The How Has to Live in the Claim: What the DeepMind Ruling Teaches Founders About Patent Drafting

Authored by Babak Akhlaghi on June 16, 2026.

In September 2025, in his first week as USPTO Director, John Squires vacated a Patent Trial and Appeal Board decision that had rejected DeepMind’s AI patent on eligibility grounds. The ruling became precedential in November.

This wasn’t just another administrative decision. It was a signal.

The message to patent examiners: stop reflexively rejecting AI innovations under Section 101. The message to founders: your software patents are as eligible as hardware patents, but only if you draft them correctly.

But here’s the critical lesson from this case:

The DeepMind application could have easily failed if the claims had only contained broad functional language. The specification described the technical problem, the solution, and the advantages in detail.

But that wasn’t enough. The claims had to capture the “how” too.

The Mistake That Kills AI Patents

The DeepMind case turned on one sentence in the claim. Not the specification. Not the detailed technical description. One claim limitation that explained how the system actually worked:

“adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task.”

That single limitation, supported by the technical problem, solution and advantages of the solution, was the difference between patent eligible and patent ineligible.

The invention addressed a real technical problem called catastrophic forgetting—when neural networks lose knowledge of previous tasks while learning new ones. The specification explained this beautifully. And, the claim captured the solution.

The how lived in the claim, not just the spec.

Director Squires pointed directly to this limitation when he vacated the Board’s rejection. The claim didn’t just say “train a machine learning model.” It explained the mechanism: preserve performance on the first task while optimizing for the second.

That’s technical. That’s specific. That’s patentable.

The Three-Part Framework You Need Before Filing

Here’s what founders should be doing with their patent attorneys before any application gets filed (and how this fits into your broader patent strategy for startups):

1. Identify the technical problem

Not the business problem. Not the user experience problem. The technical problem.

In DeepMind’s case: neural networks catastrophically forget earlier training when learning new tasks. That’s measurable. That’s a performance issue in the system itself.

2. Define the technical solution

What specifically does your invention do differently? What’s the core mechanism?

DeepMind’s solution: compute an approximation of a posterior distribution over parameter values, then use that to determine which parameters are important to the first task, then adjust parameters for the second task while protecting those important to the first.

That’s not a broad functional statement. That’s a process.

3. Capture the advantages in technical terms

The USPTO now includes DeepMind as an example in MPEP § 2106.05(a): reduced storage requirements, reduced system complexity, preservation of performance on earlier tasks.

Those aren’t abstract benefits. They’re measurable improvements to how the computer system functions.

What “The How” Actually Looks Like

Broad functional statements don’t survive Section 101 scrutiny. “Training a machine learning model” isn’t enough. “Optimizing parameters” isn’t enough.

You need sufficient technical detail in the claim itself.

Compare these two claim limitations:

Doesn’t work: “training the machine learning model on a second task”

Works: “training the machine learning model on the second training data to adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task”

The second version explains the mechanism. It tells you what happens to the parameters and why. It connects the solution to the problem.

That’s what the Director meant when he wrote that the claim “integrates an abstract idea into a practical application.”

The Question to Ask Your Patent Attorney

Before you file, ask this: “Is the how in the claim?”

Not “Is the invention described well in the specification?” That’s table stakes. The specification in the DeepMind case was excellent. The Board still rejected it.

The claim is what matters.

Your attorney should be able to point to specific claim limitations that explain:

• What technical problem the invention solves
• What mechanism it uses to solve it
• What measurable improvement results

If those elements live only in the specification, you’re vulnerable.

The Standard That Actually Protects You

Here’s something most founders don’t know: when an examiner issues a Section 101 rejection, they need to support it with a preponderance of evidence.

That means if you’re 51% right, the examiner should withdraw the rejection.

But you can only hit that threshold if your claim gives you something to argue with. If your claim just recites “training a machine learning model” and “computing a result,” you have nothing to work with.

The examiner will say it’s abstract. You’ll point to the specification. The examiner will say the claim doesn’t recite the improvement. And you’ll be stuck.

The DeepMind claim survived because it recited the improvement.

What Changed (and What Didn’t)

The DeepMind ruling is precedential at the USPTO. It binds examiners. Director Squires also issued a memo telling examiners to focus on Sections 102, 103, and 112—novelty, obviousness, and written description—not Section 101 eligibility.

That’s the good news.

The reality is more complicated. Some examiners still issue Section 101 rejections at an 80% rate. The message from the Director hasn’t fully reached the ground floor.

And the ruling isn’t binding on courts. If your patent gets litigated, you need more than USPTO guidance. You need Federal Circuit precedent.

That means when you’re drafting claims and responding to rejections, cite both. Use the DeepMind decision to get through the USPTO. But anchor everything in controlling Federal Circuit authority like Enfish and McRO.

That way, if your patent ever gets challenged in court, you can point to arguments the examiner considered under controlling precedent—not just internal USPTO policy.

The Counterexample: What Not to Do

In April 2025, the Federal Circuit held in Recentive Analytics 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.”

That’s the flip side of DeepMind.

Recentive’s claims described applying machine learning to a problem space. They didn’t explain how the machine learning itself improved. The claims were broad and functional.

The court said no.

DeepMind’s claims explained a specific improvement to how the model trains. The claims were technical and detailed.

The Director said yes.

The difference wasn’t the technology. It was the drafting.

The Real Cost of Getting This Wrong

Fighting a Section 101 rejection is brutal for founders. You’re not arguing about prior art. You’re not arguing about whether your invention is obvious. You’re arguing about whether your invention is even the kind of thing that deserves a patent.

And the statistics are terrible. Reversal rates on Section 101 appeals hover around 20%. That means if you go to the Board, you’re likely to lose.

Big companies can afford that fight. They have patent portfolios in the hundreds or thousands. They can take a few losses.

Founders can’t. You’re operating on limited runway. Your patent portfolio is how you compete, how you attract investment, how you prove your idea has been validated by the patent office.

Losing that fight means walking away from the money you’ve already spent. It means abandoning the application. It means your investors don’t get the “approved” seal they were counting on.

The time to fix this is before you file, not after you get rejected.

The Door Is Open Right Now

The pendulum is swinging in favor of AI founders. Director Squires sent a clear signal in his first week. The USPTO added DeepMind as an official example in the examination manual. Examiners are being told to focus on novelty and obviousness, not eligibility.

This is the environment you want to file in.

But only if you draft strategically. Only if the how lives in the claim. Only if you can point to specific limitations that explain the technical problem, the technical solution, and the measurable improvement.

If you’re sitting on an AI invention right now, this is the time to file. And if you’re working with a patent attorney, make sure they understand what the DeepMind case actually teaches:

The specification tells the story. The claim has to prove it.