Amgen Enablement Lessons: Disclosure Depth vs. Claim Breadth

The Amgen Mechanism

Amgen v. Sanoi established a mechanical constraint that now strangles AI enablement: the specification must teach a person of ordinary skill in the art (POSITA) to make and use the claimed invention without undue experimentation across the entire scope of the claims, not merely at the point of disclosure. The Supreme Court rejected the argument that a "roadmap" sufficed when 26 claimed PCSK9 antibodies were enabled by only three disclosed structures. The holding forces a direct mapping onto machine learning specifications. A filing that discloses one trained transformer—complete with weights, a single dataset, and one loss curve—while claiming "a model configured to perform X across any domain" creates the identical scope-to-disclosure gap as the 26-antibody failure. In this context, retraining on new domains is the functional analogue of isolating new antibodies; the POSITA cannot reproduce the full scope without engaging in the very experimentation the patent law forbids the inventor from shifting onto the public.

The legal mechanism enforcing this constraint is the Wands test, derived from In re Wands (858 F.2d 731, Fed. Cir. 1988), which examiners now apply rigorously to AI specifications under the post-Amgen framework. The Federal Circuit's factors require analysis of the quantity of experimentation needed, the amount of guidance provided in the specification, the presence of working examples, and the predictability of the art. According to IP Boutique Law (March 2026), an enablement rejection challenges whether a third party could reproduce the invention without undue experimentation, while a written description rejection challenges whether the specification demonstrates inventor possession of the claimed invention. Under Wands, a thin AI spec fails on multiple fronts: it provides minimal guidance for generalization, lacks working examples across the claimed domain space, and operates in an unpredictable art where the quantity of experimentation required to reach every embodiment is excessive. Commercial success or FDA approval does not insulate a patent if the specification forces skilled artisans into trial-and-error discovery across the full claimed scope, as noted in JDSupra/Knobbe post (2026).

The 2026 examination cycle serves as the practical cutoff for these disclosures. USPTO examination guidance and the 2024 Federal Register AI guidance set the disclosure baseline examiners apply to applications entering the 2026 examination queue. Applications filed in 2024-2025 with thin AI specifications will hit this tightened standard at their first office action. According to LegalClarity (July 2026), patent rules flow from Title 35 USC, 37 CFR, and the Manual of Patent Examining Procedure (MPEP), administered by the USPTO. The 2024 USPTO Guidance Update explicitly addresses how AI-related inventions are evaluated, reinforcing that the Alice/Mayo two-step test (§ 101 eligibility) is distinct from the enablement requirement (§ 112); relying solely on § 101 arguments ignores the fatal enablement defects exposed by Amgen. Purely theoretical concepts with no identified real-world practical application fail the utility requirement under § 101, but even practical applications fail if they lack reproducible disclosure.

Drafting failure modes are quantified by the behavior of AI-assisted drafting tools. Large language model drafting pipelines used by filers generate claim language that generalizes beyond the specification's working examples at scale. A drafter can produce a 40-claim family from one working example—a ratio precisely penalized by Amgen. This automation amplifies the scope-to-disclosure gap, creating families where the breadth dwarfs the teaching. According to IP Boutique Law (March 2026), the written description requirement mandates demonstrating that the inventor actually possessed the claimed invention at the time of filing. Best mode requires disclosure of the inventor's known preferred embodiment at filing; intentional concealment fails this standard. To survive the 2026 cycle, applicants must abandon prophetic generalizations and disclose the architecture, datasets, and hyperparameters that allow POSITA reproduction, ensuring the specification supports the full scope rather than hiding behind the myth that Amgen applies only to antibodies or that AI patents are governed solely by eligibility frameworks.

The Evidence

Amgen Inc. v. Sanofi, 600 U.S. 436 (2023) dismantled the prophetic drafting strategy that AI practitioners have relied upon for a decade. The Supreme Court held claims to a genus of PCSK9-binding antibodies unenabled because the record demonstrated that isolating and structurally characterizing new antibodies outside the disclosed three required substantial time and resources. This is not a biological exception; it is a mechanical constraint on claim scope relative to disclosure depth. When you claim a functionally defined genus in machine learning—such as "a neural architecture optimized for low-latency inference"—without disclosing the training data provenance, hyperparameters, and validation results that define the boundaries of that optimization, you are asserting a scope that a person of ordinary skill cannot reproduce without undue experimentation. The Court's reasoning establishes that enablement requires the specification to teach the POSITA how to make and use the full scope of the claimed invention, not just a single working example or a vague aspiration.

The Federal Circuit has explicitly extended this constraint beyond biotechnology. In Becton, Dickinson & Co. v. B. Braun Medical (Fed. Cir. 2024), the court reaffirmed that Amgen's full-scope enablement requirement applies outside the antibody context, extending the risk to any genus-style claiming. This decision signals that the USPTO and courts will apply the same rigorous scrutiny to AI patents claiming broad model families or general-purpose architectures. If your application relies on 'the model will generalize' language without disclosing the specific dataset distributions, loss functions, and regularization techniques that constrain that generalization, you face the same fate as the unenabled antibody claims in Amgen. The 2026 examination cycle represents the practical cutoff where pending applications must be re-drafted around these disclosed architectures, datasets, and hyperparameters, or be abandoned.

The USPTO's own guidance confirms this trajectory. The February 13, 2024 Federal Register notice on AI-assisted inventions explicitly confirms that AI inventions remain subject to the same § 112 disclosure requirements as other technologies. Furthermore, the July 2024 USPTO request for public feedback on AI patentability signals that examination standards are tightening through 2026, with examiners increasingly demanding reproducible training disclosures. This is not a suggestion; it is a directive that aligns with the canonical rule: draft and prosecute AI patent applications so that a person of ordinary skill can reproduce the claimed model's full scope from the specification alone before the 2026 examination cycle.

Empirical evidence from litigation outcomes reinforces this urgency. Post-Amgen district court and PTAB decisions through 2025 show that § 112(a) challenges are replacing § 101 challenges as the lead invalidity theory in biotech and software-adjacent patents. Enablement invalidations are rising sharply where claim scope exceeds disclosed working examples. This trend indicates that the era of relying on Alice/§ 101 eligibility frameworks to save AI patents is over; the real threat now comes from insufficient enablement disclosures. According to IP Boutique Law (March 2026), 35 U.S.C. § 112(a) imposes three legally distinct requirements: written description, enablement, and best mode. Examiners issue written description and enablement rejections separately because response strategies differ fundamentally. Defining models primarily by what they do rather than what they are increases the risk of rejection. Post-Amgen decisions, including Seagen v. Daiichi Sankyo and OssiFi-Mab v. Amgen, invalidate claims directed to broad functionally defined genera lacking sufficient working embodiments. In Teva v. Lilly, the Federal Circuit held that isolating specific working embodiments for a targeted treatment using a well-known genus is a routine exercise akin to 'extra credit,' not a 'research assignment' requiring undue experimentation. When a patent claims a novel method of using a well-known genus, the required scope of enablement is narrower than Amgen suggests. However, when the genus itself is novel and unpredictable, the scope must be narrowed to match the disclosure. The data is unambiguous: AI patents drafted with model outputs instead of reproducible training disclosures fail enablement across their full claim scope. The 2026 USPTO examination cycle is the deadline to re-draft or abandon.

Disclosure Depth vs. Claim Breadth

Amgen v. Sanofi did not merely tighten written-description doctrine; it recalibrated the enablement calculus for functional genus claims across all technology sectors. The Supreme Court’s reasoning applies with equal force to machine-learning architectures, and the Federal Circuit’s 2024–2025 AI docket has already treated cross-domain generalization as a textbook example of undue experimentation. When examiners apply the Wands factors to AI specifications, they no longer accept “the model will generalize” as sufficient guidance. Instead, they demand reproducible training disclosures that map directly to the full breadth of the claimed scope. This structural shift forces practitioners to choose between claim breadth and disclosure depth—a trade-off that now dictates prosecution strategy through the 2026 examination cycle.

The three viable drafting postures diverge sharply on how they satisfy the Wands framework. Posture A relies on broad functional claims anchored by a single working example and roadmap language. Under Amgen, this posture collapses because it fails both the “amount of guidance” and “working examples” factors: a POSITA cannot extrapolate from one architecture to an unbounded genus without resorting to trial-and-error optimization. Posture B introduces an exemplar ladder—three to five concrete implementations spanning the claimed range, paired with a reproducible training recipe. Where the underlying art is predictable (e.g., fine-tuning standard transformer-class models on domain-specific corpora), posture B satisfies Wands because routine hyperparameter sweeps and established validation pipelines bridge the gap between examples and the full scope. Posture C demands complete transparency: every claimed domain receives its own hyperparameter set, dataset provenance chain, and validation metrics. It passes all Wands factors but forces immediate public disclosure of proprietary training data, effectively surrendering the dataset advantage that often constitutes the invention’s commercial moat.

The mechanism is straightforward: enablement measures whether a person of ordinary skill can make and use the invention without undue experimentation, and Amgen confirmed that claiming a functionally defined genus while disclosing only a fraction of what is claimed risks invalidation under both written description and enablement. Proactive portfolio audits against the Wands framework are now standard practice for life sciences IP counsel managing broad composition-of-matter claims, and the same audit logic applies to algorithmic genera. If your pending application still leans on prophetic language or single-example scaffolding, re-draft around disclosed architectures, datasets, and hyperparameters before the 2026 examination cycle closes. Otherwise, abandon the overbroad genus and prosecute narrower, fully enabled claims.

What the Data Doesn't Tell You

The 'AI enablement crisis' remains a projection from Becton Dickinson remands rather than settled Federal Circuit precedent, creating a structural risk for practitioners who treat the analogy as binding. No court has yet squarely applied Amgen v. Sanofi to a pure machine-learning patent, leaving the threshold for reproducible training disclosures untested in the algorithmic context. The analogy between antibody isolation and model training could fracture if the court determines that trained-model weights are inherently more reproducible than biological isolates; unlike the unpredictable nature of epitope mapping where minor mutations alter binding affinity, neural network inference is deterministic given fixed parameters. If the Federal Circuit views weight matrices as sufficient functional disclosure, the genus claims currently under siege might survive where the biotech roadmap failed. However, relying on this divergence is speculative. The canonical rule demands disclosure of training data provenance, hyperparameters, and validation results regardless of this theoretical distinction, because the burden of proof shifts to the applicant to demonstrate reproducibility across the full claim scope.

Predictability acts as a double-edged sword in post-Amgen prosecution, particularly for fine-tuned architectures. While Amgen penalized unpredictability in biotech, post-decision jurisprudence suggests that arts with high predictability may retain broader claim survival. Fine-tuning a disclosed base architecture on a new dataset is arguably far more predictable than isolating a novel antibody, as the underlying loss landscapes and gradient descent mechanisms follow known mathematical constraints. AI genus claims built on standard pipelines with disclosed hyperparameters may withstand enablement challenges where antibody claims collapsed, provided the specification proves that the claimed variations fall within the expected performance envelope. This creates a bifurcation: claims relying on 'black box' outputs without architectural transparency will fail, while those demonstrating predictable generalization through rigorous validation metrics may endure. Prosecutors must distinguish their applications by emphasizing the deterministic nature of the pipeline rather than treating all AI inventions as equally unpredictable.

Claims that enablement has supplanted § 101 as the primary rejection theory lack empirical verification due to a critical measurement gap. There is no public dataset quantifying how many AI patents were invalidated on § 112(a) grounds versus § 101 eligibility rejections during the 2024-2025 cycle. Assertions that enablement is now the 'lead rejection theory' rest entirely on litigation-pattern inference rather than audited counts. Without granular data on examiner rationales, practitioners cannot accurately calibrate prosecution strategies based on assumed rejection frequencies. According to LegalClarity (July 2026), Title 35 U.S.C. § 101 still governs patentability for new and useful processes requiring real-world practical application, meaning eligibility remains a foundational hurdle. Conflating written description requirements with enablement leads to failed prosecution responses that miss the examiner's actual ground of rejection, as noted by IP Boutique Law (March 2026). Until comprehensive tracking emerges, the shift toward enablement must be treated as a strategic assumption rather than a statistical certainty.

Examination outcomes exhibit significant variance across USPTO Technology Centers, undermining the notion of a uniform 2026 cutoff. TC 2100 examiners apply the Wands factors with a baseline expectation of predictability inherent to computer architecture, whereas TC 1600 examiners operate under the low-predictability standards established in biotech precedents. A specification disclosing only model outputs may pass examination in TC 2100 if the underlying architecture is standard, yet trigger an enablement rejection in TC 1600 where the same disclosure would be deemed insufficient for a biological genus. This fragmentation means the effective deadline for re-drafting pending applications depends on the assigned art unit. Applicants cannot assume a single regulatory horizon; instead, they must assess the specific predictability baselines of their assigned technology center and adjust disclosure depth accordingly.

Worked Case

The enablement math reveals a four-domain gap structurally identical to the problem in Amgen. The claims cover anomaly detection across five named domains—fraud, network security, manufacturing telemetry, medical monitoring, and energy load—but the specification enables only one. Under the Wands framework, the burden of experimentation required to bridge this gap is undue because the specification offers no roadmap for adapting the model to new domains. The examiner determined that a person of ordinary skill in the art would need to engage in extensive trial-and-error to determine appropriate preprocessing, feature extraction, and hyperparameter tuning for each unexemplified domain. This outcome confirms that Amgen's technology-neutral enablement test applies with equal force to machine-learning patents; the Federal Circuit's 2024-2025 docket has consistently rejected arguments that AI models are inherently unpredictable and thus exempt from reproducible disclosure requirements. The myth that Amgen governs only antibody cases is debunked by the court's explicit reliance on the principle that a specification must enable a POSITA to make and use the full scope of the invention without excessive experimentation.

To cure the rejection, the applicant employed an exemplar-ladder fix, adding two additional working examples and a disclosed fine-tuning recipe. The first added example demonstrated network intrusion detection on a public benchmark (CIC-IDS2017), disclosing input dimensions, normalization techniques, and convergence metrics. The second added example covered manufacturing telemetry, specifying sliding window sizes, missing-data imputation strategies, and a retraining cadence tied to concept drift thresholds. Crucially, the specification included a fine-tuning recipe: frozen encoder layers, a ten-epoch schedule with a decaying learning rate, and validation AUC thresholds above 0.90 on each exemplar. This recipe reduced the undue-experimentation burden for the remaining two domains—medical monitoring and energy load—to routine adjustment. According to the evolving body of case law following Amgen, as analyzed by AZBio, IP counsel can leverage such predictable adaptation methods to satisfy enablement even when not every embodiment is fully exemplified. The fine-tuning protocol provided a clear pathway for a POSITA to apply the core architecture to new domains using standard techniques, thereby bridging the enablement gap without requiring exhaustive disclosure of every possible variation.

The re-drafting process involved a strategic claim-scope trade. The independent claim was narrowed from "any time-series domain" to "time-series domains of the type exemplified, wherein the model is fine-tuned per the disclosed recipe." This amendment cut literal scope by roughly the two unexemplified domains but converted a near-certain § 112(a) rejection into an allowable claim set. By anchoring the claim to the exemplified types and incorporating the fine-tuning method, the applicant satisfied the enablement requirement while preserving meaningful protection. The amended application survived the 2026 examination posture with three exemplars enabling a five-domain genus because the fine-tuning method was disclosed and predictable. In contrast, the original prophetic drafting would have faced abandonment or a significantly narrowed, litigation-vulnerable claim. The lesson is clear: AI patent applications must be drafted and prosecuted so that a person of ordinary skill can reproduce the claimed model's full scope from the specification alone, disclosing training data provenance, hyperparameters, and validation results before the 2026 examination cycle cutoff. Relying on model outputs without reproducible training disclosures will continue to fail enablement across the full claim scope.

How to Choose Well

Amgen v. Sanofi does not merely tighten written-description doctrine; it imposes a mechanical constraint on genus claims that collapses when the specification fails to anchor the full scope in reproducible training data. For AI practitioners, the 2026 USPTO examination cycle marks the practical cutoff where prophetic drafting strategies—claiming broad model capabilities without disclosing the underlying

Frequently Asked Questions

How many disclosed antibody structures were deemed insufficient to enable the twenty-six claimed PCSK9 antibodies in Amgen v. Sanofi?

The Supreme Court rejected the roadmap argument when only three disclosed structures were provided for the twenty-six claimed PCSK9 antibodies.

Does commercial success or FDA approval protect an AI patent specification that forces skilled artisans into trial-and-error discovery across the full claimed scope?

Commercial success or FDA approval does not insulate a patent if the specification forces skilled artisans into trial-and-error discovery across the full claimed scope.

What specific ratio of claims to working examples triggers the exact scope-to-disclosure gap penalized by the Amgen holding?

A drafter can produce a forty-claim family from one working example, a ratio precisely penalized by Amgen.

Which two distinct statutory requirements must AI applicants address separately rather than relying solely on § 101 eligibility arguments?

Patent rules flow from Title 35 USC and the Manual of Patent Examining Procedure (MPEP), administered by the USPTO, where the Alice/Mayo two-step test (§ 101 eligibility) is distinct from the enablement requirement (§ 112).

What date marks the practical cutoff for pending AI applications to be re-drafted around disclosed architectures, datasets, and hyperparameters before facing tightened examination standards?

The 2026 examination cycle serves as the practical cutoff for these disclosures.

Under what condition do purely theoretical AI concepts fail the utility requirement under § 101 even if they lack reproducible disclosure?

Purely theoretical concepts with no identified real-world practical application fail the utility requirement under § 101.

Quick answers

Also worth reading: Sanofi's $14B Trademark Acquisition Examining the Quten Research Institute Buyout Strategy: Sanofi's $14B Trademark Acquisition Examining · USPTO Director Squires signals future updates to AI patent eligibility guidance for 2025: USPTO Director Squires signals future · Micron's $445M patent loss locked in by Biden's comments: Micron's $445M patent loss locked