Australia | Claiming AI & Machine Learning Inventions – Part 2: Beyond The Model

Part Two of our feature on patenting AI and machine learning inventions in Australia considers how to claim AI and ML inventions strategically once the technical implementation has been identified.

Part One From Model to Method considered patentable subject matter, including the importance of characterising the claimed invention as a whole and avoiding claims that present AI or ML merely as an abstract result, commercial scheme or generic computer implementation.

That discussion now reflects the current Australian position following Aristocrat Technologies Australia Pty Ltd v Commissioner of Patents [2025] FCAFC 131 (Aristocrat ’25), IP Australia’s response to stakeholder feedback, and the updated patent examination manual guidance for computer-implemented inventions.

Once the technical implementation has been identified, the next challenge is ensuring that the patent application supports claims that are enforceable, adaptable and commercially meaningful.

1. Identify the centre of technical gravity

The first drafting question is where the real value of the invention lies.

Is the contribution in how the model is trained? In how data is selected, labelled or transformed? In how inference is performed under real-time or resource-constrained conditions? In how the model is embedded in a device, sensor network, communications system, clinical workflow or industrial control process? Or in how the model output causes a downstream technical action?

Identifying the centre of gravity helps avoid overclaiming the invention as a generic AI result. It also helps ensure that the application supports the characterisation of the invention as a practical implementation, not merely a model output, prediction or business objective.

For AI and ML inventions, the most persuasive patent story is often not “we used AI to make a better decision.” It is “we implemented this AI system in this environment, using these data and processing steps, to produce this practical result.”

2. Support multiple claiming pathways

AI and ML inventions often provide value at more than one point in the system. A strong patent specification should preserve options by supporting different claim pathways where appropriate.

In practice, this means thinking beyond a single claim to “the model” and considering where protection may be needed across the full AI lifecycle, from training and data preparation through to deployment, inference, system integration and downstream action¹.

1. Training methods

If the invention lies in preparing, selecting, labelling or transforming training data, or in training under particular constraints such as distributed data, sparse data, privacy restrictions or low-quality inputs, those aspects should be disclosed in detail.

Training claims may be valuable where the inventive contribution lies before deployment, particularly in regulated fields or environments where data access is fragmented or sensitive.

2. Inference methods

If the invention lies in how a trained model is used operationally, inference claims may be important. This may include real-time decision-making, edge deployment, streaming inputs, adaptive thresholds, confidence-based processing, or triggering downstream technical actions.

3. System architecture

AI models rarely operate in isolation. If the invention involves integration with hardware, sensors, communications modules, medical devices, control systems, cloud platforms, mobile devices or feedback loops, those features should be described as part of the practical system.

This can be particularly important in Australia, where the characterisation of the claim as a whole may depend on how the claimed features interoperate².

4. Data flow and processing pipelines

In many AI systems, the value lies in how raw data is transformed before it reaches the model, or how model outputs are processed after inference. Compression, feature extraction, anonymisation, encoding, synchronisation, confidence filtering and feedback-based retraining may each provide important claim pathways.

5. Apparatus and system claims

Where the AI model is embedded in a configured product or technical environment, apparatus and system claims may help tie the invention to a practical implementation. Examples include diagnostic imaging devices, drones with onboard ML control, autonomous vehicles, agricultural sensor arrays, clinical monitoring platforms, cybersecurity gateways and industrial controllers.

6. Downstream action or real-world effect

In some inventions, the model output matters because it changes the operation of another system. The output may control a device, adjust a process parameter, allocate network resources, trigger authentication, change image acquisition, configure a communication channel, or initiate a safety response.

Where that downstream action is central to the invention, it should be disclosed and claimed where possible.

This approach preserves flexibility across jurisdictions and can improve enforceability by tying the claims to the features that drive the practical result.

3. Build in support for practical result

For AI and ML inventions, it is valuable to explain not only the model architecture or algorithm, but the practical result produced by the configured system.

The focus should be on practical outcomes, not abstract aspirations.

This does not always require production-scale experimental results. However, where the invention improves speed, efficiency, reliability, accuracy, latency, power consumption, bandwidth use, image quality, control stability, privacy preservation or security, the patent specification should explain why those benefits arise from the claimed implementation.

The key is to connect the result to the mechanism³. Unsupported statements that a model is “more accurate” or “more efficient” are less useful than a description of the particular features that produce that improvement.

Examples may include:

• reduced latency where the specification describes a local inference process on an edge device using compressed features
• improved privacy where the specification describes a federated training process that maintains sensitive datasets within separate secure environments
• improved imaging quality where the specification describes the image-acquisition or reconstruction pipeline in which model-based noise reduction is applied
• improved equipment control where the specification describes how real-time sensor streams are processed to adjust operating parameters
• improved cybersecurity response where the specification describes how a model output is used within a network gateway to initiate secondary authentication.

These details assist not only with patentability, but also with support, clarity, sufficiency and enforcement.

4. Draft for characterisation

In view of Aristocrat ’25, IP Australia’s stakeholder feedback response, and the updated Manual guidance, claim characterisation is likely to remain an important battleground in examination.

That makes it important to identify, early, how the invention is likely to be characterised by the Patent Office. If the application presents the AI system only as a business idea, mathematical model or information-processing result, it may be difficult to defend later by pointing to generic computer implementation⁴.

For that reason, when preparing AI and ML patent applications, focus is to be placed on capturing the features that give the invention its practical character so it is not easily reduced to a business idea, mathematical model or information-processing result by an Examiner. That also does not mean adding routine computer steps for their own sake. Those features assist only where they form part of the practical operation of the claimed combination⁵.

For example, rather than claiming only that an AI model generates a risk score, a stronger claim may define the relevant data source, preprocessing steps, model operation, deployment environment and system response triggered by the score. However, those features should not be treated as a checklist. If they are merely routine or generic, they may not materially improve the patentability position. They are most useful where they explain how the claimed invention operates in a particular practical context.

Similarly, rather than claiming only that a model recommends an action, a stronger claim may define how the recommendation changes operation of a device, network, imaging system, monitoring platform or control process.

The goal is to make the interoperation of the claim features part of the invention’s substance, rather than relying on generic computer implementation to carry an otherwise abstract idea.

5. Thinking globally

AI and machine learning inventions often have international relevance, but the patentability standards are not harmonised.

For example, the European practice places strong emphasis on technical character and technical effect. In many cases, applications drafted with European-style technical detail will provide a strong foundation for Australian prosecution.

United States practice can permit broader functional claiming in some circumstances, but subject matter eligibility and written description issues can still arise, particularly where claims are framed at a high level of abstraction.

Therefore, for international filings, the safest approach is to describe the invention in a way that supports multiple jurisdictional arguments: what the model does, how it is trained, how it is deployed, how the data flows, how the system components interoperate, and what practical result is achieved.

Pulling it together | Securing meaningful protection for AI innovation

For businesses and research teams developing AI and ML technologies, patents are not just legal instruments. They are tools for protecting commercial advantage, investment and technical differentiation.

However, AI inventions often raise difficult questions because the value may sit between a model, a dataset, a software workflow and a real-world system. If the application describes only the model output or commercial objective, the invention may be vulnerable. If it explains how the model is implemented, how the system components interact, and what practical result is produced, the prospects are stronger.

Following Aristocrat ’25, IP Australia’s stakeholder feedback response and the updated patent examination manual guidance, the position in Australia is helpful, but not a free pass. Conventional computer implementation is not automatically fatal, and known technical features should not be ignored when they are part of the claimed combination. At the same time, an abstract idea, business method, prediction or information scheme does not become patentable merely because AI is used.

The drafting task is therefore to move from model to method: from an AI concept to a claimed invention that can be characterised, as a whole, as producing an artificial state of affairs and a useful result.

How can we help

At Spruson & Ferguson, our patent attorneys work with clients across sectors where AI and machine learning are driving innovation, including medical devices, diagnostics, therapeutics, pharmaceuticals, the broader life sciences sector, industrial automation, communications, cyber security and autonomous systems.

Footnotes

1 Aristocrat Technologies Australia Pty Ltd v Commissioner of Patents [2025] FCAFC 131 at [127], [136].

2 See Aristocrat [2025] FCAFC 131 at [127], [131], [136].

3 See Aristocrat [2025] FCAFC 131 at [132].

4 See Aristocrat [2025] FCAFC 131 at [126],[131].

5 See Aristocrat [2025] FCAFC 131 at [136].