In the previous part, we left with a question that does not settle easily:
If AI inventions are difficult to define, how do we even begin to protect them? Before one attempts to draft, argue, or claim, there is something more fundamental that must be done i.e., the invention must be understood in its true structure.
To understand what structure truly is, one must first strip away what it is not. Not the convenient abstraction. Not the polished summary in a pitch deck. Not the easy sentence “an AI system that does X” which hides more than it reveals.
Structure is not the outcome. It is not the interface. It’s not even the “intelligence” we often talk about. It’s what sits underneath. A simple arrangement of parts put together in such a way that the result naturally follows. Because in AI, the invention is rarely one thing. It is often a composition. And unless this composition is broken down clearly, the invention itself remains blurred.
Why Structure?
One may ask why structure is important when drafting or understanding AI inventions. Does it really matter, as long as all the information is present? More often than not, anything that involves a deep and non-trivial technical understanding must be structured. Only then can the reader or in this context, the examiner, the Controller, or even the Judges who decide and grant patents properly comprehend the invention.
In traditional inventions, a machine has parts, a circuit has components, and a process has steps. The structure is tangible, visible, and easier to map. In AI inventions, however, the structure is not always physical; it is conceptual, layered, and interdependent.
This is where the difficulty begins.
An applicant or inventor may simply state:
“We have developed an AI model that predicts fraud.”
But the real question arises, where does the invention actually lie? Is it in the model architecture? Is it in the way the data is processed? Is it in the training method? Or is it in how the output is ultimately used?
Without answering this, novelty becomes difficult to demonstrate, and inventiveness becomes even harder to argue.
This is precisely why breaking down the AI system becomes essential.
The Four Core Elements of an AI Invention
At a practical level, most AI inventions can be understood through four basic elements:
Data → Model → Training → Application
These are not merely technical stages. They are points where invention may exist. Let us look at each of them carefully.
Data: The Starting Point of Intelligence
Data is the essential "fuel" and foundation of artificial intelligence (AI), enabling models to learn, make predictions, and perform tasks through patterns identified in vast datasets. The role of data extends beyond just feeding models; it is also crucial for data validation and quality control to ensure accuracy
Every AI system begins with data. But not all data is equal and may not necessarily be part of the novelty or the inventive steps. The invention may lie in:
- How the data is collected;
- How it is structured or labelled;
- How irrelevant or noisy data is filtered; or
- How multiple data sources are combined.
Therefore, in many cases, what appears to be a simple dataset may actually involve a highly non-trivial process of preparation. Since the efficiency of AI systems is generally dependent on the data more particularly on clean data, or datasets that are noise-free and bias-minimized the manner in which such data is obtained and refined becomes critical.
For example, a system that generates synthetic training data, a method that improves data quality for better prediction, or a technique that reduces bias in datasets are not merely pre-processing steps. They may carry independent inventive value and can form a core part of the invention.
Model
An AI model is a program trained on vast datasets to recognize patterns, make decisions, or generate new content without constant human intervention.
The model is what most people instinctively associate with AI. Neural networks, decision trees, transformers etc. these are the visible engines of learning. However, from a patent perspective, merely stating a known model is not enough. The invention must pass the muster of the technical effect as doctrine…?.
But in the case of the AI models, the invention may lie in:
- A modification to an existing architecture;
- A combination of multiple models (subject to the technical effect test);
- A new way of structuring layers or parameters; or
- A constraint or optimization applied within the model.
But caution is necessary here. If the model is claimed purely as a mathematical construct, it risks falling within exclusions relating to algorithms or mathematical methods. This is where the reasoning adopted in Ferid Allani v. Union of India becomes relevant. The focus must not remain on the abstract model, but on what the model achieves in a technical sense.
Training
AI model training is the iterative process of teaching a system to recognize patterns in data so it can make accurate predictions or decisions.
Training is often the most underappreciated part of an AI invention. Yet, in many cases, this is where the real innovation lies.
For example, a new loss function that improves accuracy, a training pipeline that reduces computation time, a method that allows learning from limited data, a technique that stabilizes training in complex environments etc.
These are not superficial improvements. They directly affect performance, efficiency, and reliability factors that are often crucial in demonstrating technical effect which the Patent Offices expect from the inventions.
Therefore, training is not just about feeding data into a model. It is about how learning is shaped. And shaping learning can itself be an invention if a technical effect could be shown in such inventions.
Application
The user-facing component, such as chatbots or AI assistants. This is where the AI system interacts with reality. It is also where patentability often becomes clearer. Because here, the questions shift from:
“What is the model doing?”
to
“What problem is being solved?”
Therefore, in this layer, the invention may lie in:
- How the AI output is used in a system;
- How it controls or improves a device;
- How it enhances efficiency in a process; or
- How it produces a tangible result.
For example, AI controlling energy consumption in a smart grid, AI improving image processing in a camera system, AI optimizing network traffic in telecommunications and so on.
At this stage, the invention begins to move away from abstraction and towards technical application. And this is where patent law becomes more receptive.
Where Does the Invention Really Lie?
The question is still unclear where exactly does the invention lie? While we understand that it may reside in the data, the model, the training, or the application, which of these elements actually contains the invention?
The answer is not fixed. It varies from one invention to another, as we often say in law, “it depends on the facts and circumstances of the case.” At the outset, however, it must be recognized that the invention may sometimes lie in one of these four layers, sometimes in a combination of them, and sometimes in the intersection between two or more layers.
This is where careful analysis becomes critical. Claiming everything may render the invention vague; claiming the wrong aspect may lead to rejection; and overlooking the true core of the invention may weaken the application altogether.
Therefore, the role of the attorney becomes crucial to identify, isolate, and draft the invention in a manner that captures its true essence, while ensuring that all essential aspects are properly defined, supported, and protected.
Conclusion:
An AI invention is not a single entity. It is a layered system, where data feeds learning, models shape understanding, training refines intelligence, and applications deliver results. Within this layered structure, the invention may lie sometimes in plain sight, and sometimes in unexpected places.
Understanding this structure is not merely a technical exercise. It is the first step towards identifying novelty, demonstrating inventiveness, and ultimately securing protection.
In the next part of this series we shall understand how do we present it in a way that satisfies patent law?
*The author is a practicing advocate and IP attorney who regularly advises on intellectual property and technology matters. The views expressed in this article are personal and are intended for informational purposes only, and should not be construed as legal advice.