The Patent Office Is Embracing AI. The Good and the Bad for Applicants.
By Brandon Furdock, Registered Patent Attorney, Chicago Patent Law LLC
The Patent Office made two moves on AI in the past year. It made AI inventions easier to patent, and it started teaching its examiners to search with AI. The first is good news for anyone building with AI. The second changes how the application has to be written. Here is both, and what to do about it.
The good: AI inventions are easier to patent than they were
For years a software or AI claim ran into the Section 101 wall. The examiner called it an abstract idea on a generic computer and the conversation ended there. Three moves in the last year are changing that, if only slightly so far. The numbers are below.
In November 2025 the Office made Ex parte Desjardins precedential: training a machine learning model is eligible when the math is tied to a practical application that improves the model itself. In December 2025 it wrote that into the examination manual. In December 2025 and again in April 2026 it told examiners to weigh sworn evidence that an invention is better, faster or cheaper in a real technical way. And on September 29, 2026, Director Squires told a Senate subcommittee the Office is taking a more receptive approach to AI applications.
The numbers have started to move. For years the Board upheld examiners' Section 101 rejections on appeal roughly nine times in ten. In the last quarter of 2025, after Squires took office, that dropped to about three in four, and in the software art units to about half. Still long odds, but the first real movement in years.
The scrutiny did not disappear, it moved. Examination data from March 2025 to March 2026 shows written description rejections under Section 112 rising relative to Section 101 rejections on AI applications after Desjardins. Examiners are asking harder how the model is trained, what goes in and comes out, and how inference actually runs. A thin description of a black box now draws a 112 rejection where it used to draw a 101.
One caution. The courts have not moved as far. The Federal Circuit still applies its own test, and this year it affirmed a rejection in In re McFadden of a software method that, in its view, did nothing more than run on ordinary hardware. The Office's guidance binds examiners, not judges. A patent that clears examination can still be challenged later, so the claims have to be built for both rooms.
The bad: the examiner may soon have an AI search in hand
In October 2025 the Office began a pilot it calls ASAP, the Artificial Intelligence Search Automated Pilot. An applicant who volunteered had the application run through the Office's internal AI tool before an examiner was assigned. The tool read the specification, claims and abstract and returned up to ten prior art documents, ranked by relevance, in a notice that went into the public file.
Few volunteered. By mid April, after the Office had waived the fee and doubled its target, under two hundred applicants had asked and under a hundred had been let in. That is not surprising. Nobody should opt into a search aimed at their own application, with the results published, in exchange for nothing. I would not have recommended it to a client and did not.
The pilot still matters, for a different reason. It shows where examination is going. Director Squires has described it as the first of many planned AI pilots in examination, and the Office's own page says it will be updated if the program is renewed or replaced.
An application filed today will likely be examined two to three years from now. It is reasonable to expect that by then an examiner may have an AI search in hand, whether or not anyone asked for it.
That changes two things. A thin application that leans on the examiner not finding something is a worse bet than it was, because the finding may be about to get cheap. And the closest art will show up earlier and more completely, in the first action rather than the third, so the application has to say up front what is different about the invention and why that difference matters. Arguing it later, after the examiner has the references in hand, is the expensive way.
What to do about it
If you are filing on something with AI inside, four things.
Write the improvement, not the label. "Uses machine learning" earns nothing. "Trains with half the labeled data because of this specific change to the loss" is what Desjardins rewards. The application has to say what got better and how.
Keep the evidence. Benchmarks, before and after numbers, the failed approaches. Those become the declaration that forces an examiner to engage, and they are hard to reconstruct a year later.
Assume the examiner will have an AI search. Draft as if the ten closest references are on the table from day one, because by the time the application is examined they may well be. Claims that only hold up against an incomplete search will not hold up.
Build for the courts, not just the Office. The friendlier line helps you get a patent. It does not by itself make the patent hold up. Claims aimed at a real technical improvement do both.
If you are deciding whether and what to file on an AI product, the strategy session on the engagement page is built for that question, and it ends with a written plan.
General information, not legal advice. Your own attorney is the one to rely on for your situation.