Whitepaper: Updated Strategies for Patenting AI Subject Matter

James Denaro / CipherLaw

On April 18, 2025, the Federal Circuit issued a significant ruling in Recentive Analytics, Inc. v. Fox Corp. that establishes clear boundaries for AI patentability. This decision represents the first time the Federal Circuit has directly addressed whether claims that merely apply established methods of machine learning to new data environments are patent eligible under 35 U.S.C. § 101. The court's unambiguous answer: they are not.

The court's holding is clear: "[P]atents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101."

The Case and Its Holdings

In Recentive Analytics, the Federal Circuit affirmed a district court's dismissal of patent infringement claims, concluding that patents directed to using generic machine learning techniques to generate “network maps” (which determine content displayed by a broadcaster’s channels within certain geographic markets at particular times) and event schedules for television broadcasts were ineligible under § 101. The court found that the patents were directed to the abstract idea of "using a generic machine learning technique in a particular environment, with no inventive concept."

Judge Dyk, writing for the court, emphasized that the patents employed conventional machine learning technology, as evident from the specifications themselves, which described using "any suitable machine learning technique" including gradient boosted random forests, regression, neural networks, decision trees, support vector machines, and Bayesian networks. The patents failed to disclose any technical improvements to these machine learning approaches.

Most critically, the court emphasized that the patents failed to "delineate steps through which the machine learning technology achieves an improvement" or "disclose a specific implementation of a solution to a problem in the software arts." This directly supports the need for both a problem-solution framework and detailed technical descriptions of improvements, as discussed in more detail below.

Other Supporting Federal Circuit Cases

Several other Federal Circuit cases reinforce these principles:

Enfish, LLC v. Microsoft Corp. (2016): The court found claims eligible because the specification included a specific technical problem (inefficiencies in conventional databases) and detailed how the self-referential table structure provided specific improvements.

McRO, Inc. v. Bandai Namco Games America Inc. (2016): The court emphasized the importance of the specification explaining how the claimed rules enabled automation of a task that previously could not be automated.

Koninklijke KPN N.V. v. Gemalto M2M GmbH (2019): The court found claims eligible where the specification detailed specific technical problems with error detection and how the claimed permutation technique provided a specific solution.

Key Takeaways for AI Patent Eligibility

The Federal Circuit's decision provides several important guidelines for AI patent eligibility:

1. Generic Applications Are Not Eligible: Patents that do no more than apply generic machine learning to new data environments, without disclosing improvements to the machine learning models, are ineligible under § 101.

2. Field of Use Restrictions Don't Create Eligibility: The court rejected the argument that applying machine learning to a new field (event scheduling and network mapping) transformed an abstract idea into patent-eligible subject matter, citing established precedent that "[a]n abstract idea does not become nonabstract by limiting the invention to a particular field of use or technological environment."

3. Efficiency Improvements Alone Are Insufficient: The court held that methods are not rendered patent-eligible merely because they perform tasks previously done by humans with greater speed and efficiency, consistent with its computer-assisted methods jurisprudence.

4. Lack of Technical Specificity Is Fatal: The patents failed to "delineate steps through which the machine learning technology achieves an improvement" or "disclose a specific implementation of a solution to a problem in the software arts."

Alignment with USPTO's Examples 47-49 from July 2024 Guidance

The Federal Circuit's decision aligns with the USPTO's July 2024 guidance update on AI patent eligibility. The USPTO's July 2024 guidance update on Patent Subject Matter Eligibility focused specifically on artificial intelligence inventions in response to Executive Order 14110 on the "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence." The guidance maintained the fundamental Alice/Mayo framework for patent eligibility under 35 U.S.C. § 101, but provided clearer instructions on how to apply this framework to AI-related inventions. Most significantly, it introduced three new hypothetical examples (numbered 47-49) that illustrate the application of eligibility analysis to AI technologies. We revisit those specific examples in light of the Recentive Analytics decision.

Example 47: Anomaly Detection using Neural Networks

Example 47 from the USPTO guidance addresses anomaly detection using artificial neural networks, illustrating when hardware implementations can be patent-eligible:

Claim 1 (Eligible):

1,. An application specific integrated circuit (ASIC) for an artificial neural network (ANN), the ASIC comprising:

a plurality of neurons organized in an array, wherein each neuron comprises a register, a microprocessor, and at least one input; and

a plurality of synaptic circuits, each synaptic circuit including a memory for storing a synaptic weight, wherein each neuron is connected to at least one other neuron via one of the plurality of synaptic circuits.

This claim is eligible because it recites specific hardware architecture for implementing the neural network rather than merely applying a generic neural network to a problem. The claim features detailed technical elements (registers, microprocessors, synaptic circuits with memory) that represent a technological implementation.

In contrast, Claim 2 (ineligible) merely describes a generic machine learning process without technical specificity:

2. A method of using an artificial neural network (ANN) comprising:

(a) receiving, at a computer, continuous training data;

(b) discretizing, by the computer, the continuous training data to generate input data;

(c) training, by the computer, the ANN based on the input data and a selected training algorithm to generate a trained ANN, wherein the selected training algorithm includes a backpropagation algorithm and a gradient descent algorithm;

(d) detecting one or more anomalies in a data set using the trained ANN;

(e) analyzing the one or more detected anomalies using the trained ANN to generate anomaly data; and

(f) outputting the anomaly data from the trained ANN.

Example 48: Speech Separation

Example 48 demonstrates that applications involving real-world technological improvements can be eligible:

Claim 1 (Ineligible):

1. A speech separation method comprising:

(a) receiving a mixed speech signal x comprising speech from multiple different sources sn, where n ∈ {1, . . . N};

(b) converting the mixed speech signal x into a spectrogram in a time-frequency domain using a short time Fourier transform and obtaining feature representation X, wherein X corresponds to the spectrogram of the mixed speech signal x and temporal features extracted from the mixed speech signal x; and

(c) using a deep neural network (DNN) to determine embedding vectors V using the formula V = fθ(X), where fθ(X) is a global function of the mixed speech signal x.

This claim merely applies mathematical operations without specific technical improvements. However, Claim 2 (eligible) shows how adding specific technical implementations can transform an abstract idea into eligible subject matter:

2. The speech separation method of claim 1 further comprising:

(d) partitioning the embedding vectors V into clusters corresponding to the different sources sn;

(e) applying binary masks to the clusters to create masked clusters;

(f) synthesizing speech waveforms from the masked clusters, wherein each speech waveform corresponds to a different source sn;

(g) combining the speech waveforms to generate a mixed speech signal x' by stitching together the speech waveforms corresponding to the different sources sn, excluding the speech waveform from a target source ss such that the mixed speech signal x' includes speech waveforms from the different sources sn and excludes the speech waveform from the target source ss; and

(h) transmitting the mixed speech signal x' for storage to a remote location.

Example 49: Fibrosis Treatment

Example 49 illustrates how AI applications in healthcare can include specific treatment protocols to qualify for eligibility:

Claim 1 (Ineligible):

1. A post-surgical fibrosis treatment method comprising:

(a) collecting and genotyping a sample from a glaucoma patient to a provide a genotype dataset;

(b) identifying the glaucoma patient as at high risk of post-implantation inflammation (PI) based on a weighted polygenic risk score that is generated from informative single-nucleotide polymorphisms (SNPs) in the genotype dataset by an ezAI model that uses multiplication to weight corresponding alleles in the dataset by their effect sizes and addition to sum the weighted values to provide the score; and

(c) administering an appropriate treatment to the glaucoma patient at high risk of PI after microstent implant surgery.

This claim merely suggests administering a treatment without specificity. In contrast, Claim 2 (eligible) specifies a particular treatment:

2. The method of claim 1, wherein the appropriate treatment is Compound X eye drops.

Practical Implications for AI Patent Applicants

In light of Recentive Analytics and the USPTO examples, patent applicants should focus on:

Technical Improvements to AI Models

Claims should articulate specific technical improvements to the machine learning models themselves, not just benefits from applying existing models to new data. The Federal Circuit clearly rejected patents that merely apply generic machine learning to new data environments without disclosing improvements to the models themselves.

Specification Drafting Strategies

The specification plays a crucial role in establishing eligibility and should include:

• Problem-Solution Framework: Clearly identify specific technical challenges in existing technologies and explain precisely how the AI implementation overcomes these limitations, rather than providing general narrative prose about benefits.

• Detailed Technical Description of Improvement: Thoroughly document how the AI innovation provides technical improvements over conventional approaches. The patents in Recentive Analytics failed specifically because they did not "delineate steps through which the machine learning technology achieves an improvement."

• Implementation Details: Provide substantial technical descriptions of the AI architecture, data preprocessing techniques, training methodologies, and hardware configurations. Avoid purely functional language that merely describes what the AI does without explaining how it achieves its results.

• Alternative Implementations: Include multiple embodiments and alternative implementations to strengthen the application and demonstrate that you're claiming a technical solution rather than an abstract idea.

• Domain-Specific Integration: For domain-specific applications, explain how the AI technology integrates with and improves that specific field, documenting adaptation to domain constraints and technical integration points.

Specific examples of specification drafting strategies that might be written into AI patent specifications are provided below. These are concrete examples that could help strengthen patent eligibility under the current framework.

Example 1: Problem-Solution Framework

[0012] Conventional machine learning approaches for speech separation typically rely on generic deep neural networks (DNNs) that perform adequately in controlled environments but suffer significant degradation in real-world scenarios with multiple overlapping speakers and background noise. Specifically, existing systems experience a 45-60% reduction in speech recognition accuracy when the signal-to-noise ratio drops below 10dB, making them impractical for many commercial applications.

[0013] The technical problem arises from the inability of conventional DNN architectures to effectively distinguish between temporally overlapping speech signals with similar frequency characteristics. Standard DNNs process the entire spectrogram globally, failing to capture the fine-grained local temporal-spectral relationships that differentiate overlapping speakers.

[0014] The present invention addresses this technical problem through a novel neural network architecture that implements hierarchical attention mechanisms operating at multiple temporal and spectral scales simultaneously. Unlike conventional approaches that process the entire spectrogram as a single unit, the disclosed architecture employs a cascade of attention layers that progressively refine the focus from coarse-grained speaker identification to fine-grained phonetic separation.

Example 2: Detailed Technical Description of Improvement

[0025] The improved neural network architecture implements three key technical innovations that enable superior performance over conventional systems:

[0026] First, the dual-pathway attention mechanism processes spectral and temporal information through separate but interconnected channels before merging them, allowing the model to capture cross-domain dependencies that are lost in traditional single-pathway networks. Figure 2 illustrates the detailed structure of this dual-pathway architecture, showing how information flows through parallel spectral and temporal attention modules before merging at the integration layer.

[0027] Second, the adaptive granularity control module dynamically adjusts the receptive field size of each attention head based on signal characteristics. In regions with high speaker overlap, the system automatically decreases the receptive field to focus on finer details, while expanding it in regions with less overlap to capture broader contextual information. This adaptive mechanism is implemented through a feedback loop from the overlap detection module to the attention scaling parameters, as detailed in Equations 4-7.

[0028] Third, the residual embedding refinement technique iteratively refines the embedding vectors through multiple processing stages, with each stage receiving both the original input and the output from the previous stage. This approach preserves information that might otherwise be lost in deep networks and enables more effective separation of similar voices. Figure 3 shows the performance improvement achieved with each additional refinement stage, demonstrating significant gains up to four stages before diminishing returns.

Example 3: Implementation Details

[0042] The custom attention mechanism is implemented as follows:

[0043] Input to the attention mechanism is the time-frequency representation X of dimensions T×F, where T represents time frames and F represents frequency bins.

[0044] The time-frequency representation X is first processed by three separate 1×1 convolutional layers to generate query (Q), key (K), and value (V) representations, each with dimensions T×F×D, where D is the embedding dimension.

[0045] Unlike conventional attention mechanisms that compute attention scores between all time-frequency bins, our frequency-selective attention mechanism restricts attention computation to frequency regions that are likely to contain the target speaker's voice.

[0046] Specifically, a frequency importance estimator network FIE(X) outputs a weight vector w of dimension F, where each element represents the importance of the corresponding frequency bin for the target separation task.

[0047] The attention matrix A is then computed as:

A = softmax((Q·K^T)⊙M)/√D

[0048] Where M is a mask derived from the frequency importance weights w such that M_ij = sigmoid(w_i + w_j - θ), and θ is a learnable threshold parameter.

[0049] This selective attention mechanism reduces computational complexity from O(T²F²) to approximately O(T²Fk), where k is the average number of relevant frequency bins, significantly improving both efficiency and separation performance by focusing computational resources on the most informative spectral regions.

Example 4: Alternative Implementations

[0078] While the primary embodiment utilizes the hierarchical attention architecture described above, alternative embodiments may implement variations of the core technology to address specific deployment scenarios:

[0079] In a first alternative embodiment, optimized for edge devices with limited computational resources, the full attention mechanism is replaced with a lightweight approximation using separable convolutions and grouped attention. This implementation reduces memory requirements by 78% and computational complexity by 65% while maintaining separation quality within 5% of the full implementation. This embodiment is particularly suitable for mobile and IoT applications.

[0080] In a second alternative embodiment designed for server-side processing of multiple simultaneous streams, the architecture is modified to process batches of overlapping speech signals in parallel, sharing early-layer representations across streams to improve computational efficiency. This parallel processing capability enables the system to scale linearly with the number of concurrent separation tasks up to hardware memory limits.

[0081] In a third alternative embodiment adapted for real-time processing, the architecture incorporates causal attention mechanisms that only attend to current and past time frames, eliminating the need to buffer future frames and enabling streaming operation with latency under 50ms. While this configuration shows a minor reduction in separation quality (approximately 7% relative to the non-causal system), it enables applications requiring immediate feedback such as live communication systems.

Example 5: Domain-Specific Integration (Healthcare Example)

[0092] The integration of the AI system with existing clinical workflows occurs through three specific technical interfaces:

[0093] First, the system implements a FHIR-compliant (Fast Healthcare Interoperability Resources) data integration layer that securely accesses patient electronic health records to extract relevant clinical parameters, including previous medication history, contraindications, comorbidities, and genetic test results. This integration layer maintains full HIPAA compliance through end-to-end encryption and role-based access controls.

[0094] Second, the treatment recommendation engine incorporates domain-specific constraints derived from established clinical guidelines. Specifically, the system encodes the complete set of contraindications and drug interaction rules from the American Academy of Ophthalmology's Post-Surgical Inflammation Management Guidelines (2024 edition), implementing them as hard constraints within the recommendation algorithm rather than learned parameters. This approach ensures that all recommendations strictly adhere to established medical protocols regardless of patterns in the training data.

[0095] Third, the system features a clinician feedback mechanism that allows ophthalmologists to provide real-time corrections to recommendations, with these corrections used to continuously refine the model through a domain-constrained reinforcement learning process. This feedback loop operates on a dual-timescale: immediate updates to the decision boundary for the specific patient case, and periodic batch updates to the global model after validation by a medical advisory board.

These examples illustrate how to move beyond generic claims of using AI to solve problems by instead documenting specific technical improvements, implementation details, and integration approaches that demonstrate patent-eligible innovations in the AI itself rather than merely applying known AI techniques to new domains.

Hardware Integration

Where possible, claims should focus on specific hardware implementations or technical steps that go beyond merely applying AI to a problem. Example 47 demonstrates that specific hardware architecture for neural networks can establish eligibility.

Avoiding Field-of-Use Limitations Alone

The court explicitly rejected the argument that applying machine learning to a new field transforms an abstract idea into patent-eligible subject matter. Technical specificity, not novel application domains, is the key to eligibility.

Specific Treatments in Healthcare Applications

For medical AI technologies, specify particular treatments or methods to transform abstract data analysis into a patent-eligible invention. Example 49 illustrates that merely suggesting "administering an appropriate treatment" is insufficient, while specifying "Compound X eye drops" can establish eligibility.The Future of AI Patentability

The Future of AI Patentability

The Federal Circuit recognized that "machine learning is a burgeoning and increasingly important field and may lead to patent-eligible improvements in technology." The court limited its holding to patents that apply generic machine learning to new data environments without disclosing improvements to the machine learning models.

This decision establishes a clear framework for AI patentability that rewards genuine technological innovation while preventing overly broad patents on the mere application of known machine learning techniques to new fields. Applicants need carefully consider these guidelines when drafting and prosecuting AI-related patent applications to ensure they meet the increasingly well-defined eligibility standards.