.png)
Building an AI system is more than creating algorithms, code and data. It also comes with responsibilities to ensure that no part of the system is created from stolen data, copied without permission, or could potentially cause significant harm if the system is hacked or used inappropriately.
Model Training and Optimization – Trade Secrets
After collecting data, companies use it to train AI models with proprietary methods they protect as trade secrets. Fine-tuning algorithms, hyperparameters, and neural network structures are key to optimizing these models, and the techniques used can be extremely valuable. To remain competitive, it's important to ensure ownership and protection of these valuable items.
Copyright Protection
AI software and its source code are protected by copyright laws. To ensure the company owns the copyright, they must make sure that developers sign contracts to assign copyright ownership to the company. Without these agreements, employees or contractors might claim ownership of all or parts of the copyright in the AI model.
Open-Source
Many AI projects rely on open-source libraries like TensorFlow, PyTorch, and Scikit-learn. However, if developers don't follow the licensing terms, the AI company can run into legal issues. Even if it is Open Source, there may be some conditions of use. Conditions in less permissive licenses may include limitation to research use, prohibited commercial use, requirement to make available (to the public) the complete source code of licensed works, including all modifications. Learn more about Open Source licenses: Licenses | Choose a License
Trade Secrets
Although software code is typically protected by copyright, keeping it secret is a very common practice. These are often kept confidential using technological ways to prevent copying the code and, to limit loss of a trade secret through a person, use Non Disclosure Agreements and strict access controls.
Mastering Go with Trade Secrets
DeepMind’s AlphaGo is an example of how trade secrets are kept. AlphaGo is an AI system that uses deep neural networks and advanced search techniques to master Go: a complex game of strategy and creativity. Its unique training process was kept as a trade secret, making it hard for competitors to copy its techniques.
Deep Genomics’ Secure Approach
Deep Genomics (Toronto) creates AI-driven drug discovery platforms and protects its machine-learning methods as trade secrets. By restricting access and using encryption, Deep Genomics stops competitors from copying its innovations.
Patentable AI Methods
High-value, technological breakthroughs that can be sourced (so that copycats can be detected) may be good candidates to protect with patent rights. Software code on its own isn’t patentable; it’s too abstract. However, a unique method where computer-implemented software code and data processing results in a technical effect or improvement may be patentable. Just remember that your patent application will be made available online, so patent isn’t a good option if you can’t prove that someone has copied you.
To qualify for patent protection inventions must be useful, novel and non-obvious. AI solutions usually phase challenges by the patent office if the invention basically describes computer instructions resulting in a computer doing what the computer is instructed to do (grouping text, compute numbers) – ideally the invention results in some sort of technical effect or outcome, often physical in nature. Around the world, IP professionals and patent offices are grappling with guidelines and evolving law for how AI can be patented. Head over to our Intellectual Property Deep Dive: A Beginner's Guide to Patenting Software and AI to learn more. Note: AI companies should check patent office requirements before applying for patents. For example, details can be studied at subject matter eligibility (USPTO) and Artificial intelligence and machine learning (EPO).
Protecting the functionalities and features of AI systems is key to maintaining a competitive edge and avoiding legal risks. From safeguarding trade secrets and securing copyrights to navigating open-source licenses and exploring patent options, each step must be taken with care. In the final part of our series, we’ll look at protecting the user interface and outputs, an essential layer of your AI IP strategy.
About the Author:
Allessia Chiappetta is a second-year JD candidate at Osgoode Hall Law School with a keen interest in intellectual property and technology law. She holds a Master of Socio-Legal Studies from York University, specializing in AI regulation.
Allessia works with Communitech’s ElevateIP initiative, advising inventors on the innovation and commercialization aspects of IP.
Allessia regularly writes on IP developments for the Ontario Bar Association and other platforms. Allessia is trilingual, speaking English, French, and Italian.
