How AI Startups Are Building Their Patent Portfolios: Lessons From the Coming IP Wars in Artificial Intelligence

Artificial intelligence is already transforming every sector of the economy — and it is about to transform patent law itself. The most revolutionary startup sector of the current decade is simultaneously the one with the most unsettled IP landscape.

The AI Patent Filing Surge

According to WIPO data, AI-related patent applications grew by an average of 28% annually in recent years. Large technology companies still dominate in total AI patent grants. But emerging AI startups are building portfolios aggressively — covering large language model training methodologies, reinforcement learning from human feedback (RLHF) techniques, and inference optimization methods. The strategic logic mirrors earlier technology waves: establish IP positions during the founding period before competitors can independently develop similar approaches.

What AI Patents Actually Cover

Training methodologies covering specific techniques for training neural networks — particular loss functions, novel architectures, or optimization algorithms — can provide strong protection because they affect the quality of any model trained using them.

Model architectures such as the structural configuration of a neural network can be patented if they constitute a non-obvious technical advance. The original transformer architecture described in "Attention Is All You Need" (2017) was not patented — an enormous lost opportunity. The lesson was not lost on subsequent AI companies.

Application-specific implementations — even where the underlying model is generic, applying AI to a particular technical problem using a specific pipeline can be patented if the implementation constitutes a non-obvious technical advance.

The § 101 Eligibility Challenge

AI patents face the same Alice challenges as other software patents, amplified. A neural network is, at its mathematical core, a system of linear algebra operations — abstract mathematical operations. The strongest AI patents navigate this by: (1) tying the claimed method to specific hardware configurations; (2) claiming the interaction between the model and a physical environment (a robot arm, a medical imaging device, an autonomous vehicle sensor array); or (3) demonstrating a specific technical improvement in computational performance measurably better than prior approaches. The USPTO's 2019 AI subject matter eligibility guidance, incorporated into MPEP § 2106, is heavily tested on the patent bar exam.

Who Can Be a Named Inventor? The AI Authorship Problem

The USPTO addressed whether an AI can be a named inventor in 2020, in the context of Stephen Thaler's DABUS AI system. The USPTO denied the applications, holding that under 35 U.S.C. § 100(f), an "inventor" must be a natural person. The Federal Circuit affirmed in Thaler v. Vidal (2022). AI-generated inventions currently require a human inventor who directed the AI's work, recognized the inventive output, and contributed meaningfully to conception.

Trade Secrets as an Alternative

Many AI companies are choosing trade secret protection as their primary IP strategy. Under the Defend Trade Secrets Act (DTSA) of 2016, training data, preprocessing methods, model weights, and alignment techniques can be protected indefinitely as long as reasonable confidentiality measures are taken and they derive independent economic value from their secrecy. The tradeoff: trade secrets evaporate if independently discovered or disclosed, while patents provide 20-year exclusivity even against independent invention. For AI model weights that are practically irreproducible by reverse engineering, trade secrets may provide stronger practical protection. For specific training techniques that might be independently discovered, patents provide a stronger position. The strategic choice between these two regimes is one of the most important IP decisions a founding AI team will make.