Can Software Be Patented in 2026? The Alice/§101 Test, Simplified

Can software be patented in 2026? Yes, but not the way most people assume, and two Federal Circuit rulings in the last eighteen months have put much harder edges on the answer. What’s changed is not the law itself but how clearly the courts have now drawn the line for AI and machine-learning claims specifically, and that line sits closer to the mechanism, and further from the result, than a lot of applicants expect.

So, can software be patented in 2026? The test you’re up against

Every software or computer-implemented claim gets run through the same two-step framework the Supreme Court set out in Alice Corp. v. CLS Bank back in 2014, sometimes called the Alice/Mayo test.

Step One asks whether the claim is “directed to” an abstract idea, a law of nature, or a natural phenomenon. Courts have never defined “abstract idea” with real precision, which is part of why this area of the law stays unsettled. Claims describing collecting data, analyzing it, and presenting a result routinely land here.

Step Two asks whether the claim contains an “inventive concept,” something beyond the abstract idea itself, sufficient to transform it into something patent-eligible. This is where most software claims get decided, and it’s the step both recent cases turned on.

That two-step framing is the shape of the test, not the whole of it. The real analysis has more moving parts than a two-box diagram suggests. The USPTO runs the same inquiry through its own structure in MPEP § 2106: Step 1 asks whether the claim falls into a statutory category at all; Step 2A Prong One asks whether it recites a judicial exception; Step 2A Prong Two asks whether the claim integrates that exception into a practical application; and Step 2B asks whether anything else in the claim amounts to significantly more.

Courts keep the two-step form, but do a good deal of unstated sub-analysis inside it. If you are reading an office action, Prong Two is usually where the argument sits. This article deliberately stays at the two-step level. A real eligibility opinion does not.

Neither step has a bright-line definition, and the Supreme Court has repeatedly declined to take up cases that would clarify one. That’s left the Federal Circuit to draw the boundary case by case. It did exactly that, twice, on machine-learning claims in the last year and a half.

Case one: Recentive Analytics v. Fox Corp.

In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), the patents at issue covered using machine learning to generate optimized event schedules and network maps. The Federal Circuit found the claims ineligible. Training a machine-learning model and applying it to a new dataset, even a dataset from a field no one had applied ML to before, wasn’t itself an inventive concept. The court’s reasoning was direct: applying a known method to a new field of use doesn’t make the method inventive, even when the results are genuinely useful and the field is new.

Case two: Dental Monitoring SAS v. Align Technology

Dental Monitoring SAS v. Align Technology, Inc., No. 24-2270 (Fed. Cir., July 7, 2026), pushed the same reasoning further. The patents (U.S. 11,049,248 and 10,755,409) covered using deep-learning devices to analyze dental images and assess orthodontic aligner fit. Before this technology, orthodontists did that work by hand.

The Federal Circuit affirmed a district court’s summary judgment that the claims were ineligible. At Step One, the court held the claims were directed to the abstract idea of “collecting information, analyzing it, and displaying certain results.” Deep learning didn’t change that categorization. At Step Two, the court found no inventive concept: training a “generic” deep-learning device on a specific set of tooth images, the opinion said, is “incident to the very nature of machine learning.” In other words, that’s just what machine learning is, not something inventive layered on top of it.

The court also rejected the argument that the claims were eligible simply because this particular application of deep learning was novel when the patents issued. Novelty and eligibility are two separate questions, and one doesn’t answer the other. IPWatchdog’s coverage of the ruling has the full breakdown if you want the litigation play-by-play.

What fails, almost every time

Read together, these two cases sketch a pattern that’s fairly reliable to test claims against. The live question is never whether software can be patented in the abstract. It is whether this particular claim survives step two.

Claims fail when they describe applying an off-the-shelf machine-learning technique (one built by a major AI provider, or one that’s standard in the field) to a new domain, dataset, or business problem, with the claimed advance being the result (more accurate predictions, better schedules, faster analysis) rather than any change to how the system produces that result. If a competent engineer could swap in any commercially available ML model and get a similar outcome, that claim is exposed. If your invention is a mobile app rather than backend ML infrastructure, the same logic applies to your claims too. Our guide on how to patent an app walks through the filing side of that.

What tends to survive

Claims fare better when they describe a concrete technical mechanism: a new way data is structured before it’s processed, a new architecture for how components interact, a new method for training or constraining a model that isn’t itself generic, or some other change to the computer system’s actual operation rather than its output. The distinguishing question examiners and courts keep coming back to is whether the claim is doing something to the technology, or just doing something with it. Asked in the abstract, “can software be patented” has no useful answer. Asked of a specific claim, it usually has a clear one.

The one piece of good news for applicants

None of this means every AI-related Section 101 rejection is correct. In August 2025, USPTO Commissioner Kim issued an internal reminder memo to examiners, quickly nicknamed the “Kim Memo,” cautioning against over-broad Section 101 rejections of software and AI claims. The memo itself is public, and the short version is this: it reminds examiners that a mental-process abstract idea shouldn’t be found in claim limitations that couldn’t practically be performed in a person’s head, that claims have to be evaluated as a whole rather than picked apart limitation by limitation, and that a Section 101 rejection needs to clear a “preponderance” standard. If the facts underlying it are genuinely debatable, the rejection should be withdrawn rather than maintained.

That memo doesn’t reverse either court decision above. It does mean applicants facing an examiner who’s applying Section 101 too aggressively now have an internal USPTO document to push back with. That’s a real, usable argument in prosecution, not just a courtroom-only defense.

A practitioner’s view: clarification, not upheaval

Coverage of these cases tends to treat them as a shock to the system. After twenty-five years of writing software patents, that is not how they read from this side of the desk.

Alice and the decisions built on it are a crystallization, not a break. They add data points that fill in gaps and, on the whole, improve certainty about where the boundary sits. The principles underlying software protection before Alice are the same principles operating after it. Section 101 exists to stop the patent system from issuing claims on results, laws of nature and mental steps, and to keep it pointed at what it is for: innovative, engineered solutions. A solution can produce a result. The result itself is not the invention, and it is not what you get to claim.

If your drafting is built on those first principles, Recentive and Dental Monitoring do not change your practice. They give you two more data points to weigh, which is what any professional does with new authority. If a pair of decisions like these lands hard on a portfolio, that is the useful signal: it usually means the drafting was tracking case headlines rather than the reasons the framework exists in the first place. So the honest answer to whether can software be patented after these rulings is the same answer as before them. It depends on whether the claim recites a mechanism or a result.

What this means for how you draft claims

The safest place to fix a Section 101 problem is in the claim drafting, before filing, not in litigation after a competitor challenges the patent. If your invention uses AI or machine learning anywhere in it, the claim needs to say something specific about how your system works, not just what it accomplishes better than before.

FAQ

Does this mean AI-related inventions can’t be patented anymore?

No. Both cases turned on how the claims were written, not on machine learning being categorically unpatentable. Claims describing a specific technical improvement to the system itself are a different fact pattern than the ones that lost.

What’s the difference between Section 101 and Section 103 (obviousness)?

Section 101 asks whether the claimed subject matter is eligible for a patent at all. Section 103 asks whether an eligible invention is too obvious over existing prior art to deserve one. The Dental Monitoring court’s finding on eligibility meant it didn’t need to reach the defendant’s separate written-description arguments. Eligibility can end a case before other defenses are even addressed.

Is this only a problem for AI/machine-learning patents?

No. The Alice/Mayo framework applies to all software and computer-implemented claims. AI and ML claims have simply been where the Federal Circuit has focused most recently, in part because “we applied a known model to a new field” is a claim pattern that comes up often as more industries adopt off-the-shelf AI tools.

So, can software be patented, or not?

Yes. Software is not excluded from patent protection, and neither is machine learning. What the two cases above decide is narrower: a claim that describes a result gets refused, and a claim that describes a mechanism does not. The subject matter is eligible; the drafting is what fails.

How do I know if my claim would survive this test?

There’s no shortcut that replaces a claim-by-claim review, and it depends on language a court hasn’t seen yet in your specific case. This is exactly the kind of judgment call worth getting in front of a patent attorney before you file rather than after an office action.

Whether your software can be patented is decided by how the claim is written, and that is a drafting skill rather than a checklist you can run yourself. It takes a patent attorney who can tell the difference, in your specific invention, between a claim that describes a result and a claim that describes a mechanism, and who’s tracking how examiners are actually applying both the Dental Monitoring reasoning and the Kim Memo right now. That’s the review worth getting before you file, not after a Section 101 rejection forces the question.

Get Your Free Patent Needs Assessment