Insight
For the last decade, the story of AI patent prosecution has been a Section 101 story. Alice. Mayo. Abstract ideas. Inventive concepts. If your claim survived that gauntlet, you assumed you were mostly through the hard part. That assumption is now out of date, and the data says so.
Since the USPTO's Ex parte Desjardins decision was made precedential on November 4, 2025, examiners have been quietly moving the fight. Section 101 rejections haven't disappeared, but Section 112 — written description — is doing more of the work than it used to, and AI medical device applications are squarely in the path of that shift.
Desjardins is the decision that found a Google machine learning training claim patent-eligible, reversing a PTAB rejection by holding that a genuine improvement to how a model itself operates can satisfy Alice step two. That part of the story is good news for AI applicants, and it's the part that got the most attention when the decision came down.
But the panel didn't stop at eligibility. It told examiners, in plain terms, where to focus instead. The decision states that Sections 102, 103, and 112 are the traditional tools for keeping patent protection within its proper bounds, and that examination should center there. That is not a throwaway line. It is a redirection, and the data since November shows examiners took it seriously.
Patent prosecution data tracked from March 2025 through March 2026 shows a clear inflection point around the precedential designation. Before November 4, 2025, Section 112 written description rejections ran at roughly 15% of the volume of Section 101 rejections, across all applications. After that date, the ratio climbed to roughly 22% — an increase of about 47% in how often 112 shows up relative to 101.
Narrow the data to AI-related applications specifically, and the same pattern holds, though the baseline sits differently. Pre-Desjardins, Section 112 rejections ran at about 5% of Section 101 volume in AI cases. Post-Desjardins, that climbed to roughly 7% — a 35% increase. The two datasets tell the same story from different angles: examiners are leaning on written description more, and AI applications are part of that trend.
None of this means Section 101 rejections went away. It means the ratio shifted, and the direction is consistent with what Desjardins explicitly asked for.
The shift isn't abstract. Recent office actions give a concrete picture of what "insufficient written description" looks like when applied to an AI claim. Examiners have identified specific categories of missing disclosure, including model architecture, the number and type of layers, how data moves through the model, the decision logic used to reach an output, weighting details, whether techniques like regression are involved, how loss is minimized, and how the model resolves a given computational problem.
The underlying objection is consistent across these rejections: a specification that describes what a model does — classifies, predicts, flags, scores — without describing how it does it mechanically is being treated as a black-box disclosure. Examiners are asking whether the inventors actually possessed the claimed implementation as of the filing date, not whether a skilled engineer could eventually build something that works. Those are different legal questions, and Section 112 asks the second one with more teeth than it used to.
It would be easy to read this as a general software prosecution issue and assume medical device applicants are somewhat insulated, since the field has long been used to detailed technical specifications for other reasons. The opposite is closer to true.
AI medical device claims tend to describe clinically meaningful outputs in detail — sensitivity, specificity, risk stratification, time-to-detection — because that language is what regulators and clinicians care about. That is precisely the kind of outcome-focused language that reads as a black box to an examiner applying post-Desjardins scrutiny. A specification that thoroughly explains what a sepsis prediction model achieves clinically, without explaining how the model's architecture, training process, or inference pipeline gets there, is now a more exposed target than it was a year ago.
There is a silver lining here, and it connects directly to something AI medical device applicants already have on hand: FDA documentation. A 510(k) submission, a validation study, or an algorithm change protocol prepared for Software as a Medical Device clearance typically already contains the layer of mechanical detail a post-Desjardins examiner is looking for — training data composition, preprocessing steps, model architecture, performance validation methodology. That documentation was underused for patent drafting before; it is now closer to a requirement than a nice-to-have. If the specification and the regulatory submission are drafted by different teams working from different vocabularies, this is the moment to close that gap.
When you or your attorney reviews a specification for an AI medical device claim, check it against categories examiners are now flagging by name.
Model architecture. Does the spec describe the actual structure — layer types, layer count, how they connect — or just name the model family ("a neural network," "a convolutional model")?
Data flow. Does the spec explain how patient data moves through the system, from raw input through preprocessing to the point where the model consumes it?
Decision logic. Does the spec describe the mechanism by which the model reaches its output, or only the output itself?
Training methodology. Does the spec explain how the model was trained — loss function, weighting scheme, how sparse or imbalanced clinical data was handled — or does it simply assert that the model "was trained on patient data"?
Inference and post-processing. Does the spec walk through what happens between the model's raw output and the claimed clinical result — thresholding, calibration, post-processing steps — or does it jump straight from input to conclusion?
Alternative implementations. Does the spec offer more than one way to achieve the claimed function, or does it describe a single, narrow embodiment that could later look like the only thing the inventors actually possessed?
A specification that answers these categories in narrative form and backs it up with a flowchart or pseudo-code where the mechanism is genuinely novel is built for both sides of the current landscape — the eligibility fight under Alice and Recentive, and the possession fight now more likely under Section 112.
Desjardins was covered, correctly, as a win for AI patent applicants on Section 101. What's received less attention is that the same decision handed examiners a new instruction: stop leaning on abstract-idea rejections and start testing whether the specification proves the applicant actually possesses what the claims describe. For AI medical device applications in particular, where the tendency is to write in the language of clinical outcomes rather than engineering mechanism, that instruction lands directly on a common weak spot.
The practical response is not new advice so much as a reason to take old advice more seriously. Write the specification like an engineer, not a marketer. Say how the model works, not just what it achieves. The teams that already have this habit — often because they've been through an FDA submission and are used to disclosing mechanism in detail — are the ones best positioned for what comes next in prosecution. The teams that don't are about to find out the hard way that "the model detects the disease" was never a description of an invention. It was a description of a result.