The US Patent Office tested eight AI patent claims. Five passed the patent-eligibility test. Three failed. The surprising part? The claims covered just three underlying technologies: neural-network security, speech processing, and AI-assisted medical treatment. In each case, the USPTO showed how an algorithm-based invention could be claimed in ways that produced different legal outcomes. The difference was not simply whether the invention used artificial intelligence. It was what the patent claim actually required the technology to do. PatentPC analyzed these examples alongside federal court decisions, USPTO examination data, and global patent publications. Our findings point to a simple rule:

You generally cannot patent a mathematical algorithm in the abstract. But you may be able to patent a specific technological invention that uses an algorithm.

And for software and AI companies, the difference can determine whether a core competitive advantage receives meaningful protection.

The short answer: Can you patent an algorithm?

Yes, in some circumstances. US patent law does not ban software algorithms as a category. But an abstract mathematical procedure, claimed by itself, generally cannot be patented. A qualifying invention must meet several requirements:

Requirement	What it means
§101 — Eligibility	The claim must cover patent-eligible subject matter, not merely an abstract idea.
§102 — Novelty	The claimed invention must be new.
§103 — Nonobviousness	It cannot be an obvious development of earlier technology.
§112 — Disclosure	The application must adequately explain and support the claimed invention.

An algorithm that controls a machine, improves data processing, changes how computers operate, or solves a technical problem may qualify. Merely putting a formula on a computer is not enough. The USPTO explains these distinctions in MPEP §2106. PatentPC covers the broader framework in What Can Be Patented in Software? and How to Patent an Algorithm.

Original Research #1: Three AI Algorithms, Eight Claims, Five Eligible

In July 2024, the USPTO published three detailed hypothetical AI patent examples, numbered 47, 48, and 49. PatentPC classified all eight claims by outcome and by the technical features that changed the analysis. Here is the result.

USPTO example	Ineligible claim	Eligible claims	What changed?
47: Neural-network anomaly detection	Detect and report anomalies using a neural network	A defined neural-network circuit; a method that drops malicious packets and blocks traffic	Specific hardware or improved network security
48: Speech separation	Calculate speech embedding vectors using a neural network	Reconstruct separated speech; generate a transcript from separated speech	Technical signal processing and transcription
49: AI-assisted medical treatment	Calculate patient risk and administer an unspecified appropriate treatment	Use the risk calculation to administer a specific treatment	A particular medical treatment

Total:

• Three ineligible claims.
• Five eligible claims.
• Eight claims across three technologies.

PatentPC calculation: 5 ÷ 8 = 62.5% eligible 3 ÷ 8 = 37.5% ineligible Source: USPTO AI Patent Eligibility Examples 47–49. Important: These are deliberately constructed teaching examples, not a random sample of real patent applications. The 62.5% figure is not a patent approval rate. The useful discovery comes from comparing the claims inside each example.

The three-for-three pattern

All three examples contain an initial algorithm-based claim the USPTO considers ineligible. In all three, the agency also provides a more specific claim that qualifies as eligible. That’s a three-for-three illustration of eligibility changing with the claimed implementation. The important change is not simply adding more words. It is claiming a meaningful technological application of the underlying calculation. Let’s examine the differences.

Example 47: Detecting malicious traffic versus stopping it

An AI system receives network data and identifies anomalies. That sounds useful. But the USPTO finds the example’s claim to detecting anomalies through mathematical processing ineligible. Now add a specific technical process: The system identifies malicious packets, detects their source, drops those packets in real time, and blocks future traffic from that address. The USPTO finds the resulting claim eligible because it integrates the abstract idea into an improvement in network security. The patent lesson: Do not stop your invention disclosure at “our model detects threats.” Explain precisely how your system changes network operation. PatentPC explores similar issues in The Role of Technical Improvements in Software Patentability.

Example 48: Calculating a speech vector versus processing speech

An algorithm receives mixed audio and uses a neural network to calculate embedding vectors. The USPTO finds that claim ineligible. But a more detailed claim uses those embeddings to separate speech sources and reconstruct speech waveforms. Another processes separated speech to generate a transcript. Both qualify as eligible under the example’s facts. The patent lesson: A useful mathematical representation is not automatically a patentable technical invention. What the system does with that representation can matter enormously.

Example 49: Predicting disease risk versus specifying treatment

A model processes genetic data and calculates a patient’s risk of post-surgical fibrosis. The initial claim calls for administering an unspecified appropriate treatment. The USPTO finds it ineligible. A dependent claim specifies a particular treatment using Compound X eye drops. That claim qualifies as eligible because it integrates the calculation into a specific medical treatment. The patent lesson: A prediction and a claimed treatment are legally different things. However, a particular treatment does not automatically satisfy novelty, obviousness, safety, or any other patent requirement. The agency’s examples establish something more useful than a general statement that algorithms can be patented:

The claim must protect how an invention works, not merely describe the information an algorithm produces.

Original Research #2: Patent Rejections Fell 25% After USPTO Guidance Changed

There is another surprising finding in the government’s examination data. The Supreme Court’s 2014 Alice decision made software patent eligibility more difficult and uncertain. In 2019, the USPTO changed its examination guidance. Its Office of the Chief Economist later studied the results. For technologies particularly affected by Alice, the USPTO found that one year after the new guidance: The likelihood of receiving a first-office-action §101 rejection fell 25%. And: Uncertainty in first-action eligibility decisions fell 44%. Source: USPTO — Adjusting to Alice. Those findings are not estimates. They come from official examination research.

Why this matters

The underlying patent statute did not suddenly change in 2019. The examination guidance did. And patent examination outcomes changed significantly. That tells us two things. First, algorithm patentability is not determined by a simple list of approved and prohibited technologies. Second, claim analysis and the correct legal framework matter greatly. There is an important limitation: USPTO examination guidance does not bind federal courts. A patent that survives examination can still face an eligibility challenge in litigation. This is why PatentPC treats the Alice test as one part of a much larger patent strategy—not a formality.

Original Research #3: One AI Patent Won on Eligibility but Still Faced Rejection for Obviousness

In September 2025, the USPTO’s Appeals Review Panel decided Ex parte Desjardins. The invention concerned machine-learning technology, including improvements to continual learning. The panel found the claims patent eligible under §101. The USPTO designated the decision precedential in November 2025. But here’s the part founders need to understand: A separate rejection under §103 remained. The invention could qualify as patent-eligible technology and still fail the test of nonobviousness. The USPTO explicitly highlighted this distinction in its Desjardins announcement. This produces a simple framework.

PatentPC’s Two-Gate Test

Gate 1: Is this the kind of technological invention patent law can protect? That’s principally §101. Gate 2: Is this particular invention new, nonobvious, and adequately disclosed? That’s principally §§102, 103, and 112. Passing Gate 1 does not mean passing Gate 2. A company could have a genuinely technical algorithm and still be unable to patent it because somebody already disclosed the same method. Or because the claimed advance was obvious. Or because the application does not explain how to implement the invention sufficiently. That is why PatentPC recommends combining software patent eligibility analysis with proper prior-art searching.

The Federal Circuit Cases Reveal the Same Pattern

Four important decisions show the distinction.

Case	Technology	§101 outcome	Main lesson
Enfish v. Microsoft (2016)	Self-referential database table	Eligible	A specific improvement to computer operation can qualify.
McRO v. Bandai Namco (2016)	Rules for automated facial animation	Eligible	Particular rules improving animation technology can qualify.
SAP v. InvestPic (2018)	Statistical analysis of financial data	Ineligible	Improved mathematical analysis alone may remain abstract.
Recentive v. Fox (2025)	Generic ML applied to scheduling and network maps	Ineligible	Applying existing ML to a new field is not automatically a technological improvement.

Read the decisions: Enfish, McRO, SAP v. InvestPic, and Recentive. The Supreme Court declined to review Recentive in December 2025. The distinction is especially important for AI. Suppose you apply an ordinary neural network to a new business task. You might produce an excellent product. But that does not necessarily mean you invented new technology. Now suppose you create a new training method that reduces memory use through a specific change in model architecture. That is a different patent story. The first describes where AI is used. The second describes how technology is improved. Neither guarantees a patent. But the second presents a much stronger starting point for technical eligibility analysis.

A Real Algorithm Patent: Google’s Original PageRank

There is a famous counterexample to the claim that algorithms and formulaes can never be patented. US Patent 6,285,999: Method for Node Ranking in a Linked Database. The patent names Lawrence Page as inventor. It was filed in January 1998, originally assigned to Stanford University, and issued in September 2001. It describes a method for ranking nodes in a linked database using relationships between documents. The original patent has since expired. Read the actual PageRank patent. This does not mean an identical claim would necessarily qualify under today’s eligibility rules. It establishes a narrower historical fact: Specific algorithm-driven methods have been patented in the United States. And it illustrates why claim drafting matters more than merely naming an algorithm. PatentPC also discusses how software patents work and what is patentable in software.

Original Research #4: GenAI Software/Code Patent Publications Grew 4.77× in Two Years

While the legal rules remain demanding, the commercial race is accelerating. WIPO identified: 339 published GenAI patent families in the software/code output category in 2023. By 2025: 1,616. PatentPC calculation: 1,616 ÷ 339 = 4.77× An increase of: 376.7%. But compare that with the overall GenAI landscape.

WIPO category	2023	2025	Growth
All GenAI patent families	14,080	37,808	2.69×
GenAI software/code output	339	1,616	4.77×

The smaller software/code category grew about: 4.77 ÷ 2.69 = 1.77× as fast as the overall GenAI patent-family total. Its share of that total rose from: 2.41% to 4.27%. That is a relative increase of about 77.5% in share. Sources: WIPO’s 2014–2023 patent landscape and WIPO’s 2025 GenAI update. A crucial distinction: WIPO’s software/code category concerns the type of content handled or generated by GenAI. It is not a count of every patent claiming a software algorithm. Still, the rate of growth is striking. Patent publications concerning code-related GenAI applications are expanding much faster than the overall GenAI landscape. For founders, the implication is practical: Search the patent landscape before assuming your software implementation is new. And do not limit searches to words such as “AI agent,” “LLM,” or “RAG.” Older patents may describe relevant mechanisms using terms such as information retrieval, resource scheduling, data transformation, or automated decision systems. PatentPC covers this in AI-powered patent searching and AI patent portfolio management.

What Changed in 2026? Technical Evidence Now Has a Clearer Route Into §101 Examination

There has also been an important procedural development. In 2025 and 2026, the USPTO clarified its approach to Subject Matter Eligibility Declarations, or SMEDs. These are voluntary declarations under Rule 132 that can provide factual evidence relevant to eligibility. In April 2026, the agency updated its applicant guidance. In September 2026, it added a flowchart and further best practices for reviewing declarations. See the April 2026 guidance and September 2026 update. These memoranda do not create a new legal eligibility test. But they reinforce a useful practice: Collect technical evidence before you need it. Suppose your algorithm improves a database. Preserve measurements of:

• query latency;
• memory use;
• throughput;
• storage requirements;
• processor time;
• failure rates.

Also preserve the architecture and test conditions that explain the improvement. A 40% speed increase, standing alone, does not establish eligibility. A specific technical mechanism supported by evidence is much more useful. And a later declaration cannot repair missing technical disclosure in the original application. PatentPC covers practical drafting in Tips for Drafting Patent Claims for AI and Machine Learning Inventions.

The Algorithm Patentability Checklist

Before paying to patent an algorithm, answer these seven questions.

Question	What a useful answer contains
1. What technical problem exists?	A clear weakness in an existing system
2. What was the old mechanism?	The prior architecture or method
3. What exactly did we change?	The new steps, data flow, or structure
4. Why was that change needed?	Failed alternatives or technical constraints
5. What improved?	Relevant technical measurements
6. What makes it different?	Specific differences from prior art
7. Could competitors work around it?	Practical alternative implementations

The most important question is number three. What exactly changed? If the answer is merely “we used AI,” you probably have not identified the real invention. If the answer describes a new way of organizing memory, processing signals, scheduling compute, training a model, controlling hardware, or validating outputs, you may have something worth investigating. PatentPC’s guides to strong software patents, technical effects, and broad but supportable patent claims explain how those facts become an application.

How to Patent an Algorithm: The Practical Filing Process

First, identify the invention—not just the product. A chatbot, recommender, or trading platform may contain many algorithms and formulaes. Identify the specific technical mechanism you believe is new. Second, search prior art. Use the USPTO Patent Public Search alongside technical papers, earlier patents, public software, and other relevant disclosures. Search what the algorithm does, not just what you call it. Third, prepare an engineering disclosure. Include the old architecture, new mechanism, data structures, flowcharts, alternatives, failure states, and relevant test results. The USPTO’s enablement guidance explains why an application must teach the claimed invention adequately. Fourth, choose the filing strategy. A provisional application can secure an initial filing date for adequately disclosed subject matter, but it is not examined and normally must be followed by a nonprovisional application within 12 months. See the USPTO provisional application rules and PatentPC’s guide to writing provisionals. Fifth, review public disclosure before filing. The US has a limited grace period for certain inventor-originated disclosures. Many other countries are less forgiving. The safer approach when foreign patent rights matter is to evaluate filing before publishing the algorithm, giving a technical presentation, or releasing detailed documentation. See the USPTO’s international filing guidance. Finally, remember that owning a patent does not guarantee freedom to use the technology. Other patent owners may hold earlier claims covering parts of your system. PatentPC explains that separate question in How to Conduct a Freedom-to-Operate Analysis.

Should You Patent the Algorithm or Keep It Secret?

Sometimes secrecy is stronger.

Situation	Protection worth considering
Competitors can inspect or independently reproduce the technical mechanism	Patent
Valuable algorithm remains hidden inside your servers	Trade secret
You mainly want to stop copying your source code	Copyright
A public architecture depends on hidden tuning or data	Patent plus trade secret
Algorithm is routine or commercially unimportant	Possibly no filing

The US Copyright Office explains that copyright can protect code expression, but not Algorithms and formulaes, methods, or program logic as such. WIPO explains that trade-secret protection generally cannot stop lawful independent development. A valid patent can reach independent implementation of the claimed invention. That difference can be particularly important when AI tools make software easier to recreate. PatentPC’s research on patents as AI moats explores that wider economic shift. For invisible technical know-how, read The Role of Trade Secrets in Protecting AI and Deep Learning Code.

Can You Patent an Algorithm Created With AI?

Using AI to help develop an algorithm does not automatically make the invention unpatentable. The USPTO’s November 2025 guidance confirms that the ordinary inventorship standard applies to AI-assisted inventions. But only natural persons can be named as inventors. AI systems are tools, not legal inventors. See the USPTO’s revised AI inventorship guidance. For AI-heavy engineering teams, the practical rule is to document which humans conceived the claimed invention and what AI tools contributed. PatentPC also covers this in The Impact of AI and IP Rights.

Why PatentPC Focuses on the Technical Mechanism

The mistake in many algorithm patent applications happens before the lawyer starts drafting. The founder explains what the software accomplishes. Nobody explains exactly how the technology accomplishes it. That missing explanation can undermine eligibility, patent scope, enablement, and the ability to distinguish prior art. Good software and AI patent counsel should be able to examine the algorithm’s architecture, identify the technical contribution, review earlier technologies, evaluate competing implementations, and decide whether patenting is commercially worthwhile. PatentPC works with software, AI, and advanced-technology inventions through that technical and commercial lens. The objective is not to patent every algorithm. It is to protect the technical mechanisms competitors would have a real reason to reproduce. Speak with PatentPC about a software or AI patent.

The Bottom Line

Algorithms and formulaes can be patented in the United States when the claimed invention meets the legal requirements. But patenting an abstract calculation is not the same as protecting a technical process that uses it. PatentPC’s analysis found:

• Five of eight USPTO hypothetical AI claims were eligible, while three were not.
• All three USPTO examples contained both an ineligible algorithm-based claim and a more specifically applied eligible claim.
• USPTO guidance was followed by a 25% reduction in first-action §101 rejection likelihood and a 44% reduction in examination uncertainty for Alice-affected technologies in the agency’s historical study.
• A precedential AI decision found claims eligible under §101 even though a §103 obviousness rejection remained.
• WIPO’s GenAI software/code output category grew 4.77× from 2023 to 2025—about 1.77× the growth multiple of the overall GenAI patent-family total.

Together, those findings reveal a critical distinction.

The question is not whether your algorithm is clever. It is whether you can claim and support a new technological invention that uses it.

That is where valuable software patent protection begins.

PatentPC Research Methodology

This article combines three research methods. 1. Claim-by-claim classification. We reviewed USPTO Examples 47–49, classified all eight hypothetical claims, and compared the features associated with their §101 outcomes. The five-out-of-eight result is descriptive, not an estimated allowance rate. 2. Examination and case-law comparison. We used the USPTO’s 2020 Adjusting to Alice report, the precedential Ex parte Desjardins decision, and Federal Circuit rulings. Court outcomes were analyzed for their legal reasoning, not counted as a statistically representative sample. 3. Patent-family calculations. We compared WIPO’s 2023 and 2025 GenAI publication counts and its separate software/code output-mode series. Calculations:

Measure	Calculation	Result
Eligible USPTO example claims	5 ÷ 8	62.5%
Ineligible USPTO example claims	3 ÷ 8	37.5%
GenAI software/code growth	1,616 ÷ 339	4.77×
All GenAI family growth	37,808 ÷ 14,080	2.69×
Relative growth multiple	4.77 ÷ 2.69	~1.77×
Software/code share, 2023	339 ÷ 14,080	2.41%
Software/code share, 2025	1,616 ÷ 37,808	4.27%

The WIPO figures count published patent families, not granted US patents. They do not establish how many algorithms were allowed under §101 or how many enforceable claims exist. The 25% and 44% examination findings come from a historical USPTO study, not from a new 2026 allowance-rate analysis. Core legal and research sources: USPTO eligibility guidance, USPTO AI examples, USPTO examination research, PTAB precedential decisions, 2026 subject-matter-eligibility declaration guidance, WIPO GenAI patent landscape, and WIPO’s 2026 update. This article provides general educational information, not legal advice. Patent eligibility, novelty, obviousness, enablement, ownership, filing strategy, and infringement depend on the particular invention and its claims.