The Specialised Business Section of the Court of Milan has decided an Italian computer-vision infringement case on a point that will matter to anyone prosecuting or enforcing AI patents: a system that learns to recognise an object is not, in patent terms, doing the same thing as a system that compares what its camera sees against pictures held in memory. In decision No. 7233/2026, published on 22 September 2026 in proceedings RG 9864/2021, the court declared Italian patent IT 1397341 invalid and dismissed the infringement and unfair-competition claims brought against driver-assistance systems built on the EyeQ vision chip — identified in Iusletter’s report of the decision as Mobileye/EyeQ technology. The parties are anonymised in the judgment.
The patent, filed in July 2008 and granted in 2013, claimed a driver-assistance system that acquired images of the vehicle’s surroundings, recognised obstacles — “preferably” using fuzzy logic, and by comparison with a second plurality of sample images stored in a static memory — detected the kinematic parameters of the vehicle and the obstacles, compared a clear area against a passable area, and raised a collision alarm. The proprietor sued the vehicle manufacturer and an Italian group company of the chip designer for direct and contributory infringement under Article 66(2) of the Italian Industrial Property Code, adding unfair competition under Article 2598 of the Civil Code. The defendants counterclaimed for invalidity, and the court appointed a technical expert on both validity and interference.
The counterclaim won. On Iusletter’s reading, no single prior document anticipated every claimed element; the patent fell for lack of inventive step under the problem-and-solution approach, with US 2004/0181338 already teaching the comparison of required against available driving space — the court declining to treat “passable area” and “clear area” as technically distinct merely because the terminology differed, and treating comparison against pre-stored images as an obvious alternative to neural-network recognition that was already known. IPKat’s account of the same decision also records a novelty objection succeeding on one prior document, so read the judgment itself before relying on either framing. Both reports agree the two post-grant limitations filed in 2018 did not save the patent. On infringement, literal infringement was excluded because the accused technology relies on a network trained in advance and performs no constant comparison against sample images and uses no static memory. Equivalence was excluded too: applying the triple-identity, or function-way-result, test, the court held that neural networks and machine learning are a different way of reaching the same result, and that the doctrine of equivalents exists to stop formal circumvention, not to pull known or obviously variant solutions into the monopoly.
The claim language that no longer reaches the product
The operative lesson is about mechanism words. A claim that recites comparing captured images with stored reference images, or a static memory holding samples, describes how the recognition happens. Once that is in the claim, a trained model sits outside it — and outside the doctrine of equivalents as well, because the differing step is the way the result is achieved rather than a swapped-out component. Every pre-2015 machine-vision portfolio is full of exactly this language, drafted when template matching was the obvious implementation. If you are counting on those grants to cover today’s neural-network products, the Milan reasoning says you are counting on the wrong claims. The remedy at drafting stage is to claim the recognition step by its result rather than its mechanism, and to disclose and claim the trained-model route expressly instead of leaving it to a “preferably” clause.
Generic description cuts both ways
The sufficiency analysis is the sharper warning for software and AI filings. The patent named no algorithm at all, and that was survivable: by 2008, assisted-driving literature was common general knowledge, so the skilled person could fill the implementation gaps. The same genericity then destroyed inventive step, because the claim amounted to a functional statement of a technical idea already available, with no algorithmic contribution to distinguish it. Being sufficiently described does not make an invention sufficiently inventive — and the broader the functional language, the more prior art it collides with. IPKat’s report adds a related trap: offering fuzzy logic as a merely “preferred” solution stretched the claim across alternative techniques the description never disclosed. Name the technique you actually enabled.
What to do with this
Three practical moves. First, if you hold legacy computer-vision claims, audit them for stored-reference and comparison limitations before you send a letter, not after a counterclaim lands — an assertion that invites a validity attack on a 2008 functional claim is how a portfolio asset becomes a liability, which is the case for running the invalidity analysis on your own patent first. Second, if you are shipping a machine-learning product into Europe, this decision is a usable data point in a freedom-to-operate assessment: claims built on template matching are weaker reads against a trained network than a bare claim-chart comparison suggests, though the Milan court’s reasoning binds nobody outside Italy and the EPO and national courts apply their own equivalence tests.
Third, the evidence point, which decided more than it should have. Iusletter reports a failure of proof: the proprietor assembled the accused system’s features from documents relating to different products without showing that those features co-existed in any one system. A claim chart stitched together from datasheets, marketing pages and a supplier’s technical notes across model years is not proof that a single accused unit practises every element. One accused configuration, one teardown or one verified specification per element — that is the standard the court applied here, and it is cheaper to meet before filing than to repair after a court-appointed expert has reported. Note finally that this was an Italian national patent, so it stayed with the Specialised Business Sections; the Unified Patent Court’s jurisdiction runs to European and unitary patents, and a parallel national title can still be litigated on these national equivalence rules.
