While tech companies scramble to patent every AI breakthrough they can dream up, the USPTO decided it was time to throw them a bone—or maybe a curveball. In August 2025, the patent office issued new guidance on AI and machine learning patent eligibility under 35 U.S.C. § 101. The goal? Better innovation outcomes. The reality? It's complicated.

The guidance leans heavily on the Manual of Patent Examining Procedure sections 2103-2106.07 and emphasizes the Alice/Mayo two-step test. That's the framework that asks whether a claim involves a judicial exception like abstract ideas, then whether it integrates into a practical application. Sounds simple enough. It's not.

The Alice/Mayo test sounds straightforward on paper—judicial exceptions plus practical applications—but patent reality is messier than legal theory suggests.

Federal Circuit rulings have made one thing crystal clear: slapping machine learning onto new data doesn't magically create patent-eligible subject matter. Courts want genuine technological advancements in ML models themselves, not just creative new uses for existing techniques. Generic applications get rejected. Hard.

The Recentive decision hammered this point home. Conventional ML training alone won't cut it anymore. Patent applicants need to show their trained models perform tasks in technically novel ways. Simply confining ML claims to a particular field isn't enough without describing non-generic computing improvements. The message is blunt: explain how your innovation works, not just what it does.

This creates a fascinating tension. The USPTO's 2025 guidance aims to clarify software-related inventions and help examiners distinguish patent-eligible AI inventions from abstract ideas. They even included AI-specific examples to promote consistency. Notably, complex AI computations are not automatically categorized as mental processes or abstract ideas, giving sophisticated AI claims a better chance at eligibility.

Meanwhile, Federal Circuit decisions keep raising the bar for what constitutes genuine technological advancement. Patent protection remains crucial for maintaining the integrity of AI innovations across the industry. Patent attorneys now face a tricky balancing act. Claims must detail specific improvements to ML models rather than broad applications. Technical implementation matters more than ever. The guidance establishes that claims training neural networks involve rather than recite abstract ideas.

The Alice/Mayo test remains central, with its two-pronged analysis of judicial exceptions and practical applications. The result? AI patent strategy requires surgical precision. Blanket claiming risks rejection. Success depends on demonstrating concrete technological improvements, not just novel use cases. The USPTO may have clarified the rules, but playing by them just got harder.