Can You Patent an AI Algorithm? A Deep Dive into 2025 Regulations

The Engineering Decision That Changed How AI Gets Patented

In 2012, a small team at Google Brain made a specific choice that would later define the front line of AI patent strategy. Training a convolutional neural network on ImageNet, they did not patent what the network predicted — top-5 image classification accuracy. They patented the specific architectural constraints and normalization sequences that made that prediction possible at scale: the dropout regularization schedule, the GPU parallelization scheme, the data augmentation pipeline. That distinction — between a model's emergent capability and the deterministic engineering scaffolding that shaped it — is the central tension every AI founder faces when asking whether their algorithm can be patented in 2025.

The answer is: yes, but almost certainly not the part you think.

Why the Alice Decision Still Governs Everything in 2025

The 2014 Supreme Court ruling in Alice Corp. v. CLS Bank International remains the controlling framework for software and algorithm patent eligibility under 35 U.S.C. §101. Alice established a two-step test that USPTO examiners apply to every AI-related application. Step 2A asks whether the claim is directed to an abstract idea — mathematical concepts, mental processes, or certain methods of organizing human activity. Step 2B asks whether the claim adds "significantly more" than that abstract idea: a concrete, practical improvement to a specific technical field, not merely applying the idea with generic computer hardware.

For AI algorithm claims, Step 2A is where most applications quietly die. A claim written as "a method for predicting patient readmission risk using a trained neural network" describes an abstract mathematical process — pattern recognition applied to data. An examiner applying the 2019 USPTO Revised Guidance (which refined Alice into a more structured three-prong analysis under Step 2A) will identify the claim as directed to a mathematical concept and look for a "practical application" that goes beyond the concept itself. Pointing to the accuracy of the prediction does not clear this bar. The output of a model is not a technical improvement; it is a result.

This is where the Weight-Surface Inversion becomes strategically decisive. In AI algorithm patents, founders instinctively claim the model's predictive output or learned capability — which reads as abstract pattern recognition under Alice Step 2A — while the genuinely Alice-resistant claim surface is the deterministic computational scaffolding (loss function constraints, normalization pipeline, inference-time transformation sequence) that shaped what the weights could learn, because these are deliberate engineering decisions, not emergent properties, and they constitute the "particular machine or transformation" that Alice Step 2B demands.

The learned weights themselves are not patentable: they are a mathematical object that emerges from training, not a deliberate engineering choice the inventor can fully specify in a claim. The training procedure that constrained the weight space — that is what the inventor actually built, and that is what a well-drafted claim can protect.

What USPTO Examiners Are Actually Looking For in 2025

Following a series of AI-specific examination guidance updates culminating in the 2024 USPTO guidance on AI-assisted inventions, examiners now scrutinize three dimensions of an AI patent application simultaneously.

1. The Claim's Technical Character

A claim that recites specific computational steps — a defined preprocessing transformation, a particular feature-weighting scheme, a constrained gradient-descent variant — carries technical character. A claim that recites the goal (accurate diagnosis, efficient routing, fraud detection) does not. The CLS Bank majority rejected claims that described what a system accomplished in a settlement context; analogously, an AI claim that describes what a model outputs rather than how the computation is structured will fail at Step 2A Prong 2.

Concrete example: a claim for "a method of detecting diabetic retinopathy comprising applying a convolutional neural network to retinal images" is almost certainly abstract. A claim for "a method comprising applying a depthwise-separable convolution sequence with a defined channel-expansion ratio to a normalized fundus image tensor, wherein normalization comprises a vessel-contrast equalization step parameterized by a learned per-image gamma value" has a fighting chance — because it specifies the computational transformation, not the clinical result.

2. The §112 Written-Description Requirement

Even claims with adequate technical character fail if the specification cannot support them under 35 U.S.C. §112(a). For AI patents, this has become a distinct rejection pathway: examiners increasingly issue §112 rejections when a specification describes a trained model's outputs without disclosing the training data characteristics, the loss function architecture, or the hyperparameter constraints that define what the model actually does. If your specification says "a neural network trained to classify anomalies," it does not describe an invention — it describes a category of machines. Reproducibility by a person having ordinary skill in the art (PHOSITA) requires enough specificity to train a model that behaves as claimed, which in practice means disclosing the computational scaffolding, not just the performance benchmark.

3. Novelty and Non-Obviousness Relative to the Prior Art Explosion

The AI patent filing rate has roughly tripled since 2019. The USPTO's patent examination unit responsible for AI (Technology Center 2100 and 3600) is working through a backlog that means first office actions on AI applications are arriving 18 to 24 months after filing. The practical consequence: the prior art landscape your application faces at examination is denser than the landscape at filing. A narrow provisional filed today buys a priority date; it does not buy protection from a crowded field when your non-provisional is examined in 2027. Filing a provisional without a claim strategy aligned to the Weight-Surface Inversion — i.e., anchored to your computational scaffolding, not your model's outputs — is structurally equivalent to establishing a priority date for a claim you cannot defend.

The Patent vs. Trade Secret Decision for AI Algorithms

Patent protection requires public disclosure. For AI algorithms, disclosure means specifying the computational architecture, training procedure, and data pipeline in enough detail to satisfy §112. That disclosure is then publicly available 18 months after filing, regardless of whether the patent ultimately issues.

The trade-off is asymmetric in ways specific to AI. A well-trained model's weights are not reconstructible from a patent claim — a competitor reading your issued patent cannot replicate your model without your training data and compute infrastructure. This means the patent's disclosure cost is real, but its protective value depends on whether your Alice-resistant claim surface (the computational scaffolding) is actually the competitive moat, or whether the moat is the trained artifact itself.

For most seed-stage AI companies, the trained model — the weights, the inference latency profile, the edge-case behavior shaped by proprietary training data — is the moat. That artifact is not patentable and is best protected as a trade secret combined with restrictive API access. The patent, when genuinely warranted, protects the specific algorithmic innovation: a novel preprocessing transform, a constrained architecture variant, a defined fine-tuning protocol that generalizes in a measurable way prior art does not.

Filing under the micro-entity fee tier — available to individuals and small entities meeting gross income and prior-filing thresholds — reduces USPTO fees by 80% relative to large-entity rates. For a non-provisional utility application, micro-entity fees in 2025 total approximately $830 in basic filing, search, and examination fees, versus approximately $4,140 for a large entity. That difference matters less than claim quality: an $830 application with claims anchored to abstract outputs is worth zero defensible IP. A $4,140 application with claims anchored to the Weight-Surface Inversion framework's computational scaffolding can anchor a Series A IP narrative.

International Filing: The PCT Window and Its Strategic Constraints

The Patent Cooperation Treaty (PCT) allows a single international application to establish priority in over 150 jurisdictions. For AI patents, the strategic calculus differs by region. The European Patent Office (EPO) applies its own "technical character" requirement under Article 52 EPC, which broadly excludes programs for computers "as such" but has developed a workable doctrine for AI claims that specify a technical effect going beyond normal physical interactions — a standard that maps closely to the Alice Step 2B "significantly more" inquiry. The China National Intellectual Property Administration (CNIPA) updated its AI patent examination guidelines in 2021 to allow claims that specify model architecture and training methodology, creating a filing environment that in some respects rewards the Weight-Surface Inversion strategy more directly than the USPTO does.

The PCT national-phase entry deadline is 30 months from priority date. For a startup with a U.S. provisional filed today, the PCT decision arrives in the middle of a typical seed-to-Series A transition — a moment when the cost of national-phase entry (commonly $3,000–$8,000 per jurisdiction before local counsel fees) competes directly with engineering budget. The decision to enter national phase in EPO, CNIPA, and the U.S. simultaneously should be driven by where the company's licensees, acquirers, or product markets sit — not by a generic "protect everywhere" instinct.

Building the Claim Architecture: A Decision Map

• Start with the computational scaffolding, not the output. Identify every deterministic engineering decision in your training and inference pipeline: preprocessing transforms, loss function formulation, regularization constraints, architecture search space restrictions. These are your candidate claim elements.
• Test each element against Alice Step 2B. Ask: does this element improve a specific technical process — faster convergence, reduced memory footprint, measurable generalization on a defined benchmark — or does it merely describe the goal? Only the former survives examination.
• Anchor your specification to reproducibility, not performance. Write the detailed description so that a PHOSITA could construct a training pipeline that produces a model with the claimed computational properties. Cite specific hyperparameter ranges, dataset preprocessing requirements, and evaluation metrics that are properties of the computation, not just outcomes.
• File a provisional within 12 months of any public disclosure. The 12-month statutory bar under 35 U.S.C. §102(b)(1) is absolute. A conference presentation, a preprint, a GitHub repository with a meaningful commit — any of these starts the clock.
• Do not treat the provisional as a placeholder. A provisional that describes only outputs and goals locks in a priority date for indefensible claims. The provisional's specification becomes the §112 anchor for the non-provisional; if it lacks computational specificity, the non-provisional cannot add it without losing the priority date.

FAQ: What Founders Actually Get Wrong About AI Patents

If my AI model outperforms every prior art benchmark, doesn't that novelty make it patentable?

No — and this is the single most expensive misconception in AI patent strategy. Performance improvement is an outcome, not a patentable method. An examiner applying Alice Step 2A does not ask "is this model better?" but "is this claim directed to an abstract idea?" A model that achieves 97% diagnostic accuracy via a process described only in terms of its outputs fails §101 regardless of benchmark superiority. The novelty that matters to an examiner is in the computational steps, not the result. For investors, this distinction matters acutely: a patent portfolio built on output-anchored claims is a liability, not an asset, because each claim is a pending Alice rejection waiting to surface during diligence.

Should I patent my training data pipeline, or just keep it as a trade secret?

The question contains a false binary. The training data itself is not patentable subject matter; a preprocessing transformation applied to data may be. The strategic question is whether your data pipeline's competitive advantage lies in the computational method (patentable, with disclosure risk) or in the proprietary data distribution it was built to handle (protectable only as a trade secret). If a competitor with different data could replicate your pipeline's effect, the pipeline is not your moat — the data is, and trade secret protection is appropriate. If your pipeline introduces a novel normalization or augmentation transform that generalizes across data distributions in a measurable way, that transform may be the Alice-resistant claim surface worth protecting through patent. These two strategies can coexist if the boundary between them is designed deliberately before filing.

Does filing a provisional patent protect me from a competitor who files a non-provisional on a similar method the same week?

Under the America Invents Act's first-inventor-to-file system, the provisional establishes your priority date — but only for subject matter actually disclosed in the provisional. If your provisional describes outputs and goals while your competitor's non-provisional describes the underlying computational method with §112-compliant specificity, your priority date does not help you: you have priority over nothing that will survive examination. The protective value of a provisional is precisely equal to the quality of its technical disclosure, not its filing date in isolation. This is the structural trap that generic "file a provisional first" advice consistently fails to address.

Can a patent on an AI algorithm block a competitor who independently trains a different model to do the same thing?

Only if your claims are written at the level of the computational method, not the application. A claim covering "a method comprising [specific training steps and architectural constraints]" can block a competitor who uses those specific steps, regardless of whether they trained their model independently. A claim written as "a method of [achieving result X] using a neural network" cannot reliably block anyone, because a competitor's different architecture achieving the same result does not infringe a result-based claim — and such a claim would likely fail Alice anyway. The scope of protection your patent actually delivers is a direct function of how far down the Weight-Surface Inversion your claims are anchored.

Is an AI-generated invention patentable if a human made the final engineering decision?

Under current USPTO policy following its 2024 guidance on AI-assisted inventions, AI cannot be listed as an inventor. A human must have made a "significant contribution" to the conception of the claimed invention. Using an AI tool to explore an architecture search space, then making the deliberate engineering decision to constrain the search to a specific subspace for a specific technical reason, likely qualifies the human engineer as inventor. Allowing an AI to autonomously generate a solution and then simply adopting it likely does not. The practical risk is not just inventorship challenge; it is inequitable conduct: if the AI's role is material and undisclosed, the patent can be rendered unenforceable. Document the human decision points in the engineering record — not as legal formality, but as the factual foundation for inventorship that examination and litigation will both require.

This article is for informational purposes only and does not constitute legal advice. Consult a qualified patent attorney for guidance specific to your situation.