Audit Cuts Green Patent Count by a Quarter
An error-as-signal audit combines PatentSBERTa disagreements with two-model climate-function judgments, reducing the Y02 estimate from 592,387 to 441,468 patents.
Underlying Paper
Systematic Bias in Green Patent Classification: Silent Green and False Green
Green-patent indicators based on Cooperative Patent Classification Y02 tags increasingly inform research, industrial policy, and climate-oriented investment, yet their construct validity has not been evaluated at corpus scale. We ask whether Y02 classification errors are random measurement noise or systematic, direction-specific bias. We introduce an Error-as-Signal framework in which disagreement between an administrative label and an independent model is treated as evidence of potential measurement error. Screening 9,075,421 USPTO granted patents from 1962-2024 with a fine-tuned domain model identifies 517,772 disagreements. Two independent open-weight large language models then assess whether each flagged invention has a direct climate-mitigation or adaptation function. Cross-model consensus identifies 180,384 administrative Type I errors (False Green) and 29,465 Type II errors (Silent Green). Correcting these errors reduces the measured green-patent population by 25.5%, from 592,387 to 441,468 patents. Misclassification is systematic rather than random. Atypicality predicts Silent Green in an inverted-U pattern, while reflection complexity independently increases under-recognition: controlling for atypicality and filing year, a one-standard-deviation increase is associated with 1.61 times the odds of Silent Green. Structural complexity has the opposite association. Among consensus-attributed errors, the same increase in reflection complexity is associated with 2.45 times the odds that an error is Silent Green rather than False Green. Event tests show no discrete rise in misclassification when green classification became more salient and only limited evidence of increased explicit green framing after the 2013 CPC launch. The evidence is more consistent with bounded classification capacity than with applicant gaming.
Y02 tags in the Cooperative Patent Classification system have become a convenient proxy for green inventive activity, supporting work in innovation economics, industrial policy, and climate-oriented investment. Their usefulness depends on a demanding assumption: that the administrative label tracks inventions with a direct mitigation or adaptation function rather than merely related technologies, applications, or language. This paper tests that assumption at unusual scale, treating disagreement between Y02 and an independent classifier as evidence to investigate rather than noise to discard.
Across 9,075,421 USPTO granted patents from 1962–2024, the authors compare administrative Y02 tagging with PatentSBERTa, a fine-tuned domain model. They then submit the 517,772 disagreements to two independent open-weight large language models, which assess whether the invention has a direct climate mechanism. The central result is directional: the administrative system appears to over-include far more patents than it omits. Correcting consensus-attributed errors lowers the measured green-patent population by 25.5%.
Core Contribution
The paper's contribution is its Error-as-Signal framework. Instead of accepting an administrative taxonomy as ground truth or treating a model as its automatic replacement, it isolates cases in which the two systems conflict and audits the substantive source of the conflict. That framing matters because classification error changes the quantity being measured: a green-patent count can rise through broader tagging rather than more climate-relevant invention.
The disagreement diagram makes the scale clear. Y02 and PatentSBERTa jointly classify 341,357 patents as green, while 251,030 are Y02-only and 266,742 are model-only; 8,216,292 are non-green under both systems. The study does not substantively audit the agreement cases, so its correction is deliberately limited to the disagreement pool rather than a claim that every remaining label is correct.
Technical Approach
The audit distinguishes four possibilities within a disagreement: administrative Type I error, or False Green; administrative Type II error, or Silent Green; and the corresponding two model errors. For an administrative False Green finding, an administratively Y02-green patent is removed when the audit attributes the disagreement to Y02 and finds no direct climate mechanism. For an administrative Silent Green finding, a patent missing a Y02 label is added when the audit attributes the disagreement to Y02 and finds a direct climate mechanism. Unresolved cases retain their original Y02 status. This is a conservative recoding rule: it avoids forcing a decision when the audit systems do not resolve a case, but it also means the corrected count remains conditional on the audit design.
Figure 4 shows how that attribution changes the aggregate measure: 180,384 administratively green patents are recoded as False Green, whereas 29,465 model-identified green patents are treated as Silent Green. Starting from 592,387 Y02 patents, that asymmetric correction yields 441,468 consensus-adjusted green patents. The figure also reports an extreme-allocation sensitivity interval of 390,540–508,126, indicating that unresolved cases still affect the total materially.
The authors also test whether the error pattern is consistent with strategic green framing or constrained classification capacity. Their models use lexical indicators from original titles and abstracts, filing period, and pooled primary non-Y CPC subclass. Atypicality has an inverted-U association with Silent Green, while reflection complexity is associated with under-recognition after controlling for atypicality and filing year. Structural complexity points in the opposite direction.
Results and Analysis
The clearest empirical finding is the imbalance between error directions: consensus identifies 180,384 False Green cases against 29,465 Silent Green cases. The correction therefore does not merely reshuffle patent families; it reduces the estimated population by 150,919 patents. PatentSBERTa alone identifies 608,099 green patents, higher than the Y02 count, but the consensus measure is lower than both. That gap illustrates why replacing one classifier with another would not answer the validity question: the paper's direct-mechanism criterion is doing substantive work beyond label agreement.
The field pattern is equally consequential. Figure 5 places administrative over-inclusion in digital data processing, wireless networks, digital communications, semiconductor devices, and selected medical and chemical technologies. Administrative omission clusters instead in separation processes, catalysis, engine-exhaust control, heat pumps and refrigeration, power systems, batteries and fuel cells, and water treatment. The error is thus not plausibly treated as uniform measurement noise across technological domains.
The regression evidence supports the capacity interpretation more than an applicant-gaming account, but does not settle causality. A one-standard-deviation increase in reflection complexity is associated with 1.61 times the odds of Silent Green; among consensus-attributed errors, the same increase corresponds to 2.45 times the odds that the error is Silent Green rather than False Green. Event tests show no discrete error increase around Y02's 2010 introduction or the 2013 CPC launch, alongside only limited evidence of more explicit green framing after 2013. For users of Y02 counts, the practical implication is narrow but important: comparisons across fields or periods should not assume that the label has stable, direction-neutral measurement error.
Limits of the Audit
The evidence is extensive but not a direct human adjudication of the full corpus. Two model assessments determine the mechanism-based audit, and agreements between Y02 and PatentSBERTa are not substantively checked. The paper therefore supports a strong warning about systematic error in disagreement cases and a carefully defined corrected count; it does not establish an error-free census of all green patents.
Evidence Box
moderateKey Claims
- •Administrative Y02 errors are systematic and direction-specific
- •Direct-mechanism consensus can distinguish False Green from Silent Green patents
- •Bounded classification capacity better explains errors than applicant gaming
Key Results
- •517,772 Y02–PatentSBERTa disagreements among 9,075,421 USPTO granted patents
- •180,384 False Green and 29,465 Silent Green cases under cross-model consensus
- •441,468 corrected green patents versus 592,387 Y02 patents, a 25.5% reduction
- •One SD of reflection complexity corresponds to 1.61× odds of Silent Green and 2.45× odds versus False Green
Limitations & Caveats
- •Two LLM assessments, rather than human adjudication, attribute disagreement errors
- •The 341,357 Y02–model green agreement cases are not substantively audited
- •Unresolved disagreements retain their original Y02 classification
- •Corpus covers USPTO granted patents from 1962–2024 rather than global patent systems