What Recent Patent Decisions Add Up To

The Same Demand, Arrived At Three Times

Three different doctrines, decided under three different statutes by different courts, converge on one demand: the patent has to say how.

• Eligibility. Claims reciting what a system is "arranged to allow," without disclosing how, were held abstract. 7,679,637 v. Google, GoTV v. Netflix, Recentive v. Fox.
• Enablement. A specification that hands the reader a search procedure rather than the answer was "little more than two research assignments." Amgen v. Sanofi.
• Written description. Two examples of a vast functional class did not show possession of the class. Juno v. Kite.

These are independent requirements. Each can invalidate a claim on its own, and satisfying one says nothing about the others. What they share is the shape of the failure: describing the destination rather than the route.

The Improvement Has To Be In The Claims

Across two forums, the same rule has now appeared four times.

• "It cannot be said that the claims are directed to a technological improvement when nothing in the claims requires the steps necessary to make the improvement." Columbia v. Gen Digital.
• "only features that are claimed, not unclaimed details that appear in the specification, can supply something beyond ineligible matter." GoTV v. Netflix.
• And at the Patent Office, an assertion in the specification "alone is insufficient to support a patent eligibility determination, absent a subsequent determination that the claim itself reflects the disclosed improvement." Ex parte Desjardins.
• "reliance on the specification must always yield to the claim language," in a case where claims to specific constellations survived and claims to a constellation "optimized" for a result did not. Constellation Designs v. LG.

This is where an engineering review goes wrong. An engineer reads the specification, recognizes the work, and approves it. Columbia lost a $185 million judgment because its real advances, selective emulation and model diversification, were described in the patent and not required by the claims.

There is a genuine tension in this line of cases and it is worth knowing about. 7,679,637 v. Google says a claim "need not explicitly recite the improvement," which reads against Columbia's statement that reliance on the specification "must always yield to the claim language." GoTV cites both propositions in a single sentence: the specification informs what the claimed combination is directed to, and only claimed features supply eligibility.

The Earliest Obligations Arrive Before Anyone Is Thinking About Patents

Three decisions locate the fatal error years before litigation, in documents nobody wrote with a patent in mind.

• Ownership. "Shall be the property of" was a promise to assign, not an assignment, and the company relying on it did not have what it thought it had. Omni MedSci v. Apple.
• Inventorship. A supplier's engineer contributed, was left off, could not be located years later, and two patents were invalidated. Fortress Iron v. Digger.
• Timing. Selling what a secret process makes started the one-year clock on patenting the process. Celanese v. ITC.
• Prior art dating. A reference reaches back to its own provisional only if that provisional supports one of its published claims. Dental Monitoring v. Align.

None of these is a litigation mistake. They are a contract template, a vendor relationship, and an invoice. That is the whole reason this section exists: these are the cases where the loss was already locked in by the time a lawyer was in the room.

AI Has One Settled Rule And One Very Large Open Question

Settled: an AI system cannot be named as an inventor (Thaler v. Vidal). Applying generic machine learning to a new field is an abstract idea, and the Supreme Court declined to review that holding (Recentive v. Fox). Improving how the model itself operates can be eligible, at least at the Patent Office (Ex parte Desjardins).

The Patent Office has since answered the operational half of that question by declining to create a special rule. Its guidance at 90 FR 54636, published November 28, 2025, rescinded the February 2024 version and confirmed that the Pannu joint-inventorship factors "only apply when determining whether multiple natural persons qualify as joint inventors." A lone human inventing with AI help faces no joint-inventorship question at all, only conception, which the guidance calls "the touchstone of inventorship." That is agency guidance, so it governs examination and binds no court.

There is also a split between forums. Recentive sets a ceiling the Federal Circuit enforces and the Supreme Court left standing. Desjardins opens a path that binds examiners but not courts, and the USPTO's own memorandum says its updates were "not intended to announce any new USPTO practice or procedure," which contradicts the widespread reading of it as a liberalization. Where an applicant meets the examining corps, Desjardins governs. Where a patent is litigated, Recentive does.

Winning The Doctrine And Winning The Case Are Different Events

Four decisions here show a party prevailing on the legal question at issue and losing anyway.

• Ollnova won step-one eligibility on two of its four patents, had a third remanded for the step-two analysis, and lost its entire $11.5 million verdict on jury-instruction and verdict-form defects.
• GoTV won reversal of an indefiniteness ruling, and the court used the resulting construction to invalidate the patents.
• Columbia won on claim construction, infringement and willfulness, and lost the judgment on eligibility and foreign damages.
• The Desjardins applicant won on eligibility, and the claims "stand rejected under § 103."

This is why "the patentee won on eligibility" is a misleading headline. Eligibility is one gate among several, and clearing it buys you the next one.

Three Pairs That Are Never Published Apart

In each of these, the first decision creates an exposure and the second bounds it. Reading either alone produces a confident belief that is wrong, in a direction that costs money.

• IPR estoppel. Caltech and Ingenico. Alone, the first says inter partes review is a trap and the second says estoppel barely matters.
• Patent term. Cellect and Allergan. The most consequential pair here to get wrong, because the reflexive response to the first alone is a terminal disclaimer, and once a patent issues subject to one the surrendered term is not recoverable. Ex parte Baurin is the third piece: a precedential agency panel reading Allergan narrowly enough that it "will rarely, if ever, arise in original examination."
• Willfulness and enhancement. SRI and Columbia. Together: a jury decides whether you infringed on purpose, a judge decides separately whether to punish you, and what the judge weighs is largely how you behaved after the suit was filed.

Each of those pages carries its companion above the holding, not in a footer, for exactly this reason.

Two Exposure Surfaces Software Companies Underrate

• Where copies are made. For downloadable software distributed internationally, the copy created on a customer's machine abroad was "made outside the United States," and on those claim types could not support domestic damages. Modeling patent exposure on worldwide revenue and modeling it on where copies are created are different exercises. Columbia, with Brumfield as the necessary other half.
• What your documentation says. Inducement is how a patent holder reaches a company whose customers do the infringing. The Supreme Court narrowed it: the question is whether the defendant "actively encouraged infringing uses, not merely whether doctors could plausibly read the alleged statements as instructions to infringe." Omissions do not count and ordinary distribution does not count. But encouragement need not be express, and a company can induce "through implicit encouragement" that is clear and affirmative. Hikma v. Amarin.

A third, cheaper than either. Marking. If you have licensed your patent, your licensees have to mark too, and a patentee that pleaded nothing about eleven settlement licences lost its pre-suit damages entirely. VDPP v. Volkswagen.

An Observation About What Is Missing

This is our reading of the corpus rather than anything a court said, and we are flagging it as such.

The decisions here that help a company being sued are largely procedural: estoppel boundaries, expert admissibility, pleading standards. The decisions that reshape what can be patented, and therefore what a company can own, run mostly against breadth. Someone arriving with "how do I stop getting sued" will find less in these twenty-seven cases than someone asking what it takes to have something worth owning.

That asymmetry is narrower than it first looks. Columbia requires a district court to weigh the defenses a defendant lost, holds that repeating an argument is not litigation misconduct, and treats the appellate court's own doubt about eligibility as evidence the case was close. SRI confirms that enhancement is a separate discretionary decision that a willfulness finding does not compel. Those are substantive defensive holdings, not procedure. The asymmetry is real and it is an observation about which cases got decided, not a claim about the law.

Every Decision, Newest First

One row per decision on this site, from the catalog that drives each page, so a decision added anywhere in the library appears here without anyone editing this table.

The Library

Every case above has its own page, grouped by the question that brings a reader to it.

• Can I patent this software or AI thing? Six cases on section 101 eligibility, including the Patent Office decision that diverges from the courts.
• Who owns it, and who invented it? Three cases where the answer was decided by a contract template or a vendor relationship.
• When must I file? Two cases on timing, both of which land inside ordinary commercial decisions.
• How broadly can I claim? Enablement and written description, which fail broad claims for different reasons.
• Patent term and double patenting. The pair where reading one half alone permanently costs you term.
• If you are sued, or you sue. Nine cases on estoppel, willfulness, damages, prior art and injunctions.
• LKQ v. GM. Design patent obviousness, reset, in both directions.

These are decided cases. The statutes and examination guidance underneath them are covered separately, for someone earlier in the question: is it obvious?, what counts as prior art?, how your application is put together, and software patents in Europe and the UK.

Alongside these, a separate section covers USPTO practice and procedure, written for registered practitioners. It reports what the Office and the Board require rather than what a court held, and every page there states its authority level, because agency guidance and binding precedent are not the same thing.

• Before the PTAB. Institution is the Director's decision as a whole, and there is no legal right to it.
• Filing and fees. Entity discounts, IDS size fees, AFCP 2.0, and who may sign what.
• Duties, and who may practise. Candor, reasonable inquiry, AI tools before the Office.

What This Page Does Not Do

• It does not say what you should file, disclaim, narrow, broaden, assign or document. Nothing on this page is an instruction.
• It does not assess any patent, application, contract or product.
• It does not predict outcomes. Every case here is a set of facts a court already had in front of it.

Educational, not legal advice. What courts required, and what happened to the companies that got it wrong, is what this page reports. Whether any of it applies to your situation is a determination a qualified patent attorney makes with your facts in hand.