The UK’s AI Patent Reset: When “Any Hardware” Becomes the New Gatekeeper

by ChatGPT-5.2

On 11 February 2026, the UK Supreme Court handed down a judgment that quietly rewires how AI and software patents are assessed in the UK. The case—Emotional Perception AI Limited v Comptroller General of Patents, Designs and Trade Marks—looked, on the surface, like a niche dispute about whether a particular artificial neural network (ANN)-based recommendation system can be patented.

But the real story is bigger: the Court effectively retired the UK’s long-standing “Aerotel” test for excluded subject matter and told courts and examiners to align much more closely with how the European Patent Office (EPO) understands “invention” under the European Patent Convention. That is a structural change that will ripple through many software and AI patent applications—not just this one.

Below is what happened, why it matters, and what AI stakeholders should be mindful of next.

What happened, in plain language

1) The invention: AI that “learns” similarity without human labels at runtime

The patent applicant said it had created a system that can recommend media files (music, video, text) that feel similar to a user—i.e., likely to trigger a similar emotional response.

The method, simplified:

• Use pairs of files that humans have described in words (so you can estimate “semantic similarity”).

• Also measure objective physical properties of those files (tempo, tone, rhythm, etc.).

• Train an ANN so that, over time, the “distance” between files in property spacestarts to line up with the “distance” between files in semantic space.

• Once trained, the model can recommend “similar” files using only measurable properties, without needing semantic labels at runtime.

So the claimed trick isn’t “ANNs exist” (they do), but a particular training alignment technique to connect physical features to human-meaning similarity.

2) The legal problem: UK patent law excludes “computer programs… as such”

Under the Patents Act 1977 (mirroring Article 52 EPC), certain categories are excluded from patentability, including computer programs “as such.” That phrase—“as such”—is where decades of argument lives.

This case turned on whether an ANN-based invention is:

• not a computer program at all (so the exclusion doesn’t apply), or

• a computer program, but not “as such” (so it might still be patentable), or

• just excluded subject matter dressed up (so no patent).

3) The procedural earthquake: the Court moved the UK away from Aerotel

For years, UK decision-making leaned heavily on a Court of Appeal framework from Aerotel Ltd v Telco Holdings Ltd(“Aerotel”), which used a four-step approach that asked, in effect: what is the contribution, and is that contribution technical?

The Supreme Court said: that approach can’t stand in light of how the EPO’s Enlarged Board has clarified Article 52—especially in Bentley Systems (UK) Ltd/Pedestrian Simulation (G1/19).

In other words: UK practice had drifted into a UK-specific interpretation, and the Court chose re-alignment.

4) The core doctrinal shift: “invention” first, then filter “technical features” for inventive step

The EPO approach (which the Supreme Court embraced) works more like this:

• Step A — Eligibility / invention: If the claim involves technical means (often: “any hardware”), it can qualify as an “invention” for Article 52(1). That hurdle is relatively low.

• Step B — Intermediate filter: Identify which features are technical versus non-technical (and which contribute to technical character).

• Step C — Patentability: Assess novelty and inventive step primarily on the features that contribute technically.

This matters because Aerotel tended to blur “is it an invention at all?” with “is it inventive/technical enough?” The Supreme Court endorsed the idea that these are separate questions that shouldn’t be mashed together.

5) The ANN question: ANNs can be “programs for computers,” but that doesn’t end it

The Court rejected the applicant’s broad claim that an ANN is not a program for a computer. It concluded that ANNs of the relevant kind can fall within “programs for computers.”

But crucially, that does not mean the invention is automatically excluded. The analysis must continue: the real fight becomes whether the claimed subject matter is a program “as such” (excluded), or whether it includes technical features/technical contributions that can support patentability.

6) The practical outcome: sent back for a proper “technical contribution” analysis

The Supreme Court did not simply grant the patent. Instead, it essentially said: UKIPO must reconsider under the corrected framework—particularly whether the features that contribute to the invention’s technical character involve an inventive step.

So: the gate moved. The applicant didn’t “win the patent,” but it did win a new, more EPO-aligned path to argue for one.

What this means for AI stakeholders

For AI inventors and patent applicants: the “eligibility hurdle” may be easier, but the “inventive step” hurdle may bite harder

If you’re building AI products, this decision can feel like a boost: it lowers the odds that an examiner kills your application early by saying “this is just software” under a UK-specific test.

But the trade-off is real:

• You may clear “invention/eligibility” faster.

• You will face sharper scrutiny later: what is the technical contribution, exactly?
  “It improves recommendations” or “it mimics human perception better” may be treated as non-technical aims unless you can show a technical improvement in a technical system.

Bottom line: the action shifts from Stage 1 rejection to Stage 2/3 attrition.

For patent lawyers and drafters: claim drafting must become more explicitly “technical-feature aware”

Under the EPO-flavored approach, you can’t rely on rhetorical “technicality.” You need claims and descriptions that make it easier to argue:

• What the technical problem is (not just a business/consumer problem),

• What the technical solution is (not just “use ML”),

• Which claim features deliver that technical solution,

• How those features interact with the computer/hardware to produce a technical effect.

In practice, you’ll see more emphasis on things like:

• data representation choices,

• signal processing / feature extraction specifics,

• model architecture constraints tied to performance,

• compute/memory efficiency gains,

• latency reductions,

• robustness improvements framed as technical system performance (not just “better recommendations”).

For the UKIPO and courts: expect a period of “doctrine migration pain”

When a country swaps a familiar test for another ecosystem’s logic, you get a messy phase:

• new examiner guidance,

• new argument patterns,

• inconsistent early decisions,

• a rush of “what counts as technical?” disputes.

Even if outcomes converge eventually, the reasoning will change immediately—and that changes how stakeholders litigate, settle, and invest.

For investors and acquirers: patents may become more obtainable, but their strength will depend on how “technical” they truly are

This is where I’d be cautious: a more permissive eligibility stage can inflate the number of granted patents, but not necessarily their defensibility.

So in diligence:

• look for patents that clearly anchor inventive step in technical features,

• be wary of portfolios heavy on “AI does X better” without technical mechanism,

• model your litigation risk around the “technical contribution” narrative.

For open-source and smaller competitors: the risk of patent thickets increases

If more AI/software inventions survive eligibility and reach deeper patentability analysis, you may see:

• more granted patents in crowded ML-adjacent spaces,

• more “defensive” filing behavior,

• more licensing pressure,

• more uncertainty for open implementations.

That doesn’t mean the system becomes pro-patent at all costs—but it can increase the volume of rights asserted, especially in the short term.

Pros and cons of the shift

Pros

• Greater alignment with Europe: closer doctrinal harmony with EPO practice can reduce cross-border unpredictability.

• Less UK exceptionalism: fewer inventions die early due to a uniquely UK framing of “excluded subject matter.”

• More coherent sequencing: separating “is it an invention?” from “is it inventive?” is conceptually cleaner and may reduce category errors.

• Potentially friendlier to AI R&D: ANN-enabled inventions aren’t dismissed just because they are “software-ish.”

Cons

• More complexity and litigation surface area: the “technical vs non-technical features” filter invites granular disputes.

• Risk of more low-quality filings reaching later stages: eligibility becomes easier to satisfy; the system must resist being flooded by vague “AI applied to X” patents.

• Patent thicket dynamics: in fast-moving AI sectors, more patents can mean more blocking positions—especially where foundational techniques are broad.

• Transitional uncertainty: until UKIPO guidance and case law settle, stakeholders will face inconsistent application.

ChatGPT’s perspective: this is a governance decision as much as a patent decision

I read this judgment as a strategic alignment move: the UK is choosing interoperability with the European patent ecosystem over maintaining a distinct domestic test that the EPO has explicitly criticized.

That’s rational. Patent systems are infrastructure. When infrastructure diverges, everyone pays a friction tax—applicants, examiners, litigators, investors. The Court is trying to reduce that tax.

But the other side is also obvious: if “eligibility” becomes a lighter touch, the system must become more disciplined downstream—or you end up with a formal “invention” gate that doesn’t do much filtering, and a market that must do the filtering through litigation, licensing brinkmanship, and consolidation.

So the success condition here is not “more AI patents.” It’s: clearer, more consistent identification of technical contribution that keeps patents tied to genuine technical advances—rather than letting “AI + desired outcome” become a rent-extraction pattern.

Recommendations: what stakeholders should do next

1) If you’re building AI products: document technical improvements like an engineer, not a marketer

Keep internal records that show:

• measurable compute/memory/latency gains,

• robustness and reliability improvements in system terms,

• specific architectural or data-processing choices,

• why those choices are non-obvious.

Those artifacts become your best ammunition for inventive step.

2) If you’re filing patents: design claims around “technical features that carry the invention”

Aim for claim sets where:

• the technical mechanism is explicit,

• the contribution isn’t just a target outcome (“better recommendations”),

• the invention can be framed as solving a technical problem in a technical system.

3) If you’re a rights/licensing stakeholder: expect more patents and sharpen FTO (freedom-to-operate)

Do more proactive scanning in AI-adjacent domains where you ship products, especially where recommender systems, embeddings, feature-space mappings, and model training workflows overlap.

4) If you’re acquiring AI companies: treat patent diligence as “technical contribution diligence”

Don’t just count patents. Test:

• how they survive a “technical vs non-technical” feature analysis,

• whether they have credible inventive-step stories anchored in technical mechanisms,

• whether the specs read like engineering or like outcome-claims.

5) If you’re policy-minded: watch whether the UK develops a clear, stable notion of “technical contribution” in AI

The next chapter won’t be about “ANNs are/aren’t programs.”
It will be about where UK practice draws the line between:

• genuine technical system improvements, and

• non-technical goals implemented via generic computing.

That line is where innovation incentives, competition, and public interest will be won or lost.