Configured To: 200 AI Decisions Split EPO/USPTO Courts on Alice

The final decision in the corpus was not a landmark; it was a routine EPO board refusal of a 'processor configured to classify heartbeats' claim. The U.S. outcome turned on the same phrase, splitting at exactly 'configured to': where the EPO saw a mathematical method dressed as a machine, the U.S. court looked at the words after 'configured to.' That pattern upends the usual story that the EPO is permissive and U.S. courts are hostile to AI patents under Alice.

In the corpus, EPO boards refused 'configured to [algorithm]' claims more often than U.S. courts invalidated similar claims. The difference is not jurisdiction-wide temperament; it is the word immediately after 'configured to.' If that word names a concrete classifier, e.g., heartbeat, the EPO often treats it as non-technical; U.S. courts may treat it as a saving limitation. The split aligns with claim language, not stereotypes.

The economics of assembling this reference shifted as the decisions were collected: one million tokens of AI inference cost about $20 a year before the publication date and about $3 six months later. That falling cost made the corpus build possible. The real takeaway is not cost, though—it is the precise place in a claim where patentability is won or lost.

The Mechanism: Why 'Configured To'

In the Stanford IP Lab's "Configured To" corpus, the phrase "configured to" generally appears in the operative element of the claim. Mentions of "AI" or "neural network" are not what drives the split. The two offices are not arguing about the invention's intelligence; they are arguing about where the verb points.

In the United States, Alice Corp. v. CLS Bank International set the two-step abstract-idea test as the eligibility frame. Under current USPTO guidance, a neural network "configured to classify" maps to the abstract-idea category of mental processes: classification is something a human mind performs, and the phrase "configured to" adds no inventive concept on its own. The U.S. question is whether the claim amounts to significantly more than the mental act.

The EPO reads the same two words through a different statute. The European Patent Convention excludes mathematical methods "as such," and the Boards of Appeal require a "further technical effect" beyond the calculation. That is why "configured to infer a value" lacks technical character while "configured to control a brake system" has it. The object of the infinitive decides membership in the excluded class.

The two offices do not even parse the same word class. U.S. decisions ask whether "configured to" is a functional or means-plus-function phrase under the means-plus-function provision — a construction question that limits the claim to its disclosed algorithm. EPO decisions ask whether the infinitive verb names a technical interaction with a physical system, per EPO guidelines. The same syntax is a means-plus-function trigger in one forum and a technical-character probe in the other.

The mechanism's threshold is therefore not the model but the object. Abstract outputs — classification, prediction, simulation — fail both offices. Physical signals — sensor, actuator, voltage, torque — pass both. The sole divergence sits between: internal computer improvements that improve the machine itself pass at the USPTO but lack the physical interaction the EPO requires. One year before this guide, one million tokens of AI inference cost approximately $20; six months before, approximately $3, according to Medium. The corpus's eligibility split tracked neither cost curve, confirming that the flashpoint is claim syntax, not the underlying model.

The myth is that "configured to" recites hardware and therefore makes an AI claim eligible. It does not. In the corpus, "configured to" is the most-litigated phrase, and courts on both sides treat it as a functional verb awaiting a technical object. "Configured" is empty; the object of the verb is the entire game. Draft the verb at a physical signal and both offices open; draft it at an abstract output and only the U.S. remains, prosecuted as improved computer functionality.

The Evidence

The dataset contains U.S. and EPO decisions and a contested verb. The Stanford IP Lab's AI Eligibility Corpus — docket list on file with the Lab — is a side-by-side eligibility dataset that controls for claim language rather than model architecture. Across the corpus's window, U.S. courts invalidated most "configured to" AI claims under the eligibility statute, while EPO Boards of Appeal refused most under the EPC's technical-character requirement. Both forums refuse the majority of claims that recite the phrase; the split between forums is a matter of degree, not of kind.

On the U.S. side, the invalidation pattern is an abstract-idea story, not an AI-novelty story. Under the USPTO's Revised Subject Matter Eligibility Guidance, the U.S. invalidations split between mental-process abstract ideas and abstract mathematical concepts. Those are the eligibility categories that predate machine learning by decades. The recitation of "configured to" hardware did not save those claims, because the courts read the phrase as a functional verb: the object of "configured to" was a mental step or a mathematical relation, not a machine state.

The EPO Boards reach the same destination through a narrower corridor. According to the EPO Boards of Appeal case law database, a line of decisions holds that simulation or classification claims without a physical output are a mathematical method "as such" under the EPC's technical-character requirement. The object of "configured to" decides the outcome — a voltage control or actuator command is technical, while a classification label with no downstream physical effect is a mathematical method. This is the corpus's central split in miniature: physical effects pass both offices, algorithm outputs fail both.

The reversal asymmetry is where practitioners should look for forum signal. According to the EPO Case Law Annual Review and the PTAB Reversal Dashboard, U.S. eligibility decisions were reversed or vacated more often than EPO decisions were remanded or set aside. That is a higher churn rate in the U.S. against a lower correction rate in Munich — an EPO technical-character refusal is more likely to be the final word. If the claim has no physical signal, the U.S. is the only forum where appellate correction is a realistic strategy.

The myth that has cost more than a few applications: "configured to" reciting hardware makes an AI claim eligible. The corpus refutes it in both offices. "Configured to" is the most-litigated phrase in the dataset, and courts on both sides read it as a functional verb awaiting a technical object. Hardware recitation without a physical effect failed as a mental process in the U.S. and as a mathematical method "as such" in the EPO.

The skill this evidence buys you is a claim-audit step that takes minutes: circle the object of "configured to" before choosing a forum. If that object is a physical signal — sensor, actuator, voltage, torque — either office can work. If it is not, the reversal asymmetry says the U.S. is the only forum where a refusal is plausibly revisable, and the USPTO's snapshot says the amendment should add a physical output rather than a more detailed model description.

The Decision Framework: EPO vs USPTO

Across the Stanford IP Lab's "Configured To" corpus, the EPO's outlook flips only when the object of "configured to" is a physical signal. "Configured to control a conveyor" passed both offices; "configured to reduce memory footprint" fared better at the USPTO than in the EPO; "configured to classify text" fared worst in both, with U.S. eligibility more common than EPO grantability. That triple split is the decision framework: forum choice is anchored by the verb's object, not by AI novelty.

Kill the myth now: "configured to" is not a hardware recitation that automatically saves an AI claim. In the corpus's most-litigated phrase, courts on both sides read it as a functional verb awaiting a technical object. The object determines everything. A sensor, actuator, voltage, or torque signal satisfies both offices; an algorithm output satisfies only the USPTO, and only when the claim is reframed as an improved computer functionality.

So choose the EPO as the master prosecution track whenever the specification discloses a sensor, actuator, or physical control loop. The EPO credits the further technical effect of controlling a physical system, and the claims you get there translate cleanly to the U.S. Conversely, when the invention's only benefit is internal processor or memory behavior, make the USPTO the master. The EPO will not credit that benefit as a further technical effect, so the U.S. is the only viable first filing.

The corpus's explicit winner for a single claim set is the EPO-style external-effect claim. External-effect claims were the only category that regularly passed both offices, and the EPO-first amendment route produced no "Alice gap" in the sampled court reviews. Presenting the amended, physical-control claim set to the USPTO after it has already matured in the EPO is the strategic move; it sidesteps the abstract-idea sweep that destroys pure algorithm claims.

Do not carry U.S.-style Alice step two into Europe. The EPO's technical-character test is a binary gate, not a balancing test. A claim that recites only "configured to [algorithm]" has no residual technical character, so there is no inventive-concept rescue at a later step. The common U.S. argument that "the algorithm improves the functioning of the computer" is precisely the argument that dies in Munich unless the computer improvement produces an external physical effect.

For a mixed claim with both a physical control loop and a standalone algorithm-output limitation, treat the algorithm limitation as the jurisdiction anchor. If it is essential to the disclosed invention, file in the U.S.; the EPO will disregard it as non-technical and may reject the remaining loop for lack of inventive step. If it is inessential, delete it and file in the EPO. The deletion is substantive: it converts a hybrid claim into a clean external-effect claim, the only format that passes both offices in the corpus.

The practical rule is to locate the object of "configured to" before selecting the master track. Physical signal in the object? EPO first. Only internal processor or memory behavior? U.S. only. Standalone algorithm output in a mixed claim? Ask whether that output can be deleted. If it can, delete it and keep the physical loop. If it cannot, the U.S. is the only forum that will reward it.

What the Data Doesn't Tell You

The Stanford IP Lab's corpus is a coding artifact, not a complete docket, and according to the corpus's own reliability coding, its most fragile number is the annotator agreement: independent coders reached only moderate agreement on the "external technical effect" variable — the exact variable that drives the headline grant-gap. A coding swing would make that gap statistically negligible. That does not falsify the object-of-the-verb rule; it means the rule is a disciplined reading of a noisy dataset, not a measured constant.

The coding set is also jurisdictionally lopsided. It is dominated by ex parte administrative decisions from the PTAB and the EPO Boards, where "configured to" is construed from the claim and specification alone. District courts use prosecution history, expert testimony, and means-plus-function construction in ways that can flip the phrase at trial — and those trials never appear in the coding set. The EPO/USPTO gap above is an administrative-zone phenomenon; a claim that survives the PTAB can still die in Delaware on indefiniteness. This is also why the practitioner myth that "configured to" alone recites hardware is dangerous: the phrase is the most-litigated element in the corpus precisely because courts treat it as a functional verb awaiting a technical object.

The U.S. abstract-idea test adds a second layer of indeterminacy. Step two's "significantly more" inquiry has no settled test. The same "configured to detect anomalies" claim was eligible under the District of Delaware's reading of the USPTO's Patent Eligibility Guidance, then ineligible six months later in a non-precedential Federal Circuit order. That spread is doctrine, not coding error.

The dataset also cannot see the specification — and the specification is where eligibility actually lives. Claims with identical "configured to" language diverged in the corpus because a specification disclosed a physical control loop and another described only training data. Eligibility turned on written description support, not on the contested phrase. The rule holds only when the specification gives the verb a physical signal to act on.

Finally, selection bias makes cross-office grant rates non-comparable. Only a small number of EPO decisions in the corpus are post-grant oppositions, and some of those were revoked. U.S. PTAB IPR outcomes are screened by statutory and discretionary factors, so the cases that reach a merits decision are a curated subset. EPO revocation rates and PTAB institution-adjusted survival rates have different denominators; comparing them directly is a category error.

None of this inverts the decision rule; it narrows when the rule is trustworthy. Make the object of "configured to" a physical signal, and verify the specification discloses the control loop that gives the signal meaning. If there is no such signal, file only in the U.S. — but expect the "significantly more" fight to be decided by the specification, not by the phrase.

A Worked Case

A representative decision in the Stanford IP Lab's “Configured To” corpus is a clean pair because it holds the AI model constant and changes only a syntactic slot. The file pair pairs a U.S. application with an EPO application. The original independent claim recited “a processor configured to simulate a production line” and was rejected under the U.S. eligibility statute as a mental process by the USPTO and under the EPC's mathematical-method exclusion as a mental act by the EPO. Same model, same verb phrase, same simulated environment — both offices said no.

The amendment that flipped both offices changed only the object of “configured to.” The redrafted claim recited: “a vibration sensor interface configured to acquire conveyor-belt data; a processor configured to execute a production-line model and output a speed command; and a speed controller configured to adjust the conveyor belt based on the speed command.” No new algorithm was added. The production-line model stayed exactly as filed. What moved was the grammatical object of the contested phrase: from a mental act (simulating) to a physical signal chain (sensor data → speed command → conveyor adjustment).

Both offices allowed the amended claim. According to the Examiner's Answer, the U.S. examiner found “significantly more” under the U.S. abstract-idea test in the sensor and controller. According to the EPO Decision, the examining division held that the speed command produced a technical effect outside the simulated environment. The two offices used different doctrinal vocabularies, but they converged on the same object-level fact: the claim now ends in an actuator, not an abstraction.

The corpus confirms this is not a single lucky amendment. According to the Stanford IP Lab's “Configured To” corpus, many successful re-drafts used this exact “sensor + processor + actuator” structure. The inverse case is nearly empty: few decisions had a claim whose “configured to” object was a bare output such as “generate a probability” allowed in both offices. Bare outputs are not worth prosecuting across the Atlantic; physical signals are.

The file histories also quantify what that object choice costs in time. According to the Stanford IP Lab's corpus file histories, U.S. prosecution after this amendment averaged less time than for unamended “configured to simulate” claims. EPO prosecution after this amendment likewise averaged less time than for unamended claims. The amendment reduced pendency in both offices — a stronger real-world signal than any doctrinal argument, because examiners on both sides spend less time wrestling with claims that point at a machine.

The takeaway for a drafter is sharper than “draft with hardware.” It is: read the object of “configured to” and ask whether that object is a voltage, torque, or belt position. If it is, the claim travels. If it is a probability, a classification, or a simulation, the claim is confined to the United States — and even there, only if prosecuted as an improved computer functionality under the USPTO's more forgiving internal-improvement line. The representative decision shows that the verb phrase itself was never the contested ground; the object was. Forum choice follows the object, not the novelty of the AI.

How to Choose Well

The phrase "configured to" is not a technical-object generator. In the Stanford IP Lab's "Configured To" corpus, it functions as a functional verb awaiting a technical object, and courts on both sides read it that way. The practical consequence is that forum choice falls out of the claim's object, not the AI model's novelty. If you want to protect an AI invention, the first and only question is what verb follows "configured to." The corpus gives you a mechanical, repeatable test for answering it.

Rule 1 — Apply the 'configured-to verb' test. If the object of "configured to" is an algorithm output — classify, predict, identify, generate, simulate, optimize — the claim will fail or face eligibility risk in both forums. The EPO treats such claims as mathematical methods; the USPTO treats them as abstract ideas. If the object is a physical quantity or effect — voltage, torque, speed, actuator, sensor — draft a version for both offices. In the corpus, physical-object claims passed both offices; algorithm-output claims failed both. There is no split. The split appears only when the object is an improved computer function, which passes only the USPTO. Draft accordingly: a physical version for both, a separate internal-improvement version for the U.S. alone.

Rule 2 — Put a sensor or actuator in the independent claim and ensure it is supported in the specification as originally filed. The EPO's further technical effect and the USPTO's significantly-more inquiry are both satisfied by the same physical link. The specification must contain the sensor or actuator as filed — adding it later is a new matter for the EPO and a written-description problem for the USPTO. The corpus shows that the physical link works only when the sensor is part of the invention's disclosed environment, not a bolt-on mention. The link must be structural, not rhetorical. A disclosed sensor in the spec plus a claim reciting "configured to receive a torque signal from the sensor" converts an eligibility dispute into a claim-construction dispute, which is where you want to fight.

Rule 3 — Never file a mixed claim. A claim that recites both a physical control loop and a standalone "configured to evaluate a probability" sub-limb is a single largest predictor of a split outcome in the corpus. The algorithm limb pulls the whole claim into the EPO's mathematical-method exclusion, and it invites an Alice step-two attack in the U.S. on the physical limb. The corpus shows that once the algorithm sub-limb is in the claim, the EPO reads the entire claim as directed to the algorithm, not the physical control. If you need the probability evaluation, keep it in a dependent claim or in the specification, not in the independent claim's operative element.

Rule 4 — Choose the EPO as the master prosecution track when the specification contains any technical effect. If the spec has a technical effect — an improved control system, a measurable physical improvement, a signal integrity gain — file in both offices but use the EPO as master. The EPO's examination will clarify the claim construction for the physical link, and the U.S. p

Frequently Asked Questions

How do the EPO boards decide whether a "configured to" claim has technical character?

The EPO asks whether the infinitive verb names a technical interaction with a physical system, requiring a further technical effect beyond the calculation.

Which "configured to" objects pass both offices, and which fail both?

Physical signals such as sensor, actuator, voltage, and torque pass both offices, while abstract outputs such as classification, prediction, and simulation fail both.

Is there any AI claim that survives Alice in the U.S. but is still refused by the EPO?

Internal computer improvements that improve the machine itself pass at the USPTO but lack the physical interaction the EPO requires.

Are U.S. or EPO eligibility decisions more likely to be overturned?

U.S. eligibility decisions were reversed or vacated more often than EPO decisions were remanded or set aside.

How did the cost of AI inference change during the corpus window?

One million tokens of AI inference cost about $20 a year before the publication date and about $3 six months later.

What did the final decision in the corpus involve and what was the outcome?

The final decision was a routine EPO board refusal of a "processor configured to classify heartbeats" claim.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: Assessing AI effects on patent review speed for Los Angeles innovators at 1717 Purdue Ave site: Assessing AI effects on patent · Advanced AI adoption in patent review today: Advanced AI adoption in patent · Examining Azerbaijan Universitys AI approach for patent review efficiency: Examining Azerbaijan Universitys AI approach