Blog

Patenting AI

By: Paul G. Nagy, Partner, Berkeley Law & Technology Group, LLP

Over recent years, and continuing through 2024–2026, U.S. federal courts, particularly the Federal Circuit, have maintained a generally skeptical approach to patent eligibility for machine-learning–related inventions under 35 U.S.C. § 101.  While courts recognize that machine learning is an important technological tool, they have repeatedly emphasized that merely invoking machine learning or artificial intelligence does not, by itself, confer patent eligibility.

Two-Step Framework From Alice Corp. v. CLS Bank

Courts have consistently applied a two-step framework established by the U.S. Supreme Court in Alice Corp. v. CLS Bank.  In a step one, claims directed to using machine-learning models for data analysis, prediction, classification, or optimization are often characterized as being directed to abstract ideas, such as mathematical relationships, data processing, or mental processes performed more efficiently by a computer.  If a claim is determined to be directed to such an abstract idea, in a step two courts may look for an “inventive concept” that transforms the abstract idea into patent-eligible subject matter.  Recent decisions emphasize that describing generic machine-learning techniques, such as training models, iteratively updating weights, or optimizing outputs, by themselves, typically do not supply an inventive concept when those techniques are used in a conventional computing environment and for well-understood business or organizational goals.

Recent Court Decisions

The recent decision by the U.S. Court of Appeals of Federal Circuit in Recentive Analytics, Inc. v. Fox Corp. affirmed dismissal of patents directed to using machine learning to generate event schedules and television network maps.  Claims at issue in this case recited actions such as collecting parameters, training machine-learning (ML) models, generating optimized outputs, and updating those outputs in real time.  In applying the U.S. Supreme Court’s two-step framework, the court held that the claims were directed to abstract ideas—namely, organizing and optimizing schedules and broadcast maps—implemented using generic machine-learning techniques in step one.  The court stressed that the patents did not claim improvements to machine-learning technology itself, but instead applied well-known ML methods to a particular business context.  In applying step two, the court found no inventive concept because the claims relied on “any suitable machine learning technique” and generic computing hardware. The opinion underscores a recurring theme: functional claiming of machine-learning results, without technical specificity as to how the model operates or improves computer functionality, is insufficient.

U.S. District Courts continue to follow similar reasoning.  Recent decisions have invalidated/dismissed AI-related claims that focus on using ML models to analyze data and produce recommendations or predictions, emphasizing that reciting training data, model types (e.g., neural networks or support vector machines), or iterative learning steps does not meaningfully distinguish claims from abstract data analysis unless tied to a concrete technological improvement.  Collectively, these cases signal that courts remain wary of claims that treat machine learning as a “black box” tool for achieving results that could be framed as longstanding human activities or business practices.

The United States Patent and Trademark Office Perspective

Recognizing the importance of AI-related inventions (and ML inventions in particular), the United States Patent and Trademark Office (USPTO) has taken steps to clarify and, in some respects, soften examination practice for AI-related inventions, while still operating within the U.S. Supreme Court’s two-step framework.  Recent USPTO guidance and training materials emphasize that AI and machine-learning inventions are not categorically ineligible.  Examiners are instructed to conduct a careful, claim-specific analysis under the Alice framework.  In particular, the USPTO has highlighted that claims may be patent-eligible when they recite a specific technological improvement, such as an improvement to computer performance, data processing efficiency, or the functioning of a machine or system.

Additionally, Patent Trial and Appeal Board’s (PTAB’s) decision in Desjardins and recent internal memoranda at the USPTO caution examiners against oversimplifying claims by merely labeling them as “using AI” or “using machine learning.”  Instead, examiners are encouraged to identify what the claims are “directed to” and whether they recite specific limitations that integrate any abstract idea into a practical application.

USPTO training examples and internal memoranda emphasize that claims are more likely to satisfy § 101 when they:

  • specify how training or inference is performed in a non-conventional way;
  • improve the operation of a computer or network itself (e.g., reduced memory usage, faster convergence, improved robustness);
  • are tied to a particular machine or technological environment in a meaningful, non-generic manner.

At the same time, the USPTO acknowledges that claims focused on organizing human activity, business decision-making, or results-oriented data analysis—implemented using conventional ML techniques—remain vulnerable to 35 USC § 101 rejections.

Key Lessons to Satisfy Both the USPTO and Federal Courts

U.S. Patent applicants should keep in mind that policy and decisions by the USPTO regarding application of 35 USC § 101 are not binding on Federal Courts.  While an applicant may succeed in overcoming 35 USC § 101 rejections by emphasizing practical applications and technical details during prosecution, issued claims may still face significant risk in court if the claims are viewed as delineating abstract ideas implemented with generic ML tools.  As such, U.S. Patent applicants are encouraged to draft claims and specifications with an eye toward both examination and litigation, providing detailed technical disclosures that go beyond high-level descriptions of machine learning.

Taken together, recent developments reflect a cautious and, in some respects, restrictive environment for patenting machine-learning inventions under 35 USC § 101. Federal courts continue to demand specificity and technological substance, rejecting claims that merely apply known ML techniques to familiar problems or business contexts. The USPTO, while more receptive in principle, increasingly aligns its analysis with these judicial standards.

For U.S. patent applicants, the key lessons are clear:

  • claims should focus on concrete technical improvements (e.g., technical solutions to technical problems), not just improved outcomes;
  • specifications should explain how the machine-learning technique operates at a technical level and why it represents more than routine or conventional implementation; and
  • reliance on broad functional language (“using a machine-learning model to optimize”) remains a significant risk.

Contact Us

If you have an AI related invention that you would like to patent and want to evaluate whether it is patentable, or need assistance preparing a patent application, experienced guidance matters. For more information about patenting AI, contact Paul Nagy at Berkely Law & Technology Group, LLP at pnagy@bltg-ip.com or 503-439-6500.