UK Supreme Court resets the legal test for AI patentability
The UK Supreme Court has replaced the long-standing Aerotel framework with the European Patent Office's Board of Appeal approach for assessing exclusions, including the "program for a computer" bar. The case returns to the UKIPO to decide whether Emotional Perception AI's recommendation system is novel and inventive.
Headline takeaway: an artificial neural network (ANN) plus its weights and biases is treated as a computer program. That does not end the enquiry. The door remains open if you can show a genuine technical contribution.
What changed
- Aerotel's four-step test is no longer the touchstone. The Supreme Court found it mixed the concept of an "invention" with other patentability requirements.
- The EPO's problem-solution approach and its case law on exclusions now set the direction: assess technical character and technical contribution, and only technical features count for inventive step.
- ANNs are treated as programs regardless of implementation detail. Hardware-embedded weights, reconfigurable hardware, or a conventional computer all lead to the same conclusion: you're still dealing with a program.
- The UKIPO must now perform the full novelty and inventive-step analysis on EPAI's application.
Why this matters for patent teams
Expect closer alignment with EPO examination practice and less forum drift between UK and European filings. For cross-border portfolios, this should reduce rework and inconsistent outcomes.
But the bar does not move uniformly. Recommender systems and content classification remain vulnerable unless you can point to concrete technical effects, not just better "semantics" or improved user satisfaction.
The practical test: what counts as a technical contribution
- Show computational-level improvements: throughput, latency, cache/memory usage, numerical stability, precision/recall improvements tied to specific compute operations, or reduced energy per inference.
- Show control of physical processes: sensors, actuators, codecs, radio stacks, storage controllers, or other real-world interfaces.
- Anchor advantages to implementation details: new network topologies mapped to hardware features, quantisation or sparsity schemes that reduce cache misses, kernel-level scheduling, or instruction-set optimisations.
- Provide evidence: benchmarks, ablation studies, profiling traces, and reproducible test data that link claimed features to the technical effect.
Claim drafting checklist
- Frame the problem and effect at a technical level, not at the level of "better recommendations" or "improved user experience."
- Draft both method and apparatus claims; include claims that reflect concrete compute or hardware interactions.
- Structure claims so the technical features carry the inventive weight; let non-technical features sit in the background.
- Prepare fallbacks: narrower claims pegged to the specific mechanism that delivers the technical gain (e.g., cache-aware matrix block size, fixed-point pipeline with bounded rounding error).
- Disclose how training/inference is realised on the machine: memory layout, dataflow, tiling, kernel fusion, or scheduler behaviour.
What this means for recommender systems like EPAI's
The Court accepted that an ANN with its weights is a program, even if "frozen" into hardware. That classification alone doesn't doom the claims. The real fight moves to whether the system delivers a technical contribution beyond producing "semantically similar file recommendations."
For recommenders, steer away from the abstract goal and point to the computing mechanics that get you there-e.g., a similarity search that reduces cache thrash through a specific graph-based index, or a quantisation scheme that stabilises gradients and cuts floating-point error on specified hardware.
Litigation context
This ruling sits alongside the UK's earlier decision in the DABUS litigation, where the Supreme Court confirmed that an AI system cannot be named as the inventor under current law. Together, the cases show UK courts are addressing AI patent law questions at the highest level, while leaving room for the UKIPO to apply technical expertise during examination.
Action items for in-house counsel and outside attorneys
- Recalibrate UK filings to the EPO playbook. Align disclosure and claim structure across jurisdictions to avoid mixed signals.
- Front-load evidence of technical effect. If you have profiler data or ablation studies, put them in the spec and be ready to file them during prosecution.
- Audit pending UK cases anchored in Aerotel reasoning. Identify where COMVIK-style arguments can be strengthened.
- For pipeline-heavy AI (training, deployment, retrieval), map each step to a concrete machine-level effect and cite it clearly in the claims.
Open questions to watch
- How the UKIPO will draw the line between "semantic quality" and a computer-implemented technical effect for AI recommenders.
- What level of empirical proof examiners will expect to credit technical advantages during prosecution.
- Whether future cases will refine how hardware-tied ANNs (e.g., baked-in weights) are treated when the alleged advance is architectural rather than application-level.
Authoritative resources
- EPO Guidelines for Examination (computer-implemented inventions and technical contribution)
- UKIPO Manual of Patent Practice (patentable subject matter and exclusions)
Bottom line: treat your ANN as a program, then earn patentability by proving what it does for the machine. If your story lives at the level of user preferences, you'll struggle. If it lives in how the computer runs, you've got a path.
Your membership also unlocks:
