EIC Support Best Practice
Understanding the Challenges of AI-generated Inventions
The intersection of artificial intelligence (AI) and intellectual property👉 Creations of the mind protected by legal rights. (IP) law presents both unprecedented opportunities and complex challenges. This blog post synthesizes insights from a CEIPI IP Business Academy lecture (part of the Diplôme Universitaire (DU) IP Business Administration) and the 🔗dIPlex digital IP lexicon’s recommendations for AI-generated inventions👉 An invention created autonomously or semi-autonomously by an AI system.. By aligning academic frameworks with real-world best practices, we show how deep tech👉 Cutting-edge innovations rooted in scientific breakthroughs and R&D. companies can navigate this evolving landscape.
Key Takeaways from the CEIPI-EPO DU Lecture
The CEIPI lecture, delivered by patent👉 A legal right granting exclusive control over an invention for a limited time. attorney Axel Karl, focuses on patenting AI innovations under the European Patent Convention (EPC). Here you can watch the lecture:
Lecture from University Diploma IP Business Administration Part 8 International Experience with Digital Patents , Chapter 7: AI fundamentals & Patenting Artificial Intelligence
Below are its core principles:
Defining AI in the Patent Context
AI is rooted in mathematical methods, machine learning (ML), and deep learning (DL). While AI systems learn autonomously from data, patentability hinges on demonstrating technical character:
- Machine Learning (ML):
Machine Learning relies on human-programmed algorithms to analyse data and build predictive models, such as recommending content on streaming platforms based on user behaviour. These systems require explicit coding for data processing and decision-making, with examples like fraud detection in finance or predictive maintenance in manufacturing. - Deep Learning (DL):
A subset of ML, Deep Learning uses neural networks to autonomously optimize algorithms through layered data analysis, enabling tasks like facial recognition in security systems. Unlike traditional ML, DL models self-adjust parameters during training, excelling at complex pattern recognition in applications like natural language processing or autonomous driving. - Technicality Threshold:
Under the European Patent Convention (EPC), purely abstract algorithms lack patentability unless applied to solve domain-specific technical challenges, such as medical image analysis for cancer detection. Patent eligibility demands a tangible technical effect, such as optimizing hardware efficiency in AI-driven robotics or industrial automation systems.
Patentability Requirements
To secure patents for AI inventions, applicants must overcome two hurdles:
- Technical Character: The invention👉 A novel method, process or product that is original and useful. must address a technical problem (e.g., optimizing GPU-CPU task allocation for neural networks).
- Technical Features: Claims must avoid abstract terms (e.g., “AI” or “neural network”) and instead focus on concrete implementations (e.g., “GPU-accelerated training sequence”).
Examples of Patentable AI Applications:
- Medical Imaging: Detecting skin cancer patterns in dermatological scans.
- Drone Surveillance: Identifying sharks in aerial footage using image classification.
- Hardware Optimization: Splitting ML tasks between GPUs and CPUs to enhance processing efficiency.
Strategic Claim Drafting
Karl emphasizes two pathways to patent AI innovations:
- Technical Applications: Protect domain-specific uses (e.g., drug discovery tools).
- Technical Implementations: Focus on novel hardware/software configurations (e.g., unique GPU training protocols)13.
Case Study: A method for generating random numbers becomes patentable when framed as a “computer-implemented method for numerical circuit simulation”.
Implementing Best Practices from digital IP lexicon 🔗dIPlex
The dIPlex AI-generated inventions topic page, authored by Dr. Laura Fe (chemistry/IP) and Dr. Malte Köllner (AI/IP law), provides actionable strategies for deep tech companies. Below, we analyze how their recommendations operationalize CEIPI’s principles:
Documenting Human-AI Collaboration
Challenge: Autonomous AI systems like DABUS blur inventorship lines, as seen in Thaler v. Vidal (U.S.) and EPO J 8/20 (EU), which reject non-human inventors.
Solution: Maintain granular records of human contributions (e.g., data curation, model tuning) to justify inventorship claims. For example, chemists guiding AI in molecular design must document their iterative input to satisfy EPO guidelines
Technical Anchoring in Domain-Specific Problems
Challenge: Pure AI Algorithms and the EPC’s Technicality Test
Pure AI algorithms, such as mathematical models for machine learning, often fail the European Patent Convention’s (EPC) technicality requirement because they are considered abstract mathematical methods under Article 52(2)(a) EPC. The European Patent Office (EPO) categorizes AI-driven innovations as computer-implemented inventions, which must demonstrate a technical effect or solve a technical problem to qualify for patentability. For example, a generic neural network trained to optimize data patterns lacks inherent technical character unless applied to a specific technical domain or hardware configuration.
Solutions: Framing Inventions Around Sector-Specific Challenges
To overcome this hurdle, patent claims must anchor AI innovations in concrete technical applications within defined industries. Below are sector-specific strategies:
- Chemistry: AI-Generated Drug Candidates
- Problem: AI models for drug discovery (e.g., predicting molecular interactions) risk👉 The probability of adverse outcomes due to uncertainty in future events. rejection if framed purely as computational methods.
- Solution:
- Disclose Training Datasets: Specify the sources and preprocessing of chemical data (e.g., public compound libraries like PubChem or proprietary datasets) to ensure reproducibility.
- Detail Model Architecture: Describe neural network layers, activation functions, and validation protocols (e.g., cross-referencing predictions with experimental toxicity assays).
- Example: A patent for AI-designed antiviral compounds might claim “a computer-implemented method using a convolutional neural network trained on SARS-CoV-2 protease structures to generate inhibitors with <90% binding affinity”.
- Materials Science: AI-Designed Alloys
- Problem: Patent claims for alloy optimization algorithms may be deemed non-technical without industrial context.
- Solution:
- Emphasize Industrial Applications: Link AI outputs to functional outcomes, such as “3D-printed titanium lattice structures for aerospace components with 50% higher strength-to-weight ratios than WE54 magnesium alloys”.
- Integrate Manufacturing Steps: Include technical implementation details, such as laser powder bed fusion parameters or post-processing heat treatments, to demonstrate practical utility.
- Example: RMIT University’s metamaterial patent highlights “GPU-accelerated topology optimization for titanium lattice designs resistant to 350°C temperatures in jet engine parts”.
- Why This Works
By grounding AI innovations in domain-specific technical problems, applicants satisfy the EPO’s requirement for a “specific technical purpose” (e.g., medical diagnostics or aerospace engineering). For instance:- In chemistry, disclosing datasets aligns with EPO enablement standards, ensuring reproducibility—a key criterion for patent validity.
- In materials science, linking AI to industrial manufacturing processes (e.g., 3D printing) demonstrates technical implementation, bypassing abstract algorithm objections.
This approach mirrors CEIPI’s guidance to “protect the application, not the AI” and aligns with Dr. Fe and Dr. Köllner’s 🔗dIPlex recommendations for deep tech firms.
The detailed article on the application of generative AI in drug development and material discovery and corresponding IP considerations by Dr. Laura Fè can you read 👉 here
Proactive IP Portfolio Management
- Audit AI Workflows: Conducting a thorough audit of AI workflows allows companies to identify distinct stages in their R&D process that may be eligible for patent protection. By breaking down complex AI systems into components like data preprocessing algorithms or novel hardware integration methods, organizations can build a more comprehensive and strategic patent portfolio.
- Leverage Trade Secrets: When certain aspects of AI development, such as proprietary datasets or specialized training methodologies, may not meet patentability criteria, companies can turn to trade secret protection as an alternative. This approach is particularly valuable for safeguarding unique assets like chemical reaction databases, which can provide a competitive edge but might not qualify for patent protection due to their nature as compilations of information.
Practice
Consider a deep tech startup using AI for chemical synthesis:
- Technical Application: Framing the invention as a “GPU-accelerated method for catalytic reaction optimization” effectively demonstrates the technical nature of the AI application in chemistry. This approach aligns with the EPO’s guidelines for patenting AI inventions by emphasizing the specific technical implementation and its practical use in a scientific field.
- Disclosure: Detailing training data sources, such as public compound libraries, and model architecture, like convolutional neural networks, is crucial for meeting patent disclosure requirements. This level of specificity not only satisfies the enablement criteria but also helps distinguish the invention from prior art and demonstrates its technical character.
- Filing Strategy: Prioritizing EU patents for technical implementations leverages the EPO’s relatively favourable stance on AI-related inventions with clear technical effects. Simultaneously, using trade secrets for proprietary datasets offers protection for valuable information that may not meet patentability criteria or could be more effectively safeguarded outside the patent system.
Conclusion
The CEIPI lecture provides a robust theoretical framework for patenting AI innovations, emphasizing technicality and precise claim drafting. Meanwhile, 🔗dIPlex’s best practices—documentation, domain-specific anchoring, and jurisdictional agility—offer a roadmap for implementation. For deep tech companies, merging these insights ensures compliance with evolving IP laws while securing competitive advantages in AI-driven markets.
