Artificial Intelligence IP Strategy

AI companies face a unique IP challenge: the most valuable AI assets (model weights trained on billions of parameters) are not naturally suited to patent protection — but a sophisticated multi-layer IP strategy can provide meaningful protection across all three IP forms: THE AI IP PROTECTION PYRAMID: LAYER 1 — TRADE SECRETS (PROTECT THE MODEL ITSELF): model weights are the single most valuable AI asset; a trained GPT-class model represents billions of dollars in compute and data curation; TRADE SECRET PROTECTION: model weights can be protected as trade secrets if: (1) they have independent economic value from not being generally known; (2) reasonable measures are taken to maintain secrecy; model architecture variants and hyperparameters not disclosed in papers; training data mixture, proportions, and curation methodology; RLHF + fine-tuning process details beyond what is published; LAYER 2 — PATENTS (PROTECT TECHNICAL IMPLEMENTATIONS): patent the technical innovations that produced the breakthrough — not the weights themselves; what can be patented: training algorithms with specific improvements (Flash Attention's IO-aware tiling algorithm; Ring Attention for long context; Mixture of Experts routing mechanisms; specific RLHF reward modeling architectures); inference optimizations (speculative decoding; KV cache compression); hardware-software co-design (custom training chip architectures; memory bandwidth optimizations); deployment systems (batching algorithms; request scheduling; auto-scaling for inference); data pipeline innovations (deduplication algorithms; quality filtering systems); evaluation systems (specific benchmark architectures; safety evaluation methods); LAYER 3 — COPYRIGHT (PROTECT TRAINING DATA AND OUTPUTS): curated training dataset collections may have copyright protection (the compilation/selection is copyrightable even if individual data points are not); LAYER 4 — CONTRACTS: terms of service restricting API use; model cards; acceptable use policies; enterprise agreements with IP ownership provisions; MODEL CARD AND SYSTEM CARD DISCLOSURE: AI labs like OpenAI; Anthropic; Meta publish model cards and system cards; these constitute public disclosures that can create prior art for others while also establishing evidence of technical priority; time the publication carefully relative to patent filing.

AI patent eligibility under § 101 is one of the most active areas of patent law — the Alice/Mayo framework creates a real barrier for abstract algorithm claims, but well-crafted AI claims tied to specific technical improvements regularly survive examination: AI INNOVATIONS THAT TYPICALLY SURVIVE § 101: SPECIFIC TECHNICAL IMPROVEMENTS: claims directed to a specific technical improvement in computer functionality (not just faster computation abstractly, but specific architectural improvement); Flash Attention (Dao et al.): IO-aware attention algorithm that reduces DRAM reads/writes through tiling; creates specific, concrete improvement in transformer training speed; this is patentable because it is tied to specific hardware behavior and produces measurable technical result; claims on specific training hardware interactions — HBM access patterns; SRAM tiling — are specific technical implementations, not abstract algorithms; SPECIFIC ARCHITECTURE INNOVATIONS: Mixture of Experts (MoE) routing with specific gating functions and load balancing is more likely patentable than a generic 'use MoE' claim; Retrieval Augmented Generation (RAG) with specific indexing and retrieval architectures; specific multi-modal fusion architectures that produce measurable quality improvements; specific positional encoding variants (ALiBi; RoPE; YaRN) with provable context extension benefits; SAFETY AND ALIGNMENT TECHNIQUES: constitutional AI (specific reward modeling approach); RLHF with specific KL penalty implementations; specific jailbreak detection architectures; AI INNOVATIONS VULNERABLE TO § 101: 'applying machine learning to [field]' without specific technical implementation; generic transformer with standard attention claiming broad scope; abstract claim to 'training a neural network on data to predict X'; claims reciting mathematical operations on abstract data without specific technical anchoring; CLAIM DRAFTING STRATEGY FOR AI: ANCHOR CLAIMS IN HARDWARE INTERACTIONS: specify memory access patterns; specify compute units (GPU; TPU; custom ASIC); specify bandwidth constraints that the innovation addresses; QUANTIFY THE TECHNICAL IMPROVEMENT: claims should be supported by specification showing measurable improvement (X% faster training; Y% less memory; Z% improvement on benchmark); this supports both patent eligibility and obviousness arguments; SPECIFIC ARCHITECTURE ELEMENTS: claim specific layer types; specific normalization techniques; specific initialization schemes that produce the technical benefit; USE DEPENDENT CLAIMS LIBERALLY: independent claim captures the concept at some breadth; dependent claims cover specific numbers of layers; specific activation functions; specific training procedures; PROSECUTION HISTORY: examiner will examine AI claims carefully; prepare detailed technical evidence of the improvement; have declarations from ML researchers available.

AI inventorship is one of the most actively developing areas of patent law — with the USPTO; Federal Circuit; and international patent offices all grappling with how to apply centuries-old inventorship doctrine to AI-assisted innovation: THE FUNDAMENTAL INVENTORSHIP RULE: THALER v. VIDAL (Fed. Cir. 2022): DABUS (an AI system created by Stephen Thaler) was designated as the inventor on patent applications in the US; UK; EPO; Australia; South Africa; Federal Circuit held: 'inventors must be natural persons'; AI systems cannot be inventors under US patent law (35 U.S.C. § 100(f): 'whoever invents or discovers any new and useful process…' = human); US SUPREME COURT declined to review; THALER v. VIDAL is settled law: AI cannot be an inventor; GLOBAL PICTURE: UK Supreme Court (Thaler 2023): AI cannot be inventor under UK Patents Act; EPO: same conclusion; South Africa: initially granted patent with DABUS as inventor; later clarified human must be listed; Australia: initially allowed AI inventor; Full Federal Court reversed 2022; WHO IS THE INVENTOR WHEN AI ASSISTS?: USPTO's February 2024 Guidance on AI-Assisted Inventions: the natural person who 'conceived' the invention is the inventor; conception = 'formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention'; KEY DISTINCTION: if a human provides the idea and uses AI as a tool to implement it → human is the inventor; if AI autonomously generates the invention without human conception of the specific invention → no valid inventorship (application cannot be granted currently under US law); PRACTICAL INVENTORSHIP CHECKLIST FOR AI-ASSISTED INNOVATIONS: (1) document that a human inventor identified the technical problem; (2) document that a human inventor formulated the approach or directed the AI's use; (3) document that a human inventor recognized and appreciated the technical result; (4) the fact that AI generated the specific code/output does not defeat human inventorship if the human conceived what the AI should produce; COMPANY POLICIES FOR AI-ASSISTED INVENTIONS: create AI use logs documenting human direction of AI in the invention process; maintain invention disclosure form fields for 'AI involvement' to identify review-required cases; when developers use Copilot or similar tools to generate code, the human who specified the technical problem and directed the AI remains the inventor; TRADE SECRET ALTERNATIVE WHEN INVENTORSHIP IS UNCLEAR: if the level of AI autonomy makes inventorship unclear, default to trade secret protection to avoid disclosure risk.

The leading AI labs have developed distinct IP strategies reflecting their different corporate structures, research philosophies, and competitive positions — and the copyright question for AI-generated content is creating new legal uncertainty globally: OPENAI IP STRATEGY: OpenAI's founding principle was openness but GPT-4 and beyond are closed-source; PATENTS: OpenAI has filed patents on specific technical innovations but maintains a primarily trade-secret-first strategy for model weights and architecture details; POLICY POSITION: OpenAI's charter commits to safety and broad benefit but patent strategy has become more aggressive as commercial competition increased; SYSTEM CARD DISCLOSURES: technical details published in system cards (like GPT-4 Technical Report) establish prior art for competitors while preserving trade secrets on implementation; GOOGLE DEEPMIND: most aggressive patent filer in AI among major labs; thousands of AI-related patent applications; broad coverage: transformer variants; reinforcement learning; specific ML architectures; AlphaFold protein structure prediction innovations patented alongside open sourcing of prediction results; strategic balance: open source research to attract talent + patent core technical innovations for competitive advantage; PATENT PLEDGES: Google has made patent pledges on some open source ML frameworks (TensorFlow); META (FAIR): Meta's AI research is heavily published through FAIR (Fundamental AI Research); LLaMA series open-released; BUT infrastructure patents for serving AI at scale; specific optimization innovations; distributed training methods are likely protected trade secrets or pending patents; META'S OPEN MODEL STRATEGY: by open-releasing LLaMA models, Meta creates a large ecosystem that is hard to compete with on closed models alone; uses 'open source as IP strategy' — making its model weights the de facto open standard for fine-tuning creates ecosystem lock-in even without patents; ANTHROPIC: constitutional AI approach (Bai et al. 2022) published as research; specific safety evaluation techniques; responsible scaling policy; IP strategy focused on maintaining competitive position through continued research advancement rather than heavy patent filing; AI-GENERATED CONTENT AND COPYRIGHT: Thaler v. Perlmutter (D.D.C. 2023): US Copyright Office; confirmed ONLY human authorship qualifies for copyright protection; AI-generated images in 'A Recent Entrance to Paradise' (DABUS artist companion) — no copyright; ZARYA OF THE DAWN case: Copyright Office granted copyright to human artist for text and selection/arrangement but refused copyright for AI-generated images within the comic; IMPLICATIONS: AI-generated text; images; code without human creative expression = no copyright protection; AI-assisted works where human makes creative choices = copyright for the human expression elements; companies cannot copyright output of AI without human creative involvement; this creates gaps in protection for AI-generated code; documentation; or marketing materials.