Every year, when the preview of the new EPO Guidelines are released, I do the same thing: I open the PDF and immediately CTRL+F for “Artificial Intelligence”.
The preview of the Guidelines due to enter into force on 1 April 2026 are all about stability, liabilitym and the “human in the loop”. Here’s my summary:
1. The “Non-Update”: Patentability is Stable
If you were worried about a new, complex test for AI patentability, you can relax. The core section on AI and machine learning (Part G-II, 3.3.1) has not undergone major substantive changes.
This signals that the EPO is confident in its current approach. For those of you who are familiar with the EPC but haven’t dived deep into AI, the EPO handles it using the same “Two-Hurdle” approach used for all computer-implemented inventions:
- Hurdle 1 (Eligibility): The claim must have technical character. This is easy; just mentioning a computer, a network, or a processor is usually enough to pass Article 52 EPC.
- Hurdle 2 (Inventive Step): This is the real test. The EPO uses the “Comvik” approach. They look at the mathematical steps (the AI algorithm) and ask: Do these steps contribute to a technical solution for a technical problem?
- Yes: If the AI processes images to identify tumors (medical purpose) or controls a robot arm (technical control), the math counts towards inventive step.
- No: If the AI merely predicts stock prices (business method) or classifies text by sentiment (linguistics), the math is ignored for inventive step.
The Takeaway: The “Comvik” approach is here to stay. The 2026 Guidelines confirm that the EPO treats AI as a mathematical method that needs a technical application to be patentable.
2. The New Rule: “The Human in the Loop”
The biggest change in 2026 isn’t about patenting AI but about using AI. A completely new section has been added to the General Part (Section 5) titled “The use of artificial intelligence.”
This section explicitly states that applicants and their representatives are fully responsible for the content of their patent applications and submissions, regardless of whether an AI tool was used to prepare them.
Why this matters: We all know that LLMs can “hallucinate”, i.e., they can invent prior art and case law, or describe technical features that don’t exist. The EPO is drawing a line in the sand: you cannot blame the bot.
If your specification contains technical inaccuracies because ChatGPT wrote it, or if you cite a Board of Appeal decision that doesn’t exist, “the AI made a mistake” will not be accepted as an excuse. This reinforces the need for a “Human in the Loop” to verify every line of an automated draft.
3. Procedural Update: AI in the Examining and Opposition Division
Finally, the EPO is practicing what it preaches by integrating AI into its own internal workflows. Part E, Chapter III, 10.1 has been updated to formalize the use of AI in drafting minutes for oral proceedings held by videoconference.
How it works: The EPO will make a sound recording of the proceedings to help an AI tool generate a draft of the minutes. These recordings are purely for the internal purpose of drafting minutes. They are deleted after the minutes are issued and are not made available to the parties.
Conclusion
The 2026 Guidelines tell a clear story: the “Wild West” phase of AI at the EPO is ending. We have moved into a phase of stabilization and regulation.
Substantively, the rules are set: AI is a mathematical method that requires a technical purpose.
Procedurally, the rules are strict: Use AI tools all you want, but if they mess up, the liability is 100% yours.
For practitioners, the advice for 2026 is simple: Don’t stop using AI tools, but strictly treat them as drafters, never as authors. The final signature, and the final responsibility, is yours.
By the way, if you need an easy intro to AI-powered patent drafting and see my own workflow in action, join my AI for Patent Attorneys program. I’ll showcase my entire toolbox, from the exact prompts I use to the local privacy setup that keeps me compliant.
Hope this helps!
Bastian