Key Takeaways from the Patenting AI Masterclass

Artificial intelligence is reshaping whole industries, and with that comes a pressing question: how do we protect AI-driven innovations through patents?

On 5 March 2025, I hosted the Patenting AI Masterclass, where we dove deep into the challenges and strategies of patenting artificial intelligence. It was an engaging session with patent professionals from law firms and industry patent departments, all eager to understand how to navigate this complex field.

Here are my key insights from the three main topics of the masterclass:

1. Can AI be patented?
2. How to draft strong AI patent claims
3. How to ensure an enabling disclosure in an AI patent

Can AI Be Patented?

One of the most sticky misconceptions about AI patents is that AI-related inventions are automatically excluded from patent protection. While programs for computers are listed among the exclusions in Article 52(2) of the European Patent Convention (EPC), this does not mean that all AI innovations are unpatentable. In a nutshell, AI-based inventions can be patented if they solve a technical problem in a non-obvious way.

The Two-Hurdle Approach to AI Patentability at the EPO

The European Patent Office (EPO) applies a two-step test to determine whether AI-related inventions are patentable:

1. Does the invention have technical character?
   • Under the EPO’s “any hardware” approach, an invention is considered patent-eligible as long as it includes any technical means, even something as trivial as a generic computer. This is in stark contrast to the stricter “patent-eligibility” standards known from the US system.
   • If AI is merely used for business automation or abstract decision-making, it remains formally patent-eligible but may be rejected under the inventive step requirement at the second hurdle.
2. Does the AI contribute to an inventive step in a technical way?
   • The EPO applies the so-called COMVIK approach, which means that only those features that contribute to solving a technical problem are considered when assessing inventive step. Any non-technical aspects, such as business logic, administrative schemes, or abstract mathematical models, are ignored in this assessment.

Patentable AI Innovations: Two Viable Routes

The EPO generally views AI with suspicion these days, considering it to be inherently abstract mathematics, which is non-technical. AI can only become a technical tool in two exceptional ways:

1. Technical Applications: AI applied to real-world technical problems, such as autonomous driving, cybersecurity, or medical diagnostics.
2. Technical Implementations: AI optimized for specific hardware architectures to improve hardware efficiency, computing power consumption, or processing speed.

How to Draft Strong AI Patent Claims

Even if an AI invention meets the requirements for patentability, its claims must be carefully drafted to ensure enforceability. Poorly written claims can easily lead to a patent that is too narrow to be valuable.

What Makes a Strong AI Patent Claim?

A well-drafted AI patent claim should:

• Clearly define the technical contribution of the AI system.
• Be detectable by looking at the competitor’s product.
• Avoid issues of divided infringement, ensuring that at least one entity (such as a competitor) is fully responsible for infringement.

Examples of typical AI patent claims

These are some examples of useful claim formulations for an AI patent:

A computer-implemented method of <overall purpose>, comprising:
     receiving an input dataset comprising <inputs>;
     <operating step(s) specific to the invention>; and   
     producing an output dataset comprising <outputs>.

A computer-implemented method of training a machine-learning model for <overall purpose>, in particular the machine-learning model of claim <...>, comprising:
     receiving an input training dataset comprising <inputs>; and
     <training step(s) specific to the invention>.

A machine-learning model for <overall purpose>, in particular for use in the method of any one of claims <...>, comprising:
     <structural features of the model>.

I shared many more examples in the seminar. If you want the full set, I invite you to join my next Patenting AI Masterclass.

How to Ensure an Enabling Disclosure in an AI Patent

Even with strong claims, an AI patent application will be rejected if the disclosure is insufficient. AI inventions must be described in enough detail to allow a skilled person to reproduce the invention without undue experimentation.

Why Black-Box AI Patents Get Rejected

One of the main reasons AI patents are rejected is that applicants fail to disclose the details of the AI model, treating it as a black-box system.

• Example Case (T 0161/18): A patent application for an AI-based cardiac output prediction model was rejected because the training data was not disclosed, making the invention impossible to reproduce.

What Must Be Disclosed?

A strong AI patent disclosure should include:

1. Computational model details – The architecture of the AI system and the algorithms used.
2. Specific input and output variables – A clear description of what the AI processes.
3. Training methodology – How the AI model is trained, including hyperparameters and learning techniques.
4. Demonstrated technical effect – Evidence that the AI provides a measurable improvement over existing solutions.
5. Test results and benchmarks – Including experimental data improves credibility and avoids rejection.

In my masterclass, participants received a full checklist of questions to ask inventors to ensure a complete and detailed AI invention disclosure. This checklist helps streamline the process, making sure all relevant technical details, including training data, model architecture, and performance improvements, are properly captured in the patent application.

Conclusion: Protecting AI Innovations with Strong Patents

AI patents are challenging but achievable. The Patenting AI Masterclass demonstrated that success depends on three key elements:

1. AI is patentable, but applications must clearly define a technical contribution.
2. AI patents require well-written claims that are enforceable and avoid divided infringement.
3. AI patent disclosures must be fully enabling and reproducible, with sufficient technical detail.

To gain practical expertise in drafting, claiming, and disclosing AI patents, join the next session of the Patenting AI Masterclass.

I hope I see you there!
Bastian