An illustrated mountain landscape at dusk

The foundation of AI patent eligibility begins with 35 U.S.C. § 101, which appears deceptively simple: “Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.” This broad language seems welcoming to AI innovations, but decades of judicial interpretation have carved out three deadly exceptions that kill patents before they begin: abstract ideas, laws of nature, and natural phenomena. For AI inventors, the abstract ideas category is the silent assassin — courts routinely classify machine learning techniques as mathematical relationships or automated mental processes, making them prime targets for rejection.

The key isn’t whether your AI actually works or has commercial value. Courts don’t care about business success — rather, this inquiry involves whether your claims describe an abstract idea dressed up with computers or whether it is a technical solution to a technical problem.

The Alice/Mayo Two-Step Dance of Death

The current framework comes from two Supreme Court cases that fundamentally changed patent law: Alice Corp. v. CLS Bank (2014) and Mayo v. Prometheus (2012). Think of this as a two-step elimination process designed to weed out weak patents.

Step One asks the threshold question: “Is your claim directed to a judicial exception?”

For AI and computer-related patents, this means asking whether you’re essentially claiming an abstract idea. Neural network training? Abstract. Data processing and analysis? Abstract. Automated decision-making? You guessed it — abstract. The courts look at whether your claims focus on what your invention does (dangerous territory) versus how it accomplishes specific technical improvements (safer ground).

The $64,000 Question: Does your invention improve computer functionality or provide a technical solution to a technical problem that is integrated into a practical application? Or, on the other hand, does it consist of a straightforward application of known AI components and architectures?

Step Two becomes relevant only if Step One identifies the claim as being directed to a judicial exception, such as an abstract idea. At this stage of the Alice/Mayo framework, the question is whether the claim includes an “inventive concept” sufficient to transform the nature of the claim into a patent-eligible application. The USPTO spreads this inquiry across Step 2A Prong Two (is the exception integrated into a “practical application”?) and Step 2B (does the claim add “significantly more” than the exception?). See MPEP § 2106.

This is where most AI patents falter. Merely applying a conventional machine learning algorithm on general-purpose hardware is not enough — such implementations are routinely viewed as abstract ideas lacking any inventive concept. To survive, the claim must reflect a non-routine, non-generic technical solution. This can be demonstrated through a novel model architecture, an unconventional training method, or an improvement to the functioning of a computer system itself (as recognized under cases like Enfish and DDR Holdings). The focus must be on how your invention overcomes a technological problem in a specific and meaningful way — not just that it uses AI to automate a human task.

What Has Worked: Instructive Examples

The case law reveals some successful patterns. Enfish v. Microsoft (2016) succeeded because it improved database architecture — the computer itself worked better. McRO v. Bandai Namco (2016) survived because its automated lip-syncing used specific algorithmic rules that genuinely improved animation technology, not just automated existing processes.

But the failures are more instructive. Recentive Analytics v. Fox Corp. (2025) is particularly brutal for AI practitioners — the court found that applying generic machine learning to new domains constitutes an abstract idea lacking technical innovation. This ruling directly threatens LLM applications and broad AI implementations.

The Harsh Reality: Having a working AI system that makes money doesn’t matter. Courts rejected Electric Power Group’s data analysis system, Digitech’s data manipulation methods, and Yu v. Apple’s dual-camera image enhancement — all because they failed to show technical improvements to underlying computer systems.

The Language That Kills vs. The Language That Lives

Here’s where most AI patents commit suicide: claim language. Writing “a system for recommending content using AI” or “a method of analyzing text using a neural network” is essentially writing your own rejection letter. These functional descriptions trigger Step One and almost never survive Step Two.

The Golden Rule: Rather than focus the claim language toward what your AI accomplishes for users, focus it on the technical features that effect those results, with the backdrop that the claimed technical features should be directed toward a use-case-abstracted but technically-focused recitation of your technology, with regard to its technical achievements, such as, for example, improving computer processing efficiency.

The Generative AI Reality Check

For today’s generative AI innovations, Section 101 eligibility isn’t just a hurdle — it’s the primary barrier to patent protection. Most AI techniques get scrutinized under Step One as potential abstract ideas, and the recent case law trend is increasingly hostile. The courts are essentially saying: “We don’t care how clever your AI is or how good the results are. Show us how you made computers work better, not how you made computers do human-like tasks better.”

Success in this environment requires a fundamental mindset shift. Instead of thinking about your AI’s end-user benefits or business applications, you must identify and articulate the underlying technical mechanisms that represent genuine advances in computer science or system architecture.

The Uncomfortable Truth: Patent eligibility for AI has nothing to do with innovation in the traditional sense and everything to do with whether you can frame your innovation as a computer science advance rather than an application of computer science to solve human problems.

This legal framework forces AI inventors to dig deeper into their technical implementations, focusing on novel architectures, unconventional training methods, or system-level improvements that courts might recognize as patent-eligible subject matter. Without this technical grounding, even breakthrough AI innovations risk rejection as unpatentable abstract ideas, regardless of their revolutionary impact on society or industry.

2. The USPTO’s AI-Specific Patent Examination Guidance

Over the past three years, AI-related patent applications have surged by more than 30% year-over-year, driven largely by the generative AI revolution sparked by systems like GPT, LLaMA, Claude, and Gemini as well as agentic systems like n8n and ForGen AI’s AMCP. This explosion of filings — covering everything from large language models and retrieval-augmented generation to reinforcement learning agents and multi-modal AI systems — caught the patent office off guard. The problem? Most of these applications are algorithmically driven and data-centric, placing them squarely in the danger zone for Section 101 rejections.

The Hidden Crisis: Patent examiners were suddenly faced with evaluating probabilistic systems, black-box ML models, and non-deterministic outputs — technologies that traditional patent examination wasn’t designed to handle.

This created a perfect storm of inconsistent rejections, overly broad interpretations, and dramatically extended prosecution timelines. The USPTO’s response has been a series of guidance documents and internal reforms designed to bring order to AI patent chaos, but the results have been mixed at best.

The Guidance Documents That Shape AI Patent Fate

The foundation of current AI patent examination rests on the 2019 Revised Subject Matter Eligibility Guidance (RSMEG), which refined the Alice/Mayo framework into a more structured approach.

This guidance introduced the critical Step 2A Prong One & Two, and Step 2B analysis.

Step 2A:

  • Prong One: Does the claim recite a judicial exception (like an abstract idea)?
  • Prong Two: Is the exception integrated into a practical application?

Step 2B: Is there an inventive concept beyond the abstract idea?

The USPTO has since built on that framework with AI-specific material, including its July 2024 eligibility examples and an AI strategy published in January 2025. Those materials, along with examiner training and specialized art units, shape how AI applications are examined in practice, but individual examiners still apply them, and their interpretation can vary widely.

A Tale of Two Claims: Eligible vs. Ineligible

The USPTO’s own hypothetical examples show how fine the line can be. In Example 47 (anomaly detection), claim 2 recites training an artificial neural network, using it to detect anomalies, and outputting the results, and it is ineligible. Claim 3 uses a trained network to detect malicious network packets and then acts on them by dropping the packets and blocking future traffic in real time. It is eligible because, as a whole, it integrates the abstract idea into a practical application by improving network security.

Example 48 (speech separation) draws a similar line. Claim 1, which uses a deep neural network to analyze speech signals, is ineligible. Claim 2 recites the same judicial exception but is eligible because the claim as a whole improves speech-separation technology.

The Lesson: In each pair, the AI involved is similar. The difference is whether the claim as a whole reflects a technical improvement or practical application, rather than an abstract analysis with generic AI.

The “Practical Application” Trap

The USPTO’s “practical application” doctrine has become a minefield for AI patents. Claims must tie AI functionality to real-world technological contexts — but not just any context will do. Generic applications like “using a neural network to analyze text” or “applying ML to recommend products” fail consistently because they describe functional goals rather than technical implementations.

Winning claims look different: they describe specific model innovations, explain how architectural choices produce measurable technical benefits, and demonstrate implementation on particular hardware platforms or within novel system pipelines. The difference between “a transformer model that summarizes documents” (rejected) and “a transformer model configured with hybrid self-attention blocks that reduces memory footprint by 40% in edge computing environments” (potentially eligible) isn’t just semantic — it’s existential.

The Examiner Lottery Problem

Despite all the guidance documents and training programs, examiner inconsistency remains the biggest problem in AI patent prosecution. Applications get routed to specialized AI art units, and examiners receive ongoing education in deep learning concepts and generative models. But interpretation varies dramatically between individual examiners.

The Uncomfortable Reality: Some examiners treat LLMs as inherently abstract and unpatentable, while others allow patents for marginal model tweaks. Your patent’s fate often depends more on which examiner you draw than on your invention’s technical merit.

This inconsistency stems from the fundamental challenge of applying decades-old patent law to rapidly evolving AI technologies. Litigation trends and Federal Circuit appeals influence internal interpretation, making USPTO practice fluid and unpredictable. The recent Recentive Analytics decision, for example, will likely make examiners even more skeptical of AI applications in new domains.

Strategic Drafting in the USPTO Reality

Understanding this guidance landscape requires a complete rethinking of AI patent strategy. Avoid general-purpose functional language at all costs — claims like “a system that uses an LLM to generate summaries” are essentially rejection letters written by applicants themselves. Instead, winning claims look like: “A transformer model configured with hybrid self-attention blocks trained using loss-variant X to reduce hallucination rate by Y%.”

The key insight is linking AI function to measurable technical benefits. Claims should generate “reduced token-sequence representations to reduce memory footprint in real-time conversation AI” rather than just “better AI responses.” This isn’t just better claim drafting — it’s survival strategy in an examination environment that’s fundamentally hostile to functional AI applications.

The Strategic Imperative: Use the USPTO’s own examples as precedent within your specification. Explicitly contrast your invention with prior rejected examples, and map your claims directly to the eligibility framework that examiners must follow.

The Evolving Battlefield

The USPTO’s AI guidance represents an ongoing experiment in adapting patent examination to revolutionary technology. Examiners are learning AI concepts in real-time while trying to apply legal frameworks designed for mechanical inventions. The result is a system in transition, where understanding examiner expectations and drafting with surgical specificity becomes more important than the underlying innovation itself.

Success requires staying current with updates in AI art unit practice, Federal Circuit decisions, and internal USPTO policy shifts. But most importantly, it requires recognizing that AI patent prosecution isn’t about protecting AI innovations — it’s about translating AI innovations into language that survives a hostile examination process designed for a different technological era.

3. The Recent Recentive Case and Its Implications for AI Patent Eligibility

The Federal Circuit’s decision in Recentive Analytics, Inc. v. Fox Corp. (April 18, 2025) represents a seismic shift in AI patent law — and not in a good way for inventors. This case has become the poster child for everything wrong with current AI patent eligibility doctrine, establishing a precedent that makes generic AI applications virtually unpatentable. For anyone trying to patent AI innovations, Recentive isn’t just another case to cite — it’s a cautionary tale that could kill your patent before it’s even filed.

The patents covered two kinds of invention: “Machine Learning Training” patents for generating optimized schedules for live events, and “Network Map” patents for determining how live events are broadcast across television stations in different cities. Using iteratively trained models to adapt to real-time changes looks like the kind of application that should merit patent protection, particularly since the system solved real business problems.

The Fatal Flaw: The claims described what the machine learning accomplished (optimized schedules and network maps) rather than how a technical implementation advanced machine learning or computer systems.

The Claims That Doomed the Patents

The court described claim 1 of the ’367 patent as containing a collecting step (receiving event parameters and target features), an iterative training step for the machine learning model, an output step (generating an optimized schedule), and an updating step (detecting changes to the data inputs and iteratively generating new, further optimized schedules). The representative network map claim followed a similar structure of collecting, analyzing, updating, and using. This is a classic “data in → analysis → decision” framework that describes functional outcomes without technical specifics.

The specification made matters worse. It stated that the disclosed method uses generic computing equipment with “any suitable machine learning technique.” Recentive also acknowledged that the patents did not claim a machine learning technique itself, and that they did not claim a specific method for improving the mathematical algorithm or making machine learning better.

The Court’s Assessment: “We affirm because the patents are directed to the abstract idea of using a generic machine learning technique in a particular environment, with no inventive concept.”

How the Federal Circuit Ruled

Recentive sued Fox for infringement. The U.S. District Court for the District of Delaware granted Fox’s motion to dismiss, holding the patents ineligible under Alice, and the Federal Circuit affirmed. It framed the case as a question of first impression: “whether claims that do no more than apply established methods of machine learning to a new data environment are patent eligible. We hold that they are not.”

The court rejected Recentive’s main arguments. Iterative training and dynamic adjustment did not represent a technological improvement because they are “incident to the very nature of machine learning.” A new data environment did not make the claims eligible. And it was not enough that the methods “perform a task previously undertaken by humans with greater speed and efficiency.” The court also found nothing in the claims, individually or in combination, that would transform them into something significantly more than the abstract idea.

The court distinguished McRO and Koninklijke, the cases Recentive relied on. Rather than disclosing what Enfish called “a specific implementation of a solution to a problem in the software arts,” the claims here did not delineate steps through which the machine learning technology achieves an improvement.

The court was also careful about the limits of its decision, noting that machine learning “may lead to patent-eligible improvements in technology” and that “[t]oday, we hold only that patents 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.” (Read the opinion.)

The New Rules of AI Patent Survival

Recentive establishes brutal new standards for AI patent eligibility. The decision makes clear that domain-specific applications of generic ML are unlikely to survive Section 101 scrutiny, no matter how sophisticated or commercially successful they might be. This creates an almost impossible standard: you must demonstrate concrete technical contributions to AI itself, not just successful applications of existing AI.

The Recentive Standard: Your AI patent must advance machine learning technology, not just use machine learning technology. The difference between these two approaches determines patent life or death.

The case forces applicants to demonstrate three critical elements: concrete technical contributions to AI systems, non-obvious implementation details that go beyond standard practice, and clear linkage between the invention and measurable system-level technical benefits. Generic claims about “using neural networks” or “applying machine learning” are now essentially suicide notes.

Strategic Responses to a Hostile Landscape

In the post-Recentive world, claim drafting must completely avoid “black box” ML references. Claims must explain how models are constructed and function differently from existing approaches. Instead of claiming outcomes, successful patents must frame improvements in concrete system terms: latency reduction, model compression, pipeline optimization, or specific architectural innovations.

Specification-level responses become equally critical. Patents must describe training methodologies, datasets, model layers, loss functions, and unique pre/post-processing steps. Include benchmarks and empirical data whenever possible — courts want proof of technical improvement, not just assertions. The specification must clearly differentiate the invention from known ML applications in prior art, anticipating the examiner’s inevitable abstract idea rejection.

The New Drafting Reality: You can’t just say your AI works better — you must prove how your technical implementation makes computer systems work better, with specific metrics and architectural details.

The Generative AI Apocalypse

For generative AI patents, Recentive signals a shift toward much tighter scrutiny. Claims about generative capabilities, content personalization, or conversational AI must show significantly more than “AI applied to domain X.” Applications citing LLMs must include specific prompt architecture details, custom token processing methods, and concrete latency mitigation or training improvements.

This creates particular challenges for startups in the generative AI space. Filing early is no longer enough — companies need strong technical disclosures and carefully constructed claims from day one. Startups must budget significant time and expertise to build patentability into their engineering and R&D pipelines, recognizing that generic AI applications are now effectively unpatentable.

The Broader Implications for AI Innovation

Recentive represents more than just another unfavorable precedent — it fundamentally changes how AI companies must approach patent strategy. The decision forces a complete rethinking of what constitutes patentable AI innovation, shifting focus from successful applications to technical contributions to AI systems themselves.

The Uncomfortable Truth: Recentive suggests that most commercial AI applications — no matter how innovative or successful — cannot be patented unless they also advance the underlying AI technology.

This creates a perverse incentive structure where commercially valuable AI applications become unpatentable, while technically interesting but commercially irrelevant AI research might qualify for protection. The decision prioritizes computer science advancement over practical innovation, potentially discouraging the kind of applied AI research that drives real-world technological progress.

For practitioners, Recentive v. Fox has become the essential reference for Section 101 analysis in AI patents. It underscores that mere use of ML/AI is not patentable unless accompanied by concrete technical improvements to AI systems themselves. The case demands a fundamental shift in patent strategy: draft precisely, claim narrowly but technically, and build innovation around system-level advancement rather than functional outcomes. In the post-Recentive world, AI patent success requires proving you’ve advanced the science of AI, not just applied it successfully.

4. Strategic Claim Drafting for AI Patents in the Post-Recentive World

4.1. Drafting for Survival in the Post-Recentive Era

The Recentive decision fundamentally rewrote the rules of AI patent drafting. The Federal Circuit made it brutally clear that domain-specific use of ML is not patent-eligible without demonstrable technical innovation. This isn’t just another hurdle to clear — it’s a complete paradigm shift that places the burden squarely on applicants to demonstrate system-level contributions rather than just functional outcomes.

The New Reality: Courts don’t care if your AI solves important problems or generates millions in revenue. They only care whether your technical implementation advances computer science itself.

The strategic shift is profound: move away from “A system that uses AI to [accomplish X]” and toward “A model architecture configured to [technically improve Y] [under constraints Z] and comprised of: [its core AI-specific components].” This isn’t semantic wordplay — it’s the difference between patent life and death in the current legal environment.

4.2. Structuring Independent Claims for Eligibility

Successful AI claims must anchor themselves to technical solutions from the first word. Begin with concrete technical implementations: “A method comprising: constructing a neural network with layers A, B, and C…” rather than “A method of predicting user preferences using a machine learning model.” The former describes how your system works; the latter describes what it accomplishes.

Every claim must identify specific model improvements that go beyond generic AI application. Include attention module variations, loss function tuning for stability, or architecture adaptations to hardware constraints like token reduction for edge devices. These technical specifics transform abstract AI concepts into concrete technological contributions.

The Golden Rule: If your claim could describe any AI system with minor word substitutions, it’s too generic and will fail under Recentive.

Avoid vague or generic elements at all costs. Language like “using a machine learning algorithm” is a rejection waiting to happen. Instead, use technically precise language: “using a multi-head transformer encoder with context-aware gating to reduce inference latency by 30% on resource-constrained devices.”

4.3. Leveraging Dependent Claims to Add Technical Depth

Dependent claims provide crucial fallback positions when independent claims face challenges. Structure them around specific model design details: hyperparameters, layer configurations, novel embedding techniques, or custom attention mechanisms. These technical details demonstrate that your invention involves more than generic AI application.

Training techniques and deployment contexts offer another layer of protection. Distinguish your invention through distributed training mechanisms, online learning configurations, or hardware-accelerated inference pipelines. Each dependent claim should add technical specificity that reinforces your system-level contributions.

Application-specific optimizations can salvage claims even when independent claims falter. Show how your system improves latency in real-time systems, reduces energy use in mobile devices, or increases accuracy in constrained data environments. The key is linking functional improvements to technical mechanisms.

4.4. Claim Categories: Method, System, and Computer-Readable Medium

Method claims work best for showing sequence and logic flow, but they must be explicit in each step: data ingestion, model transformation, and result output. Avoid black-box descriptions — every step must reveal technical implementation details that distinguish your approach from generic AI processing.

System claims anchor your invention to infrastructure improvements, making it harder for examiners to characterize your invention as purely abstract. Frame these as “A computing system comprising memory and processor configured to execute a model using [specific technical approach]” rather than generic hardware descriptions.

Computer-readable medium claims emphasize the software-as-product angle but require exceptional specification support. The claims cannot feel “disembodied” — they must be backed by concrete technical disclosures that demonstrate how your software implements novel AI techniques.

4.5. Claim Language Tips and Pitfalls

Use terms with genuine technical weight: “self-attention mechanism,” “gradient clipping layer,” “asynchronous message bus.” These terms signal technical depth to examiners trained in AI concepts. Avoid marketing-speak like “AI engine,” “prediction logic,” or “intelligence module” — terms that scream abstract idea to experienced examiners.

Fatal Language Patterns: Avoid overly functional claims using terms like “configured to optimize,” “adapted to analyze,” or “suitable for summarizing.” These describe outcomes, not technical implementations.

Instead, use precise technical language: “Configured with residual vector fusion layer to reduce compression loss by 15%.” This language describes specific technical mechanisms and measurable improvements — exactly what post-Recentive examination demands.

4.6 The Disclosure–Trade Secret Balance: A Strategic Tightrope

Here’s the uncomfortable truth about AI patents: disclosure equals public domain. Once you disclose information in a patent application, even if not claimed, that information can be cited by examiners, used by competitors, and loses trade secret protection under U.S. and global law. The USPTO publishes applications approximately 18 months after filing, making the entire specification public unless filed under non-publication request.

The AI-Specific Risk: Proprietary training data pipelines, custom preprocessing workflows, model compression algorithms, performance tuning heuristics, and reinforcement learning reward shaping methods represent enormous competitive advantages that could be lost through patent disclosure.

Keep information as trade secrets when it’s: hard to reverse-engineer, not necessary for Section 112 compliance, provides operational advantages like performance gains or cost reduction, and can be used internally or licensed under NDA. Disclose information when it’s: the core innovation needed to show technical advancement, required for enablement, or essential for enforceability against competitors.

4.7. Final Thoughts

Drafting strong AI claims today requires hybrid thinking that integrates legal eligibility under Section 101, technical enablement under Section 112, and strategic IP layering to avoid over-disclosure. Your patent isn’t just legal protection — it’s a public roadmap to your innovation.

The best claim strategies integrate patent and trade secret protection, using each tool where it delivers maximum value. This means making hard choices about what to protect through patents (technical innovations that advance computer science) and what to protect through trade secrets (operational advantages and implementation details).

The Strategic Imperative: In the post-Recentive world, successful AI patent prosecution requires treating every application as both a legal document and a strategic business decision about competitive advantage and public disclosure.

The goal isn’t just obtaining patents — it’s building comprehensive IP strategies that protect core innovations while preserving competitive advantages. This requires understanding not just patent law, but the business realities of AI development and the strategic value of different types of intellectual property protection.

The battle to protect artificial intelligence innovation in 2025 is no longer just a legal exercise — it is a high-stakes strategic chess match played on a fast-evolving technological and legal board. The combined force of the Alice/Mayo framework, evolving USPTO examination practices, and the seismic impact of Recentive v. Fox has radically transformed the patent landscape for AI and generative technologies.

The message is clear: mere application of AI to real-world problems is not enough. Regardless of how commercially successful, socially valuable, or technically clever your invention may be, it will fail under Section 101 unless it clearly and convincingly demonstrates a concrete technological improvement to the functioning of computer systems or AI architecture itself.

Patent claims must be drafted with surgical precision, avoiding functional abstractions in favor of deep technical specificity. The invention must not only work but must advance how machines think, compute, or learn in measurable and non-obvious ways. This requires detailed disclosures of system architecture, model structures, training innovations, and performance benchmarks — all of which must be captured in the claim language itself, not just buried in the specification.

But this drive toward technical specificity brings its own trade-off: the public nature of patent disclosures can easily compromise valuable trade secrets. In a field like AI — where training data, performance heuristics, and system configurations can represent critical proprietary edge — over-disclosure can be as dangerous as underclaiming. The savviest inventors must now strategically balance what they reveal and what they protect. Patent what proves technical merit and is enforceable. Keep proprietary what grants sustained commercial advantage.

This new reality demands that AI startups, researchers, and IP counsel stop viewing patents as checkboxes for funding or acquisition. Instead, they must be treated as precision-engineered legal weapons — tailored to defend genuine technical contributions, ward off competitors, and strengthen market positioning. This requires interdisciplinary collaboration between legal teams, engineers, data scientists, and business strategists from the very first lines of code.

Ultimately, succeeding in the post-Recentive world isn’t just about meeting eligibility standards — it’s about aligning innovation, legal doctrine, and strategic protection into a unified intellectual property framework. The winners in this environment won’t just be those who invent better AI — but those who learn to describe, defend, and deploy that innovation in a language the law is willing to protect.

The path is narrow, but it is navigable — if you’re ready to think like a computer scientist, draft like a litigator, and protect IP like a strategist.

For the bigger picture, including eligibility, inventorship, and filing strategy, see our complete guide to patenting AI.

This article was originally published on our Medium blog on Jun 16, 2025.