If you’re building or investing in AI, here’s what you need to know:
When it comes to IP in enterprise AI models, everyone wants to race to the discussion of who is going to own the results. But before we talk about who owns what in AI collaborations, we need to start with some basic concepts.
What is an AI model?
At its core, it’s software code executing functions, to transform input data (like an image or prompt) into output data (like a label or generated video).
The functions rely on parameters – think of a giant panel of control knobs. In an untrained model, these are random or set to zero, producing meaningless results. Training adjusts those knobs, tuning the model to generate useful outputs.
The trained model is the result of combining code, methods and data – each of which is a form of intellectual property.
IP in AI Models Falls Into Three Buckets:
1. Methods – i.e. design and execution strategies.
Examples: ML techniques, optimization strategies, algorithms, business process knowledge, novel use cases.
2. Code – i.e. executable instructions that implement the methods to transform the data.
Examples: Open-source libraries, model architectures, proprietary tools, system integrations, data pipelines.
3. Data – plays a role in every aspect of AI, such as:
- Training data sets parameters
- Parameters become derived data
- Inference data includes inputs and outputs
- Metadata stores configuration
Examples: Public datasets, proprietary records, model outputs, learned parameters, derived insights.
It’s All Interconnected
Methods shape the code.
Code uses the data.
Data affects both.
This overlap matters – especially when it comes to collaboration, ownership and strategy.
Once you understand the buckets of potential IP in AI model development, then you can have a more intelligent discussion of who-owns-what.
We’ll dive into that interesting subject in my next post…