Australia. Bottom line for AI founders and inventors: the humans who direct, design, and apply the AI system are the inventors. If you use GitHub Copilot to write code that implements your invention, or use a large language model to suggest chemical compounds, you are the inventor — the AI is a tool. Name the humans who conceived the claimed invention. Document the human inventive contributions clearly for inventorship disputes.

The Alice Corp. v. CLS Bank Int'l (2014) two-step test applies to AI/ML patent claims and creates a significant eligibility risk: Step 1 (directed to abstract idea): most AI/ML methods are inherently mathematical — neural network training is a mathematical optimization process (gradient descent, backpropagation); inference is a mathematical computation (matrix multiplications, nonlinear activations). Examiners frequently characterize AI claims as directed to mathematical concepts or mental processes at Step 1. Step 2 (inventive concept): even if directed to an abstract idea, the claim survives if it has 'something more' that transforms it into patent-eligible subject matter — a specific technical application, a specific improvement to computer functionality, or integration with specific physical hardware beyond generic computing. How to survive Alice for AI patents: (1) Specificity of application: claims directed to a specific technical domain (diagnosing a specific disease from a specific imaging modality; detecting a specific class of cyberattack from specific network traffic features) are more likely to survive than generic 'classify data using neural network' claims; (2) Technical improvement to computer functionality: if the AI innovation genuinely improves how a computer system operates — faster inference through a novel hardware-software co-design, reduced memory bandwidth through specific quantization method, improved accuracy with fewer computations through architectural innovation — claim the specific technical improvement; (3) Hardware-specific claims: claims that recite specific hardware elements (GPU/TPU instructions, specific memory access patterns, specific SIMD operations) beyond generic computing infrastructure are stronger; (4) Method of treatment / medical device claims: AI medical device claims are often analyzed under the medical treatment doctrine and may have a different eligibility analysis if the claim is tied to a specific medical purpose; (5) USPTO 2019 Revised Guidance Step 2A: the USPTO's 2019 Guidance for examiners includes a Step 2A Prong 2 practical application gate — if an abstract idea is integrated into a practical application (even if the abstract idea itself is mathematical), the claim passes Step 1. Use this prong: argue the neural network classification is integrated into the specific practical application of real-time industrial defect detection, autonomous vehicle pedestrian avoidance, or targeted drug delivery optimization.

The patent vs. trade secret decision for AI innovations depends on what the innovation is: (1) PATENT these AI innovations: specific system architectures for deploying AI in a defined product (how camera + LiDAR + AI inference work together in an autonomous vehicle system); novel training procedures or loss functions that achieve meaningfully better performance and can be described in a claim with enough specificity to satisfy Alice (if it can be claimed broadly enough to be commercially valuable without being so broad it fails § 101 or obviousness); novel data pipeline and preprocessing methods with technical character; AI application to specific technical domains (specific medical diagnostic systems, specific industrial process control systems, specific scientific research automation pipelines); hardware-software co-design innovations for efficient AI computation; (2) TRADE SECRET these AI innovations: model weights (the trained parameters of any AI model); training data composition, sourcing, and curation methodology (who was in the training dataset, how data was selected, cleaned, and weighted); scaling laws and empirical observations about how model performance scales; hyperparameter optimization findings; evaluation benchmark designs and red-teaming procedures (publishing these helps attackers); (3) OPEN SOURCE STRATEGICALLY (forgoing both patent and trade secret) for: foundational framework code (TensorFlow, PyTorch, ONNX) — open-sourcing builds an ecosystem and developer adoption that benefits the platform; model architectures that become industry standards (Meta's decision to open-source LLaMA created a downstream ecosystem of fine-tuned models that reinforced LLaMA as a standard, generating commercial value through cloud services even as the weights became open). The key strategic question: can your AI innovation be reverse-engineered from the deployed product? If yes (e.g., architectural improvements in a robotics system that competitors can observe), patent. If no (e.g., training data curation that is invisible in the deployed model), trade secret.

No — the original Transformer architecture ('Attention Is All You Need', Vaswani et al., NeurIPS 2017) was not patented. Google Brain and Google Research published the paper openly, establishing academic priority and making the architecture prior art for any subsequent patent claims — which means no one else can patent the basic Transformer architecture either. This was a strategic choice by Google: (1) Publishing the paper established Google's technical leadership and recruited talent without requiring commercial secrecy; (2) Open publication enabled a massive ecosystem of Transformer-based research (BERT, GPT, T5, ViT, etc.) that benefited Google's research capabilities; (3) The alternative — patenting the Transformer — would have been difficult to enforce broadly because the attention mechanism builds on prior work in attention-based neural machine translation (Bahdanau et al. 2015, itself published without patenting), and a broad patent might have been challenged on obviousness grounds given the prior attention literature. The result: the Transformer architecture is unowned prior art. Any company can freely use the attention mechanism as described in the original paper. Subsequent innovations built on Transformers — specific fine-tuning methods (LoRA, QLoRA), specific architectural modifications (Flash Attention memory efficiency, Rotary Position Embeddings [RoPE], Grouped Query Attention [GQA]) — may be patentable if novel and non-obvious, and some have been. Flash Attention (Tri Dao et al., Stanford) — the memory-efficient attention computation algorithm that reduces VRAM requirements from quadratic to near-linear — was patented. However, the paper was also published openly, meaning competitors can understand the general method; only the specific claim language is protected by the patent.

AI patent applications face somewhat different eligibility hurdles at the EPO versus USPTO: (1) Framework difference: the USPTO applies the Alice two-step test (abstract idea → inventive concept). The EPO applies the technical character test — a CII (computer-implemented invention) must have a 'technical character,' typically meaning it produces a 'further technical effect going beyond the normal physical interactions between a program and the computer on which it runs' (T 1173/97). An AI method that classifies data for a non-technical purpose (business analytics, market prediction, aesthetic image enhancement) lacks technical character and is excluded under EPC Art. 52(2). An AI method applied to a technical domain has technical character; (2) Practical permissiveness: in practice, the EPO's technical character test for AI is often MORE permissive than Alice for the same invention. Many AI medical imaging inventions that struggle under Alice Step 2 pass easily under EPO technical character because medical diagnosis is inherently technical. Google, Microsoft, and large AI companies file broadly at both USPTO and EPO, with EPO coverage often providing the most reliable protection for AI medical and industrial applications; (3) AI/ML Guidelines (EPO, 2021/2023 updates): the EPO published specific examination guidelines for AI/ML in the CII Guidelines section. Key guidance: (a) the use of AI to solve a technical problem qualifies as a technical contribution; (b) a mathematical method used to solve a non-technical problem (customer churn prediction, ad targeting) is not a technical contribution; (c) classifying physical objects/phenomena (radar, seismic signals, medical images, spectroscopy data) = technical; (d) classifying abstract concepts (text sentiment, legal document categories, financial risk) = not inherently technical; (4) Inventive step and technical contribution: at the EPO, the problem-solution approach is used for inventive step. For an AI patent to have inventive step, the AI approach must contribute a non-obvious technical improvement — not just 'we used a deep neural network' where any skilled person would consider using a neural network, but a specific architectural choice, training methodology, or application integration that was non-obvious from the starting state of the art.