The exact contents of citations.db main_text.text for this paper — one flattened LaTeX string, title through conclusion, appendix excluded, unmodified except for removing email addresses. This is what our citation measures are computed over.
120,204 characters
Systematic Bias in Green Patent Classification: Silent Green and False Green
\maketitle
\begin{abstract}
Green-patent indicators built on the Cooperative Patent Classification Y02 tags now steer research assessment, industrial strategy, and large flows of climate-directed capital, yet their construct validity has never been audited at corpus scale. We assess whether Y02 is systematically biased and whether that bias may reinforce the documented ESG innovation disconnect, in which high-emitting firms with weak environmental ratings are prolific green patenters. We introduce an Error-as-Signal framework that treats disagreement between an administrative label and an independent model as diagnostic evidence of measurement error rather than model failure. Screening all \num{9075421} USPTO granted patents (1962--2024) with a fine-tuned domain model against the Y02 label yields \num{517772} dual-measurement disagreements; an ensemble of two independent open-weight large language models then judges, by cross-model consensus, whether each flagged invention has a direct climate-mitigation or adaptation function. We identify \num{180384} administrative Type I errors (False Green, concentrated in digital data processing, wireless networks, digital communications, and semiconductors) and \num{29465} Type II errors (Silent Green, concentrated in separation processes, catalysis, exhaust control, heat pumps, power systems, and batteries). Correcting consensus-attributed administrative errors reduces the measured green-patent population by 25.5\% (\num{592387} to \num{441468} patents; sensitivity bounds \num{390540}--\num{508126}), with ICT energy efficiency (Y02D) falling 67.6\%. Omission rises with technological atypicality in an inverted-U pattern, so complex, unconventional inventions, whose climate function lies in engineering rather than green vocabulary, are the most likely to go unlabelled. Event tests reveal no discrete jump in misclassification when green classification became salient, only a small relative rise in green framing after the 2013 CPC launch. The bias primarily reflects classification capacity rather than applicant strategy, and structurally under-recognizes heavy industry.
\end{abstract}
\textbf{Keywords:} {\bfseries\boldmath Measurement; classification; construct validity; climate innovation; ESG}
\textbf{JEL codes:} {\bfseries\boldmath O31; O34; Q55}
\section{Introduction}\label{sec:introduction}
The global effort to achieve a net-zero economy represents one of the most significant attempts at directed technical change in modern times. To guide this transition, policymakers, international organizations like the OECD and WIPO, and investors rely heavily on quantitative indicators to identify, track, and support climate-related technologies. Patent statistics have long been central to measuring inventions \parencite{Griliches1990}. Within this tradition, the Cooperative Patent Classification (CPC) Y02 scheme has become the dominant administrative indicator for identifying climate-related inventions. The Y02 classification was created as a tagging scheme for climate-change mitigation technologies and has become widely used in research, policy monitoring, and ESG-related analysis \parencite{Veefkind2012}. Because these indicators shape how scholars, policymakers, and investors evaluate the direction of green technological change, their accuracy and validity are matters of first-order importance.
However, a growing body of evidence reveals a troubling disconnect between these widely used green patent indicators and other measures of environmental performance. Recent research documents a significant ESG innovation disconnect, where firms with high environmental ratings do not necessarily produce more green patents, and conversely, firms in traditionally capital-intensive heavy industries are often the most prolific producers of green technologies despite weaker ESG scores \parencite{Cohen2026,Lan2025}. This paradox raises a fundamental question for innovation policy. Does the disconnect reflect a failure of corporate environmental strategy, or does it partly stem from the tools used to measure green inventions being themselves imperfect?
The answer matters because, to our knowledge, the construct validity of Y02 has never been audited across the full patent corpus. The scheme's own designers cautioned that inclusion in the tagging scheme does not by itself guarantee that a technology is environmentally sound \parencite{Veefkind2012}, and a growing set of sample-limited studies has documented both over-inclusion, notably in information and communication technologies, and omission across green-patent classification systems \parencite{Santarlasci2023,Rainville2025}. What has been missing is a corpus-scale audit that can separate ordinary, idiosyncratic classification mistakes from structural, direction-specific measurement bias; attribute the direction of each error to an invention's primary technical function; and adjudicate whether any bias originates in the classification system itself or in the strategic behaviour of applicants. This study proposes a complementary patent-side account of the ESG innovation disconnect. Building on construct-validity research \parencite{CronbachMeehl1955,AdcockCollier2001} and recent work on measurement in computational systems \parencite{JacobsWallach2021}, we distinguish the underlying construct, a patent's substantive contribution to climate-change mitigation or adaptation, from the administrative indicators used to track it. The administrative Y02 label is one such indicator, while a model prediction is another. Neither should be treated as unquestioned ground truth. The Y02 system can generate two opposing administrative errors. An administrative Type I error, which we call False Green, occurs when a non-green patent is included in Y02. An administrative Type II error, which we call Silent Green, occurs when a genuinely climate-relevant patent is omitted. Yet the predictive model can also make Type I and Type II errors. A Y02-labelled patent rejected by the model may therefore represent either administrative over-inclusion or a model false negative. Likewise, an unlabelled patent predicted as green may represent either administrative under-inclusion or a model false positive. Dual-measurement disagreement (model--administrative) identifies where a measurement problem may exist, but it does not, by itself, determine which system is wrong.
We therefore develop an Error-as-Signal Framework for administrative measurement. The framework treats disagreement not simply as model failure, but as diagnostic evidence of possible measurement error. Although developed and applied here to green patents, the framework
is indicator-agnostic: it applies wherever an authoritative administrative label can
be confronted with an independent measurement of the same underlying construct. PatentSBERTa first screens the full patent corpus for disagreement with Y02. To adjudicate these signals, we implement an LLM-bagging approach: two independent LLMs (Qwen\footnote{\texttt{Qwen/Qwen3-32B}: \url{https://huggingface.co/Qwen/Qwen3-32B}.} and Gemma~4\footnote{\texttt{google/gemma-4-31B-it}: \url{https://huggingface.co/google/gemma-4-31B-it}.}) evaluate the primary technical function of each flagged invention. Their direct-mechanism consensus attributes resolved disagreements to an administrative or model error, while cross-model disagreements remain explicitly unresolved. By isolating these \num{117586} ambiguous cases, the framework distills the measurement problem from 9~million totalgranted patents down to a targeted escalation set for future research. The objective is not to claim that all classification error can be eliminated. Rather, it is to distinguish ordinary classification noise from systematic errors that bias comparisons across sectors, technologies, and firms.
Applying this framework to \num{9075421} U.S. granted patents issued between 1962 and 2024, we establish a baseline of \num{592387} administratively Y02-labelled green inventions. While initial screening with PatentSBERTa yields over 8.5 million model--administrative agreements, including \num{341357} confirmed Y02 labels, it isolates \num{517772} critical disagreements: \num{251030} downward (administratively green but predicted non-green) and \num{266742} upward (administratively non-green but predicted green). Qwen and Gemma then independently assess whether the invention's primary technical function has a direct climate-mitigation or adaptation mechanism. Direct-mechanism consensus attributes \num{180384} cases to administrative Type I error (False Green), \num{29465} to administrative Type II error (Silent Green), \num{170619} to model Type I error, and \num{19718} to model Type II error. A further \num{117586} cases remain unresolved. Correcting only the consensus-attributed administrative errors yields \num{441468} substantively green patents, or 4.86\% of the corpus.
The findings reveal a pronounced sectoral bias. False Green is disproportionately concentrated in asset-light and ICT-related fields, where environmental terminology can be visible despite a weak climate mechanism. Silent Green is concentrated in capital-intensive and technically complex fields, including energy, transportation, chemistry, and industrial production, where the climate contribution is often embedded in engineering function rather than expressed through overtly green vocabulary. These opposing errors may partly offset one another in aggregate while still transferring measured green inventive activity from sectors central to deep decarbonization toward sectors in which green language is more visible. Two mechanism analyses sharpen this interpretation: within the corrected green-patent portfolio, the probability of administrative omission rises with technological atypicality and peaks at moderate-to-high levels, whereas the prevalence of misclassification shows no discrete break when green classification became salient, and explicit green framing among administratively labelled patents rises at most modestly, and only after the 2013 CPC launch. In short, the indicator structurally perpetuates the gap it is intended to diagnose. Sectors that contribute significantly to decarbonization are precisely those that the metric fails to acknowledge.
This work makes three core contributions. First, it delivers the first corpus-scale construct-validity assessment of a green-patent indicator, and it shows how to distinguish individual classification error from structural measurement bias, a distinction the green-innovation literature has asserted but rarely operationalized at this scale. Where prior work compared classification schemes or sampled sub-domains, we audit the entire USPTO granted corpus and quantify a net correction of roughly one quarter of the measured green-patent population, with explicit sensitivity bounds, providing a benchmark against which future indicators can be tested. Second, it introduces the Error-as-Signal framework for administrative measurement: a general, reusable method that converts disagreement between administrative and model-based measurements into attributable evidence of measurement error, using an ensemble of independent judges to separate attributable error from genuine ambiguity. The framework is indicator-agnostic and applies wherever an authoritative label can be checked against an independent functional assessment, from ESG ratings and sustainable-finance taxonomies to R\&D statistics. Third, it offers a complementary patent-side account of the ESG innovation disconnect: the finding that low-ESG heavy-industry firms out-patent high-ESG firms in green technology may be reinforced by an indicator that under-recognizes mechanical, chemical, and industrial decarbonization while over-crediting digitally visible green language. Alongside these contributions, we provide an open patent-level dataset\footnote{\url{AAUBS/Silent-Green}} detailing consensus-attributed error classes, model confidence scores, functional mechanisms, and unresolved cases, enabling replication and extension.
The remainder of the paper is organized as follows. Section~\ref{sec:theory} develops the theoretical background, connecting the ESG innovation disconnect to construct validity and formalizing the Error-as-Signal logic of dual-measurement disagreement. Section~\ref{sec:methodology} describes the data, the PatentSBERTa screening stage, and the bagging-inspired LLM audit. Section~\ref{sec:results} reports the error-attribution results, the technological structure of the attributed errors, mechanism tests linking administrative omission to technological atypicality and tracing green framing around the introduction of Y02, the CPC, and generative AI, and the consensus-corrected landscape of green invention. Section~\ref{sec:discussion} draws out the implications for innovation measurement, algorithmic auditing, and policy, and Section~\ref{sec:conclusion} concludes with limitations and directions for future research.
\section{Theoretical Background}\label{sec:theory}
The global shift toward a net-zero economy is a paradigmatic example of directed technical change, in which policy and market signals are intended to steer innovation toward societal goals \parencite{acemoglu2012environment}. To assess the effectiveness of these signals, policymakers and scholars require robust, valid, and reliable indicators of inventive activity. For decades, patent statistics have served this role, providing granular, long-term data on the rate and direction of technological change \parencite{Griliches1990,jaffe2002patents}. The introduction of the CPC's ``Y'' section, and specifically the Y02 scheme for climate-change mitigation technologies, was a landmark development in this tradition, creating a dedicated policy instrument to track the outputs of this directed innovation effort \parencite{Veefkind2012}.
\subsection{The ESG Innovation Disconnect: A Crisis of Measurement?}
Despite the widespread adoption of Y02-based indicators, a significant empirical gap has emerged in the literature: the ESG innovation disconnect. A recent study by \textcite{Cohen2026} documents that firms with high environmental ratings do not consistently produce more green inventions than their lower-rated counterparts. Incumbent firms in carbon-intensive, heavy industries are often among the most prolific and influential producers of green patents despite weaker ESG profiles \parencite{Lan2025}.
This paradox presents a fundamental challenge to innovation theory and policy. It shapes where green inventions are believed to occur, which firms and sectors are seen as leading the transition, and how public and private resources are directed towards climate-relevant technologies. If patent-based indicators systematically overcount some fields and undercount others, observed patterns may reflect measurement bias rather than the true distribution of green inventive activity. This paradox may arise from two broad, non-exclusive sources. First, firm-side mechanisms may create a substantive disconnect between environmental strategy, communication, and green innovation. Firms may fail to translate environmental commitments into corresponding R\&D activities, consistent with policy--practice decoupling \parencite{BromleyPowell2012,delmas2011drivers}. Even when environmental innovation activities are undertaken, sector-specific capabilities, organizational routines, and managerial cognition may weaken the link between those activities and their intended technological outcomes. This pattern can be interpreted as firm-level means--ends decoupling, although the innovation literature primarily identifies the underlying capability and inertia mechanisms rather than using that term directly \parencite{BromleyPowell2012,HendersonClark1990,TripsasGavetti2000}.
Genuine green inventions may remain under-recognized for two distinct reasons. The first is strategic: firms may deliberately downplay or withhold communication about their environmental achievements, consistent with recent work on greenhushing and brownwashing \parencite{Hilton2025,MontgomeryRobertsonSummers2026} and foundational research on greenhushing and strategic silence \parencite{font2017greenhushing,carlos2018strategic}. The second is non-strategic: an invention's climate contribution may be accurately described but embedded in specialized engineering, scientific, or legal language rather than explicitly framed in environmental terms. We refer to this mechanism as technical under-framing. Recent research documents both the distinctive linguistic complexity of patent texts and important gaps across green-patent classification systems \parencite{JiangGoetz2025,Santarlasci2023,Rainville2025}, while the original development of the Y02 taxonomy provides the foundational classification context \parencite{Veefkind2012}.
Second, the apparent disconnect may arise from measurement limitations on either side of the relationship. ESG ratings may imperfectly represent firms' environmental performance because providers differ in their scope, measurement methods, and weighting schemes \parencite{Berg2022,Chatterji2016}. Patent-based indicators may likewise exhibit a form of means--ends decoupling in administrative measurement: classification procedures are implemented, yet the resulting labels do not consistently achieve their intended objective of identifying substantively climate-relevant inventions \parencite{BromleyPowell2012}. Because substantive climate relevance is often difficult to infer from specialized patent language, classification systems operating under limited attention and codified taxonomies may rely on simplifying heuristics, privileging salient lexical cues and familiar technological categories over less visible functional mechanisms \parencite{JiangGoetz2025,Simon1955,TverskyKahneman1974}. The resulting gap between the substantive construct and its administrative representation can weaken measurement validity and generate systematic differences across green-patent classifications \parencite{JacobsWallach2021,Rainville2025}. More generally, patents and administrative taxonomies are imperfect indicators of innovation and technological constructs \parencite{Griliches1990,Nagaoka2010,Santarlasci2023,Veefkind2012}. Although measurement divergence among ESG ratings has been extensively documented, considerably less attention has been paid to whether the patent-based indicators on the innovation side are themselves systematically biased. Our study addresses this second measurement problem by examining whether Y02 over-recognizes some technological fields while under-recognizing others.
\subsection{How Green Patent Labels Are Assigned}
Green patent indicators do not directly observe whether an invention is environmentally beneficial. Instead, they rely on administrative classification systems that translate technical patent documents into standardized technology codes. The CPC Y02 scheme was developed to identify climate-change mitigation and adaptation technologies that were previously dispersed across conventional patent classes \parencite{Veefkind2012}. Y02 has therefore become a widely used proxy for green invention in research, policy monitoring, and ESG-related analysis.
However, assigning a Y02 label is not the same as measuring substantive environmental impact. It is an administrative classification decision in which patent text is mapped onto a predefined taxonomy. This process may involve examiner judgment, classification expertise, automated search procedures, and patent-office updating systems \parencite{Santarlasci2023}. As a result, Y02 should be understood as an administrative proxy for climate relevance rather than as ground truth.
This distinction matters because the environmental function of a patent is not always explicit. Some inventions use visible green language, such as emissions reduction, renewable energy, carbon capture, or recycling. Others contribute to decarbonization through technical mechanisms such as catalyst design, thermal efficiency, fuel injection, power electronics, materials substitution, or industrial-process optimization. These inventions may be substantively green even when they do not use overt environmental vocabulary.
This creates a bounded-rationality problem in green patent classification. Patent classifiers and administrative systems operate under limited time, attention, and information-processing capacity \parencite{Simon1955,Simon1957}. Under such constraints, classification may rely on simplifying heuristics, such as salient keywords, familiar technology templates, or established Y02 categories \parencite{TverskyKahneman1974}. This does not imply intentional misclassification, but it helps explain why classification errors may be systematic.
From this perspective, two forms of Y02 error can arise. Patents with visible green language may be over-included even when their substantive climate contribution is weak, producing Type I errors or False Green. Conversely, patents whose climate relevance is embedded in complex technical functions may be missed, producing Type II errors or Silent Green inventions. We therefore treat Y02 as an indicator to be validated, not as a final measure of green invention.
\subsection{Measurement Validity and the ESG Innovation Disconnect}
This study proposes a complementary patent-side account of the ESG innovation disconnect. Rather than treating the divergence between ESG scores and green patent counts as direct evidence of a failure in corporate environmental strategy, we argue that measurement bias in Y02 may reinforce this disconnect. In particular, the Y02 classification should not be treated as a neutral or self-validating measure of green invention. It is an administrative indicator whose validity must be assessed against the substantive technological function of the patents it classifies.
This argument builds on the literature on construct validity and social-science measurement. Construct validity refers to whether an empirical indicator actually measures the theoretical concept it is intended to capture \parencite{CronbachMeehl1955}. In social-science measurement, this requires a clear distinction between the underlying concept, the operational indicator, and the empirical observations generated by that indicator \parencite{AdcockCollier2001,JacobsWallach2021}. Applied to green invention, the underlying concept is substantive climate-related technological contribution, while the empirical indicator is the administrative Y02 label. The key validity question is therefore whether Y02 accurately captures patents that contribute directly and credibly to climate mitigation or adaptation.
From this perspective, the ESG innovation disconnect can be understood as a possible measurement artifact. If Y02 over-includes patents with weak or symbolic environmental relevance, raw green patent counts may make some firms and sectors appear more green than they are. Conversely, if Y02 under-includes technically complex decarbonization inventions, especially in industrial sectors, it may hide important green inventions from firms that are central to the net-zero transition. In both cases, the observed divergence between ESG ratings and green patent counts may partly reflect mismeasurement in the patent indicator rather than firm behaviour alone. \parencite{Cohen2026} document an ESG innovation disconnect in which traditional energy firms make substantial contributions to green innovation despite weak ESG recognition. Our findings suggest that measurement bias in Y02 may reinforce, and potentially widen, this disconnect by under-recognizing mechanical, chemical, and industrial decarbonization while over-crediting digitally visible technologies.
All large-scale classification systems contain some degree of error. The central validity concern is therefore not the mere existence of mistakes, but whether those mistakes are approximately non-differential or systematically related to the characteristics of the observations being measured. Random or non-differential error primarily reduces precision and weakens estimated relationships. Systematic error is more consequential because its probability or direction varies across sectors, technologies, firms, or linguistic forms. In that case, classification error becomes measurement bias. This distinction is especially important because an aggregate correction alone cannot reveal measurement bias: opposing errors may partly offset one another in overall counts while still redistributing measured green invention across the technological landscape. The relevant question is therefore not only how many patents are misclassified, but whether the probability and direction of error vary systematically across technological fields, a question we test directly in Section~\ref{sec:results}.
\subsection{Error as Signal: Dual-Measurement Disagreement}
We conceptualize green-patent identification as a dual-measurement problem. Let $G_i\in\{0,1\}$ denote the latent substantive climate relevance of patent $i$, $A_i\in\{0,1\}$ its administrative Y02 label, and $M_i\in\{0,1\}$ the model prediction. Because $G_i$ is not directly observed at scale, the empirical analysis uses the direct-mechanism consensus of the LLM-bagging audit as an audit-based reference judgment for resolved cases. Administrative and model errors are respectively defined as
\[
e_i^{A}=A_i-G_i,
\qquad
e_i^{M}=M_i-G_i.
\]
A value of zero indicates a correct classification, a positive value indicates the incorrect inclusion of a non-green patent, and a negative value indicates the omission of a genuinely green patent. The distinction between random error and systematic bias can be expressed conditionally. If
\[
E[e_i^{A}\mid X_i]\approx 0,
\]
where $X_i$ represents technological domain, sector, linguistic visibility, or complexity, administrative errors behave primarily as noise with respect to those characteristics. If instead
\[
E[e_i^{A}\mid X_i]\neq 0,
\]
the administrative indicator is conditionally biased. The same logic applies to model error. Bias therefore refers not simply to non-normality or to a high number of mistakes, but to a persistent direction or unequal probability of error across relevant groups. Model--administrative disagreement provides a scalable diagnostic signal. Define
\[
D_i=\mathbf{1}(A_i\neq M_i).
\]
When $D_i=1$, the two measurement systems cannot both be correct relative to a binary substantive truth. However, disagreement alone does not reveal which system is wrong. A downward disagreement ($A_i=1,M_i=0$) may be an administrative Type I error or a model Type II error. An upward disagreement ($A_i=0,M_i=1$) may be an administrative Type II error or a model Type I error. Substantive adjudication is therefore necessary to attribute the error source. This logic follows research on multiple imperfect observers and noisy labels, which cautions against treating any single observed label as an infallible reference standard \parencite{DawidSkene1979,FrenayVerleysen2014,Northcutt2021}. Table~\ref{tab:error_attribution_framework} summarizes the resulting attribution scheme; the empirical frequency of each outcome is reported in Section~\ref{sec:results}.
\begin{table}[H]
\centering
\caption{Attribution scheme for administrative and model classification errors under direct-mechanism LLM consensus}
\label{tab:error_attribution_framework}
\small
\setlength{\tabcolsep}{4pt}
\begin{tabularx}{\textwidth}{
>{\centering\arraybackslash}p{0.75cm}
>{\centering\arraybackslash}p{1.35cm}
>{\centering\arraybackslash}p{1.15cm}
>{\raggedright\arraybackslash}p{2.6cm}
>{\raggedright\arraybackslash}X
}
\toprule
\textbf{Y02}
& \textbf{TF}
& \textbf{LLMs}
& \textbf{Error source}
& \textbf{Interpretation} \\
\midrule
1 & 0 & 0
& Admin. Type I
& False Green: Y02 incorrectly includes a patent without a direct green mechanism. \\
0 & 1 & 1
& Admin. Type II
& Silent Green: Y02 omits a patent with a direct green mechanism. \\
0 & 1 & 0
& Model Type I
& The TF model incorrectly classifies a functionally non-green, non-Y02 patent as green. \\
1 & 0 & 1
& Model Type II
& The TF model incorrectly classifies a functionally green Y02 patent as non-green. \\
\midrule
\multicolumn{3}{c}{\textit{No cross-model LLM consensus}}
& Unresolved
& The functional benchmark is unresolved; therefore, the error source cannot be attributed reliably. \\
\bottomrule
\end{tabularx}
\vspace{0.4em}
\begin{minipage}{0.97\textwidth}
\footnotesize
\textit{Note:} Y02 denotes the administrative classification, and the TF model
refers to PatentSBERTa. For Y02 and the TF model, 1 denotes green and 0 denotes
non-green. For the LLM judgment, 1 indicates cross-model consensus that the
patent contains a direct climate-mitigation or adaptation mechanism, whereas
0 indicates cross-model consensus that such a mechanism is absent. Incidental
environmental benefits are coded as 0. Cases without cross-model LLM consensus
are classified as unresolved. Confidence is analysed separately and does not
alter the consensus label. The empirical count corresponding to each
outcome is reported in Table~\ref{tab:llm_validation_disagreement}.
\end{minipage}
\end{table}
The framework changes the role of prediction error. Under conventional evaluation, a mismatch with an observed label is counted as model failure. Under imperfect administrative measurement, the same mismatch becomes an error signal: it identifies a case in which the validity of either the administrative label or the model prediction should be examined. Once validated, the direction, confidence, sectoral concentration, and linguistic characteristics of the errors reveal whether the measurement system is affected by lexical over-inclusion, functional opacity, taxonomic under-coverage, model overgeneralization, or domain shift.
\subsection{Administrative Error Mechanisms: False Green and Silent Green}
False Green and Silent Green refer specifically to audit-attributed administrative errors. They should not be inferred directly from model--administrative disagreement. A downward disagreement becomes False Green only when the functional audit supports a non-green judgment; an upward disagreement becomes Silent Green only when the audit supports a direct climate mechanism. In the large-scale analysis, this attribution is made only under cross-model LLM consensus.
\subsubsection{Administrative Type I Error: False Green}
False Green refers to the misclassification of a patent as green when it offers no substantive climate-mitigation or adaptation benefit. These false positives often arise from lexical ambiguity, where broad terms such as environment, efficiency, hybrid, or power management are used in non-climate contexts, for example in expressions such as software environment, hybrid data architecture, or generic power management.
The literature suggests two broad explanations for this form of over-inclusion. The first is a firm-level symbolic explanation. From an institutional perspective, firms may respond to regulatory, investor, and market expectations by adopting visible green signals that enhance legitimacy, even when these signals are only weakly connected to substantive environmental change \parencite{meyer1977institutionalized,dimaggio1983iron}. This logic is consistent with the greenwashing literature, which emphasizes institutional, market, and disclosure-related drivers of symbolic environmental claims \parencite{delmas2011drivers,marquis2016scrutiny}. Recent patent-based studies provide related evidence that green finance, ESG pressure, and environmental regulation can increase visible green innovation outputs without necessarily improving the substantive quality of those inventions, suggesting a form of patent-based greenwashing or symbolic green innovation \parencite{shi2023green,zhang2022environmental,peng2024corporate}.
A second explanation is measurement-based. Y02 errors do not necessarily imply intentional firm-level greenwashing. Instead, they may arise because the indicator used to measure green invention is itself imperfect. The construct-validity literature emphasizes that observed indicators should not be treated as ground truth unless they are validated against the underlying construct they are intended to measure \parencite{CronbachMeehl1955,AdcockCollier2001}. Applied to Y02, this means that administrative green labels should be evaluated against the substantive technical function of the invention rather than assumed to capture true climate relevance. Recent green-patent classification studies support this concern by showing that standard green-patent indicators can generate both false positives and false negatives \parencite{Santarlasci2023}.
This study builds on the measurement-based view while remaining attentive to the firm-level symbolic explanation. We distinguish between firm-induced False Green, where firms strategically produce or frame patents as green signals, and indicator-induced False Green, where the measurement system over-includes patents with green-sounding but substantively non-climate technical language. The same observed outcome, an inflated green-patent count, can therefore arise from strategic symbolic behaviour or from structural bias in the classification infrastructure. Sections~\ref{sec:greenwashing_event} and~\ref{sec:post_llm_greenwashing} test this distinction directly by asking whether explicit environmental framing among administratively labelled green patents, and the stricter subset combining such framing with no direct climate mechanism, increased once an administrative green label existed to be obtained and, subsequently, once generative drafting tools became widely available.
\subsubsection{Administrative Type II Error: Silent Green}
Silent Green inventions are genuinely climate-relevant patents that are not tagged as green because their environmental contribution is embedded in technical functionality rather than explicitly described with environmental keywords \parencite{Block2024}. These errors are especially likely in engineering, chemistry, energy systems, transport, and heavy industry, where decarbonization benefits are expressed through process optimization, materials design, thermal efficiency, emissions-control mechanisms, power electronics, catalysts, or energy-conversion systems rather than overt climate-oriented vocabulary \parencite{Cohen2026}. Silent Green represents Y02 under-inclusion and can systematically disadvantage sectors central to industrial decarbonization \parencite{Cohen2026,Lan2025}.
A bounded-rationality perspective helps explain why this under-inclusion may be systematic rather than random. Patent examiners and administrative classification systems operate under limited time, attention, and information-processing capacity \parencite{Simon1955,Simon1957}. Classification may therefore rely partly on simplifying heuristics, including visible environmental keywords, familiar Y02 templates, and established technology-category mappings. A catalyst, thermal-control system, battery component, engine-control mechanism, or industrial process may contribute directly to decarbonization while being described primarily in engineering or chemical language. Technically complex green inventions can consequently remain invisible to an administrative indicator that is more sensitive to explicit environmental language than to functional climate relevance. Section~\ref{sec:atypicality_silent_green} tests this account directly by asking whether, within the corrected green-patent portfolio, the probability of administrative omission rises with the technological atypicality of the invention's knowledge combination.
Greenhushing or technical under-framing may provide a secondary firm-level mechanism. Greenhushing refers to the deliberate under-communication of genuine sustainability practices, while related work on strategic silence shows that organizations may withhold positive environmental information to avoid reputational scrutiny or accusations of hypocrisy \parencite{font2017greenhushing,carlos2018strategic}. In the patent context, firms or patent attorneys that do not explicitly articulate the environmental contribution of an invention may reduce the probability that examiners or automated systems recognize its climate relevance.
However, Silent Green does not require intentional under-communication. It can arise even when applicants accurately describe an invention because the environmental contribution is embedded in complex technical functionality that the Y02 classification system fails to recognize. Our framework therefore treats greenhushing as a possible contributing mechanism, while emphasizing structural measurement bias under bounded classification capacity.
The distinction between False Green and Silent Green matters because the two errors create different policy risks. Administrative Type I errors can generate a false sense of progress and reward low-impact or merely green-sounding activity. Administrative Type II errors can conceal strategically important inventions and may contribute to their under-recognition in research and investment decisions \parencite{Cohen2026}. Silent Green is therefore not simply missing data; it is a systematic form of under-recognition. Technological atypicality provides an observable implication of this bounded-rationality account. Patents that recombine technologies in unfamiliar ways impose greater cognitive and taxonomic demands because their climate contribution may not fit established templates or familiar Y02 category mappings. Administrative omission should therefore become more likely as inventions move away from conventional technological configurations. This relationship need not be monotonic, however. Extremely atypical inventions may be unusually salient, may trigger closer scrutiny, or may contain distinctive specialist signals that facilitate recognition. The bounded-capacity argument consequently allows for an inverted-U relationship in which Silent Green is most likely among moderately-to-highly atypical inventions rather than among either conventional patents or the most extreme technological outliers. The empirical analysis tests this nonlinear implication while treating it as an association consistent with bounded classification capacity, not as direct evidence of the underlying cognitive mechanism.
\subsection{Model Error Correction and Transmission}
The rapid growth of patent filings has increased the need for machine learning (ML)
methods capable of processing large-scale, multi-label, and highly imbalanced patent
data. Encoder-based models such as PatentSBERTa are particularly useful because they
provide deterministic and high-throughput mappings from patent text to
administrative classifications
\parencite{Bekamiri2024PatentSBERTa}. However, when a model is trained to reproduce
an administrative label, disagreement with the official classification does not
necessarily indicate model failure. It may instead reveal a measurement error in the
administrative label. Because PatentSBERTa is trained using administrative Y02 labels, its predictions are
not independent of the administrative classification process. Systematic errors in
Y02 may enter the training data as label noise and shape the model's decision
boundary \parencite{FrenayVerleysen2014,Northcutt2021}. At the same time,
PatentSBERTa may generate additional errors through class imbalance, threshold
selection, limitations in textual representation, or insufficient coverage of
emerging and atypical technologies.
We distinguish two principal relationships between administrative and model error:
error correction and error transmission. Error correction occurs when
PatentSBERTa challenges an incorrect administrative label and agrees with substantive
adjudication:
\[
A_i \neq M_i,
\qquad
M_i = G_i
\]
where \(A_i\) denotes the administrative Y02 label, \(M_i\) the PatentSBERTa
classification, and \(G_i\) the substantive adjudication of patent \(i\).
Within the resolved disagreement pool, PatentSBERTa corrects \num{209849}
administrative errors, comprising \num{180384} administrative Type~I errors and
\num{29465} administrative Type~II errors.
Error transmission occurs when distortions associated with the administrative
classification persist in, or become stronger through, the predictive model.
Transmission may take two forms: bias preservation and
bias amplification. Bias preservation occurs when the administrative classification and PatentSBERTa
produce the same incorrect label:
\[
A_i = M_i \neq G_i
\]
In such cases, the model preserves an administrative error rather than correcting
it. These shared errors remain hidden in disagreement-based auditing because the two
measurement systems generate no disagreement signal. The potential search space for preserved errors consists of all
dual-measurement agreement cases, which constitute the vast majority of the
corpus (Section~\ref{sec:results}). The potential search space for
preserved errors contains \num{8557649} dual-measurement agreement cases,
comprising \num{8216292} joint non-green classifications and \num{341357} joint
green classifications. These agreement cases should not themselves be interpreted
as preserved errors. Identifying preservation requires substantive adjudication of a
representative sample drawn from this space. Bias amplification occurs when the predictive model introduces additional errors
that reinforce and increase an existing administrative distortion
\parencite{Zhao2017,WangRussakovsky2021}. Cases in which the substantive audit supports the administrative label
while PatentSBERTa is incorrect represent potential sources of amplification,
because the model adds errors beyond those made by the
administrative classification. They constitute evidence of amplification only when
their technological distribution reinforces and increases an existing administrative
pattern, such as the over-recognition of ICT technologies or the under-recognition
of energy, transport, chemical, or industrial technologies. Amplification is
therefore a comparative group-level effect and cannot be inferred automatically from
an individual model error.
In this study, dual-measurement disagreement is used as a diagnostic signal for
potential classification error, consistent with methods that use model predictions
to identify potentially mislabelled observations
\parencite{Northcutt2021}. The subsequent evaluation and attribution of these
disagreements form part of a broader algorithmic-auditing approach
\parencite{Raji2020}. The purpose of the audit is therefore not merely to assess how
accurately PatentSBERTa reproduces Y02, but to determine whether each disagreement
is more consistent with an administrative error or a model error. PatentSBERTa
functions as a scalable diagnostic instrument rather than as a replacement ground
truth. Both administrative labels and model predictions are imperfect operationalizations
of the underlying construct of substantive climate relevance
\parencite{JacobsWallach2021,MishraGorana2021}. Independent substantive
adjudication is therefore required to determine the likely source and direction of
error. In the present study, this role is performed through consensus-based LLM
evaluation, while cases without sufficient agreement are retained as unresolved
rather than assigned a definitive error source.
The empirical analysis directly measures error correction and examines potential
bias amplification within full pool of dual-measurement disagreement
cases. Bias preservation is not directly observed because preserved errors are
located within the considerably larger agreement space. A comprehensive analysis of
error transmission would therefore require substantive audits of representative
agreement cases to identify preserved errors, together with sector-level comparisons
of administrative and model error rates to determine whether existing distortions
are maintained or amplified. Algorithmic auditing thus uses predictive disagreement
to locate possible measurement failures while recognizing that disagreement alone
cannot identify shared errors or establish amplification without comparative
group-level evidence.
\section{Methodology}\label{sec:methodology}
Figure~\ref{fig:researchworkflow} summarizes the research design. The workflow
separates four tasks that are often conflated in conventional classification
studies: learning the administrative label, detecting dual-measurement
disagreement, adjudicating the likely source of resolved disagreement, and testing
whether attributed errors are systematically distributed. PatentSBERTa first screens the full patent corpus and identifies cases in which its
prediction differs from the administrative Y02 classification. These disagreement
cases are then evaluated using a bagging-inspired independent LLM ensemble.
Qwen and Gemma receive the same patent information, classification rubric, and
direct-mechanism criterion, but evaluate each patent independently. Neither model
observes the other model's prediction, explanation, or confidence assessment.
Their judgments are subsequently aggregated through cross-model consensus. This
parallel design reduces dependence on the idiosyncratic reasoning and error
tendencies of any single LLM.
A case is substantively resolved only when Qwen and Gemma agree on whether the
patent contains a direct climate-mitigation or adaptation mechanism. Consensus
judgments are then used to attribute the disagreement to an administrative
Type~I or Type~II error, or to a model Type~I or Type~II error. Cases involving
mechanism disagreement or uncertain outputs are retained as unresolved rather than
being forced into an error category. The resolved cases are subsequently used to
test whether attributed errors vary systematically across technological fields,
filing periods, confidence levels, and patent characteristics.
The treatment of unresolved cases also suggests a possible
boosting-inspired sequential adjudication extension. In such a procedure,
only cases that remain unresolved after the independent first-stage audit would be
passed to a stronger adjudicating model. The adjudicator would receive the patent
text together with the conflicting judgments, cited technical evidence, and
reasoning produced by Qwen and Gemma, and would be asked to evaluate the competing
interpretations. This approach is boosting-inspired because later evaluation is
concentrated on difficult cases that the initial models failed to resolve, and the
information generated at the earlier stage is used to guide the subsequent
assessment. It nevertheless differs from classical statistical boosting because
the models are not iteratively retrained, observations are not formally reweighted,
and predictions are not combined through learned ensemble weights. To avoid
anchoring or order effects, future implementation should randomize the presentation
order of the competing judgments and require the adjudicating model to assess the
underlying patent evidence rather than simply select between the earlier answers.
\begin{figure}[H]
\centering
\setlength{\fboxsep}{7pt}
\begin{tabular}{c}
\fbox{\parbox{0.88\textwidth}{\centering
\textbf{Stage 1: PatentSBERTa screening}\\
Predict the administrative Y02 label from patent text and identify
\num{517772} dual-measurement disagreements}}
\\[0.8em]
$\Downarrow$
\\[-0.2em]
\fbox{\parbox{0.88\textwidth}{\centering
\textbf{Stage 2: Bagging-inspired independent LLM audit}\\
Qwen and Gemma independently classify the climate-related mechanism as
direct, incidental, absent, or uncertain; neither model observes the
other model's judgment or reasoning}}
\\[0.8em]
$\Downarrow$
\\[-0.2em]
\fbox{\parbox{0.88\textwidth}{\centering
\textbf{Stage 3: Consensus-based error attribution}\\
Direct-mechanism consensus attributes the disagreement to an administrative
or model Type~I/II error; mechanism disagreements and uncertain outputs
remain unresolved}}
\\[0.8em]
$\Downarrow$
\\[-0.2em]
\fbox{\parbox{0.88\textwidth}{\centering
\textbf{Stage 4: Bias and robustness analysis}\\
Test whether attributed errors vary systematically across technological
fields, filing periods, confidence levels, and patent characteristics}}
\\[0.8em]
$\Downarrow$
\\[-0.2em]
\fbox{\parbox{0.88\textwidth}{\centering
\textbf{Future extension: Boosting-inspired sequential adjudication}\\
Escalate unresolved Qwen--Gemma cases to a stronger adjudicating model that
receives the patent evidence together with both conflicting judgments and
their supporting reasoning}}
\end{tabular}
\caption{Research workflow for detecting, adjudicating, and testing administrative
and model errors in Y02 green-patent classification. The primary audit uses a
bagging-inspired ensemble in which Qwen and Gemma evaluate each disagreement
independently. A future boosting-inspired extension would concentrate additional
adjudication on unresolved cases by providing a stronger model with the competing
first-stage judgments and their supporting evidence.}
\label{fig:researchworkflow}
\end{figure}
\subsection{Data}
Our empirical analysis is based on the full corpus of \num{9075421} USPTO granted patents issued between 1962 and 2024. For each patent, we combine patent text, CPC classifications, filing year, assignee information, citation-based indicators, technological-community measures, and geolocation variables. The administrative green label is defined using the CPC Y02 classification. A patent is coded as $A_i=1$ if it has been assigned at least one Y02 code, and as $A_i=0$ otherwise. In the full corpus, \num{592387} patents are administratively labelled as Y02 green patents, corresponding to approximately 6.5\% of all granted patents.
The Y02 classification is widely used as a standardized patent-based indicator of climate-related technology. It covers technologies related to climate-change mitigation and adaptation, including renewable energy, energy efficiency, low-carbon transport, climate adaptation, carbon capture, and industrial decarbonization. As shown in Figure~\ref{fig:distribution_y02}, administratively labelled Y02 patents represent a relatively small share of the full patent universe, creating a highly imbalanced measurement setting.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{subclassy02.png}
\caption{Distribution of the administrative CPC Y02 label}
\label{fig:distribution_y02}
\end{figure}
\subsection{Dual-Measurement Disagreement as an Error Signal}
We fine-tuned PatentSBERTa \parencite{Bekamiri2024PatentSBERTa}, a BERT-based model pre-trained on patent data, with an MLP classification head to predict the administrative Y02 label from patent text. The model's purpose is not to replace the Y02 taxonomy or directly define substantive greenness. It provides a scalable second measurement of the same underlying construct and thereby creates diagnostic tension with the administrative label. After fine-tuning, the classifier was deployed to the full corpus of \num{9075421} granted U.S. patents. Among these patents, \num{592387} were administratively labelled Y02. Using the validation-calibrated decision threshold of $\tau=0.88$, the model agreed with the administrative Y02 label for \num{341357} Y02 patents but classified \num{251030} administratively green patents as non-green. At the same time, it classified \num{266742} patents without an administrative Y02 label as potentially green. Table~\ref{tab:full_deployment_matrix} reports the deployment matrix.
\begin{table}[H]
\centering
\caption{PatentSBERTa deployment matrix on the full patent corpus}
\label{tab:full_deployment_matrix}
\begin{tabular}{lrrr}
\toprule
& \multicolumn{2}{c}{Model prediction} & \\
\cmidrule(lr){2-3}
Administrative label & non-green & green & Total \\
\midrule
non-Y02 & 8,216,292 & 266,742 & 8,483,034 \\
Y02 & 251,030 & 341,357 & 592,387 \\
\midrule
Total & 8,467,322 & 608,099 & 9,075,421 \\
\bottomrule
\end{tabular}
\end{table}
If the administrative Y02 label is treated as ground truth, this deployment yields an accuracy of 0.943, precision of 0.561, recall of 0.576, and an $F1$ score of 0.569. These metrics measure agreement with Y02, not agreement with substantive climate relevance, and should therefore be interpreted cautiously. Let $M_i$ denote the model prediction. We define the disagreement signal as
\[
D_i=\mathbf{1}(A_i\neq M_i)
\]
The distinction between random error and systematic bias is central to the empirical design. Some disagreement is expected in any large-scale classification task and may reflect ordinary model noise. Our concern is whether the direction and frequency of consensus-attributed errors vary systematically across technological fields. We therefore use the model first to locate individual candidate errors and then use functional consensus to determine their source where possible. Patents with $A_i=1$ and $M_i=0$ form the downward disagreement pool. Before substantive validation, these cases may represent either administrative Type I errors (False Green) or model Type II errors (model false negatives). Patents with $A_i=0$ and $M_i=1$ form the upward disagreement pool. Before validation, these cases may represent either administrative Type II errors (Silent Green) or model Type I errors (model false positives). The procedure identifies \num{517772} total disagreement signals: \num{251030} downward and \num{266742} upward disagreements.
\subsection{Bagging-Inspired Independent LLM Validation and Error Attribution}
The complete disagreement pool was evaluated by two open-weight LLM judges: Qwen\footnote{The Qwen audit used the checkpoint \texttt{Qwen/Qwen3-32B}.} and Gemma\footnote{The Gemma audit used the quantized checkpoint \texttt{Gemma-4-31B-it-W4A16}.}. Each model returned a binary green judgment, a confidence level, a domain, a benefit type, and a short technical justification. The substantive criterion was stricter than the binary green-field criterion: an invention was treated as functionally green only when its primary technical function made a direct and credible contribution to climate change mitigation or adaptation. For model $m\in\{Q,G\}$, where $Q$ denotes Qwen and $G$ denotes Gemma, define
\[
L_{im}=
\begin{cases}
1, & \text{if benefit type is direct},\\
0, & \text{if benefit type is incidental or none},\\
\text{undefined}, & \text{if the mechanism output is uncertain or unparseable}.
\end{cases}
\]
A functional consensus judgment $L_i$ is assigned only when $L_{iQ}=L_{iG}$. If the two values differ, or if either value is undefined, the patent remains unresolved. Confidence is used to assess robustness but does not change the functional label. For downward disagreements ($A_i=1,M_i=0$), consensus $L_i=0$ identifies administrative Type I error, while consensus $L_i=1$ identifies model Type II error. For upward disagreements ($A_i=0,M_i=1$), consensus $L_i=1$ identifies administrative Type II error, while consensus $L_i=0$ identifies model Type I error. This procedure attributes \num{400186} disagreements and retains \num{117586} cases as unresolved. Table~\ref{tab:llm_validation_disagreement} summarizes the attribution outcomes.
\begin{table}[H]
\centering
\caption{Qwen--Gemma direct-mechanism attribution of model--Y02 disagreements}
\label{tab:llm_validation_disagreement}
\small
\begin{tabular}{llrr}
\toprule
Disagreement pattern & Outcome & Count & Share of all disagreements \\
\midrule
Y02 = 1, Model = 0 & Administrative Type I: False Green & 180,384 & 34.84\% \\
Y02 = 0, Model = 1 & Administrative Type II: Silent Green & 29,465 & 5.69\% \\
Y02 = 0, Model = 1 & Model Type I: false positive & 170,619 & 32.95\% \\
Y02 = 1, Model = 0 & Model Type II: false negative & 19,718 & 3.81\% \\
Either direction & Unresolved mechanism judgment & 117,586 & 22.71\% \\
\midrule
\multicolumn{2}{l}{Total} & 517,772 & 100.00\% \\
\bottomrule
\end{tabular}
\end{table}
The consensus-based corrected measure removes attributed administrative Type I errors from the Y02 count and adds attributed administrative Type II errors:
\[
\text{Consensus-corrected Green}
=
\underbrace{592387}_{\substack{\text{administrative}\\\text{Y02 count}}}
-
\underbrace{180384}_{\substack{\text{administrative Type I}\\\text{False Green}}}
+
\underbrace{29465}_{\substack{\text{administrative Type II}\\\text{Silent Green}}}
=
441468
\]
This equals 4.86\% of the full patent corpus, compared with 6.53\% under raw Y02. Because \num{50928} unresolved patents belong to the downward pool and \num{66658} belong to the upward pool, transparent extreme allocations produce a lower bound of \num{390540} and an upper bound of \num{508126}. The point estimate of \num{441468} should therefore be interpreted as a consensus-based correction, not as a final corpus-wide truth.
\subsection{The Recombinant Blind Spot: Why Atypical Green Technologies Go Unrecognized}
\label{sec:atypicality_method}
To examine whether Silent Green is consistent with bounded classification capacity, we test whether administrative omission varies systematically with technological atypicality. The analysis is restricted to the consensus-corrected green portfolio. The dependent variable equals one for a Silent Green patent and zero for a patent retained as green under the consensus correction. Thus, the comparison group contains Y02 patents not attributed to administrative Type~I error; unresolved downward disagreements retain their original administrative status and are not treated as substantively adjudicated. We measure technological atypicality following the atypical-combinations approach of \textcite{Uzzi2013}, adapting it from journal pairs in scientific papers to technology-class pairs in patents. An invention is atypical to the extent that it combines technological components that rarely co-occur relative to what chance would predict in the contemporaneous patent landscape. Let $C_i$ denote the set of CPC classes assigned to patent $i$, and let $W(t_i)$ denote the cohort of patents in a time window around its filing year $t_i$. For each unordered class pair $\{j,k\}\subseteq C_i$, let $O_{jk}$ be the number of patents in $W(t_i)$ jointly assigned classes $j$ and $k$. We compare this observed co-assignment frequency with a null distribution obtained from $R$ degree-preserving randomizations of the patent--class incidence matrix within the window, holding fixed both the number of classes per patent and the total frequency of each class \parencite{Uzzi2013}. Writing $\mu_{jk}$ and $\sigma_{jk}$ for the mean and standard deviation of $O_{jk}$ across randomizations, the pairwise conventionality score is the $z$-score
\[
z_{jk} \;=\; \frac{O_{jk}-\mu_{jk}}{\sigma_{jk}} .
\]
A large positive $z_{jk}$ denotes a conventional pairing that co-occurs more often than chance, whereas a negative $z_{jk}$ denotes an atypical, rarer-than-chance pairing. We orient the score so that higher values indicate greater atypicality and summarize each patent by its single most atypical combination:
\[
\tilde{A}_i \;=\; \max_{\{j,k\}\subseteq C_i}\!\bigl(-z_{jk}\bigr)
\;=\; -\!\min_{\{j,k\}\subseteq C_i} z_{jk}.
\]
We refer to $\tilde{A}_i$ as the maximum atypicality of patent $i$: the degree to which its rarest technological combination departs from contemporaneous convention, so that higher values indicate a more unusual technological configuration. The measure is standardized within the complete-case estimation sample,
\[
Z_i \;=\; \frac{\tilde{A}_i-\bar{\tilde{A}}}{s_{\tilde{A}}},
\]
so that one unit corresponds to one standard deviation. The complete-case sample contains \num{349898} corrected-green patents, including \num{19747} Silent Green patents. The reduction from the full corrected-green portfolio of \num{441468} patents is due to unavailable maximum-atypicality scores, for instance, single-class patents, for which no within-patent combination exists, rather than to missing filing years.
Because bounded classification capacity may generate a nonlinear relationship, we estimate a logistic model containing both linear and quadratic atypicality terms:
\[
\Pr(S_i=1\mid Z_i,t)
=
\Lambda\!\left(
\alpha+\beta_1 Z_i+\beta_2 Z_i^2+\delta_t
\right),
\]
where $S_i$ indicates Silent Green status, $Z_i$ denotes standardized maximum atypicality, $\delta_t$ represents filing-year fixed effects, and $\Lambda(\cdot)$ is the logistic function. An inverted-U relationship corresponds to $\beta_1>0$ and $\beta_2<0$, with an interior maximum at $Z^{\ast}=-\beta_1/(2\beta_2)$. The main specification uses heteroskedasticity-robust standard errors.
As a robustness test, we estimate a logistic model with both filing-year and firm fixed effects, clustering standard errors by firm. This specification compares patents within the same firm while controlling for common filing-year shocks. Firms without within-firm variation in Silent Green status do not identify the fixed-effect relationship and are excluded, leaving \num{136030} patents, \num{6909} Silent Green patents, and 734 firms.
\section{Results}\label{sec:results}
\subsection{Attributing Administrative and Model Errors}
Figure~\ref{fig:venn_disagreement_scope} clarifies the scope of the audit. Among all
\num{9075421} U.S. granted patents, \num{592387} are classified as green by the
administrative Y02 system and \num{608099} are classified as green by PatentSBERTa.
The two systems agree on \num{341357} patents as green and on \num{8216292} patents
as non-green, producing \num{8557649} agreement cases in total. These agreement
cases are not substantively audited in the present study and therefore constitute
the potential search space for shared or preserved errors.
The audit instead focuses on the \num{517772} administrative--model disagreement
cases. These comprise \num{251030} downward disagreements, in which Y02 classifies
the patent as green while PatentSBERTa classifies it as non-green, and
\num{266742} upward disagreements, in which Y02 classifies the patent as non-green
while PatentSBERTa classifies it as green. These non-overlapping regions define the
empirical audit scope in which one measurement system challenges the other.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.90\textwidth]{y02_venn_bw_600dpi.png}
\caption{Administrative Y02 and PatentSBERTa classifications across the full patent
corpus. The intersection contains \num{341357} patents classified as green by both
systems. The non-overlapping regions contain \num{517772} administrative--model
disagreements, comprising \num{251030} Y02-only green classifications and
\num{266742} model-only green classifications. The \num{8216292} patents outside
both circles are classified as non-green by both systems. Agreement cases are not
substantively audited in the present study.}
\label{fig:venn_disagreement_scope}
\end{figure}
Figure~\ref{fig:flow_error_attribution} shows how the disagreement cases are
attributed through the direct-mechanism LLM audit. The audit attributes
\num{400186} of the \num{517772} disagreements and retains \num{117586} cases as
unresolved. Among the \num{251030} downward disagreements, \num{180384} are
administrative Type~I errors, \num{19718} are model Type~II errors, and
\num{50928} remain unresolved. Among the \num{266742} upward disagreements,
\num{29465} are administrative Type~II errors, \num{170619} are model Type~I
errors, and \num{66658} remain unresolved. In the terminology of Section~\ref{sec:theory}, the audit therefore
documents \num{209849} instances of error correction, in which PatentSBERTa
successfully challenges an incorrect administrative label, and \num{190337}
model errors that constitute potential sources of bias amplification.
Disagreement therefore identifies a candidate measurement problem, while
Qwen--Gemma direct-mechanism consensus identifies the likely source of error only
for resolved cases. Administrative Type~I errors correspond to False Green patents,
where Y02 incorrectly includes a patent without a direct climate mechanism.
Administrative Type~II errors correspond to Silent Green patents, where Y02 omits a
patent containing a direct climate mechanism. Model Type~I and Type~II errors arise
when the LLM audit instead supports the original administrative classification.
Table~\ref{tab:new_error_counts_confidence} reports these consensus outcomes
together with the associated cross-model confidence levels.
\begin{figure}[H]
\centering
\includegraphics[width=0.92\textwidth]{y02_audit_flow_bw_600dpi.png}
\caption{Attribution of administrative and model errors within the
\num{517772} administrative--model disagreement cases. The direct-mechanism LLM
audit distinguishes administrative Type~I and Type~II errors from model Type~I and
Type~II errors. Unresolved cases retain their original Y02 classification and are
not reclassified. Recoding the resolved administrative errors produces a
consensus-adjusted green-patent count of \num{441468}. The optional extreme-allocation
sensitivity interval is \num{390540}--\num{508126}.}
\label{fig:flow_error_attribution}
\end{figure}
\begin{table}[htbp]
\centering
\caption{Direct-mechanism consensus, confidence, and unresolved cases}
\label{tab:new_error_counts_confidence}
\small
\begin{tabular}{lrrrr}
\toprule
Outcome
& Count
& Share
& Both high confidence
& High-confidence rate \\
\midrule
Administrative Type~I & 180,384 & 34.84\% & 177,047 & 98.15\% \\
Administrative Type~II & 29,465 & 5.69\% & 24,415 & 82.86\% \\
Model Type~I & 170,619 & 32.95\% & 168,319 & 98.65\% \\
Model Type~II & 19,718 & 3.81\% & 16,429 & 83.32\% \\
Unresolved & 117,586 & 22.71\% & 84,333 & 71.72\% \\
\midrule
Total & 517,772 & 100.00\% & --- & --- \\
\bottomrule
\end{tabular}
\end{table}
Most unresolved cases reflect substantive disagreement about whether the climate
mechanism is direct rather than low-confidence failure. Gemma identifies a direct
mechanism while Qwen does not in \num{115779} cases, whereas the reverse occurs in
\num{1752} cases. Only 55 cases contain at least one uncertain mechanism output.
Moreover, \num{84333} unresolved cases receive high confidence from both models.
High confidence therefore does not eliminate cross-model disagreement and should
not be interpreted as calibrated accuracy. After recoding only the consensus-resolved administrative errors, the estimated
green-patent count decreases from \num{592387} to \num{441468}. Because the
\num{117586} unresolved disagreement cases are not reclassified, they retain their
original administrative Y02 status: \num{50928} unresolved downward cases remain
provisionally green, while \num{66658} unresolved upward cases remain provisionally
non-green. Accordingly, \num{441468} should be interpreted as a
consensus-adjusted count rather than as a complete substantive adjudication of the
full corpus. For transparency, Figure~\ref{fig:flow_error_attribution} also reports the
extreme-allocation sensitivity interval of \num{390540}--\num{508126}. These values
are not alternative main estimates; they indicate how the consensus-adjusted count
would change under the most extreme possible classifications of the unresolved
cases.
\subsection{Illustrative Mechanisms from the Full-Corpus Audit}
\label{sec:illustrative_mechanisms}
Table~\ref{tab:illustrative_error_mechanisms} presents five especially
informative examples of each error type. The cases were purposively
selected from the consensus-resolved audit to reveal the substantive
mechanisms behind the errors and the technological bias identified in
the population-level analysis. They are illustrative rather than
representative, and a broader catalogue of examples is reported in
Appendix~\ref{app:illustrative_patent_cases}.
\begin{table}[H]
\centering
\caption{Illustrative False Green and Silent Green Patents from the
Full-Corpus Audit}
\label{tab:illustrative_error_mechanisms}
\scriptsize
\setlength{\tabcolsep}{3pt}
\begin{tabularx}{\textwidth}{
@{}
>{\raggedright\arraybackslash}p{1.35cm}
>{\raggedright\arraybackslash}X
>{\centering\arraybackslash}p{0.55cm}
>{\centering\arraybackslash}p{0.55cm}
>{\centering\arraybackslash}p{0.75cm}
>{\raggedright\arraybackslash}X
@{}
}
\toprule
\textbf{Patent}
&
\textbf{Primary technical function}
&
\textbf{Y02}
&
\textbf{TF}
&
\textbf{LLMs}
&
\textbf{Mechanism and bias signal}
\\
\midrule
\multicolumn{6}{@{}l}{
\textbf{Panel A. Administrative Type I: False Green}
\quad
($Y02=1$, TF$=0$, LLM consensus$=0$)
}
\\
\addlinespace[2pt]
\#11417179
&
Automated retail checkout using image and voice tracking within a
``shopping environment''
&
1 & 0 & 0/0
&
\textbf{Semantic homonymy and ICT over-inclusion.}
The term ``environment'' denotes a retail or digital setting rather
than the natural environment.
\\
\addlinespace
\#9537918
&
Encrypted data sharing within a distributed secure ``environment''
&
1 & 0 & 0/0
&
\textbf{Generic computing language.}
A secure computing environment is classified as environmentally
relevant despite having no direct climate function.
\\
\addlinespace
\#7811444
&
Oxidation and upgrading of heavy-hydrocarbon asphaltenes for fuel or
petrochemical use
&
1 & 0 & 0/0
&
\textbf{Counter-directional taxonomic spillover.}
The invention supports fossil-resource processing rather than climate
mitigation.
\\
\addlinespace
\#11572899
&
Pressure-exchange system designed for hydraulic-fracturing operations
&
1 & 0 & 0/0
&
\textbf{Counter-directional application.}
The technology facilitates fossil-fuel extraction despite receiving an
administrative green label.
\\
\addlinespace
\#9993540
&
Immunotherapy for the treatment of brain cancer
&
1 & 0 & 0/0
&
\textbf{Adaptation-category spillover.}
A general medical invention is included without a specific causal
connection to climate adaptation.
\\
\midrule
\multicolumn{6}{@{}l}{
\textbf{Panel B. Administrative Type II: Silent Green}
\quad
($Y02=0$, TF$=1$, LLM consensus$=1$)
}
\\
\addlinespace[2pt]
\#7986539
&
Maximum-power-point tracking and power conversion for photovoltaic
electricity
&
0 & 1 & 1/1
&
\textbf{Power-electronics opacity.}
The renewable-energy contribution is embedded in feedback control,
voltage regulation, and conversion architecture.
\\
\addlinespace
\#6266975
&
Heat-pump and engine system for energy-efficient heating and cooling
&
0 & 1 & 1/1
&
\textbf{Embedded building efficiency.}
The climate mechanism is expressed through thermodynamic engineering
rather than explicit green terminology.
\\
\addlinespace
\#11525437
&
Mechanical storage and delivery of electricity using gravitational
potential energy
&
0 & 1 & 1/1
&
\textbf{Enabling-technology omission.}
A grid-scale storage function is described through lifting and
mechanical-system terminology.
\\
\addlinespace
\#11878950
&
Oxy-fuel cement-production process that concentrates carbon dioxide
for capture
&
0 & 1 & 1/1
&
\textbf{Industrial-process opacity.}
A direct mechanism for reducing hard-to-abate cement emissions is
embedded in specialized process-engineering language.
\\
\addlinespace
\#8877486
&
Microalgae photobioreactor for carbon-dioxide mitigation during
wastewater treatment
&
0 & 1 & 1/1
&
\textbf{Cross-domain under-coverage.}
The invention combines biological carbon capture and wastewater
treatment but remains outside the administrative green indicator.
\\
\bottomrule
\end{tabularx}
\vspace{2mm}
\begin{minipage}{0.98\textwidth}
\footnotesize
\textit{Notes:}
Y02 denotes the administrative classification and TF denotes the
PatentSBERTa prediction at the validation-calibrated threshold.
The LLM column reports the two independent direct-mechanism judgments:
1 indicates a direct climate-mitigation or adaptation mechanism and
0 indicates that no such direct mechanism was identified. All ten
examples received matching high-confidence judgments from both LLMs.
The cases were selected for explanatory clarity and are not estimates
of the relative frequency of the mechanisms.
\end{minipage}
\end{table}
\subsubsection{Administrative Type I Error: False Green}
The False Green examples reveal several forms of administrative
over-inclusion. Patents \#11417179 and \#9537918 illustrate semantic
homonymy in ICT. In both cases, the term ``environment'' describes a
digital or commercial setting rather than the natural environment.
These examples help explain the strong concentration of administrative
Type I errors in computing and communications technologies. Patents \#7811444 and \#11572899 reveal a more consequential mechanism.
Their primary functions support heavy-hydrocarbon processing and
hydraulic fracturing, respectively. These are not merely inventions
with weak or incidental environmental benefits; they facilitate
fossil-resource production. Their inclusion, therefore, represents
counter-directional taxonomic spillover. Patent \#9993540 illustrates
a further form of category expansion, in which a general medical
invention is associated with climate adaptation despite lacking a
specific climate-related health mechanism.
False Green, therefore, cannot be reduced to a simple keyword error. It
can arise from semantic ambiguity, technological-category proximity,
overextension of adaptation categories, or insufficient consideration
of the invention's primary technical function.
\subsubsection{Administrative Type II Error: Silent Green}
The Silent Green examples demonstrate the opposite measurement
mechanism. Patent \#7986539 directly improves photovoltaic electricity
conversion, Patent \#6266975 provides energy-efficient heating and
cooling through heat-pump engineering, and Patent \#11525437 supplies
mechanical electricity storage for grid integration. Nevertheless,
their climate functions are expressed through power electronics,
thermodynamics, and mechanical engineering rather than overt
environmental vocabulary. The omissions extend beyond energy-system components. Patent
\#11878950 embeds carbon capture within an oxy-fuel cement-production
process, while Patent \#8877486 combines microalgae-based carbon
mitigation with wastewater treatment. These examples illustrate why
technologies crossing conventional classification boundaries, or
contributing to decarbonization through specialized industrial
functions, may remain administratively invisible.
\subsubsection{From Illustrative Errors to Measurement Bias}
The examples show what the errors look like, whereas the full-sample
analysis establishes whether they constitute measurement bias.
Dual-measurement disagreement first identifies a potentially
problematic classification. Independent LLM consensus then attributes
the disagreement according to the invention's primary technical
function. Finally, comparisons across technological fields determine
whether the attributed errors are systematically concentrated.
This combination reveals an opposing measurement structure: visible
digital or environmentally suggestive language can generate
over-inclusion, while technically embedded mechanisms in energy and
industrial technologies can generate under-inclusion. The method
therefore identifies not only individual errors but also how the
administrative indicator redistributes measured green invention across
the technological landscape.
\subsection{From Individual Errors to Systematic Technological Bias}
\label{sec:technological_bias}
An individual False Green or Silent Green patent represents a
classification error. Such errors constitute measurement bias when
their probability or direction varies systematically across
technological fields. We therefore examine error attribution separately
within the two directions of PatentSBERTa--Y02 disagreement. Among
resolved downward disagreements, we compare administrative Type I
errors with model Type II errors ($N=\num{200102}$). Among resolved
upward disagreements, we compare administrative Type II errors with
model Type I errors ($N=\num{200084}$). These within-direction
comparisons test whether error attribution varies across technologies
among patents already selected by the same disagreement pattern.
\begin{table}[H]
\centering
\caption{Association of technological field and filing period with
error attribution in the resolved disagreement pools}
\label{tab:resolved_bias_tests}
\small
\setlength{\tabcolsep}{5pt}
\begin{tabularx}{\textwidth}{
@{}
>{\raggedright\arraybackslash}p{4.0cm}
>{\raggedright\arraybackslash}X
r
r
r
r
@{}
}
\toprule
\textbf{Resolved disagreement pool}
&
\textbf{Patent characteristic}
&
\textbf{$N$}
&
\textbf{$\chi^2$}
&
\textbf{df}
&
\textbf{Cram\'er's $V$}
\\
\midrule
Administrative Type I vs.\ Model Type II
&
Primary non-Y CPC section
&
200,102
&
13,143.73
&
8
&
0.256
\\
&
Primary non-Y CPC subclass
($n\geq500$; smaller subclasses pooled)
&
200,102
&
31,542.97
&
81
&
0.397
\\
&
Filing period
&
200,102
&
1,045.97
&
6
&
0.072
\\
\addlinespace
Administrative Type II vs.\ Model Type I
&
Primary non-Y CPC section
&
200,084
&
9,294.09
&
8
&
0.216
\\
&
Primary non-Y CPC subclass
($n\geq500$; smaller subclasses pooled)
&
200,084
&
32,125.25
&
92
&
0.401
\\
&
Filing period
&
200,084
&
618.30
&
6
&
0.056
\\
\bottomrule
\end{tabularx}
\vspace{0.35em}
\begin{minipage}{0.97\textwidth}
\footnotesize
\textit{Notes:}
All Pearson chi-square tests reject independence at $p<0.001$.
The primary non-Y CPC section represents the broad technological field
derived from each patent's primary conventional CPC classification.
The primary non-Y CPC subclass identifies the more detailed technology
within that field. Y02 classifications are excluded from both
technology measures. For the subclass tests, subclasses represented by
fewer than 500 patents in the relevant resolved-disagreement pool are
combined into a residual category. Filing periods are $\leq1994$,
1995--1999, 2000--2004, 2005--2009, 2010--2014, 2015--2019, and
2020 onward. Cram\'er's $V$ ranges from zero, indicating no association,
to one, indicating perfect association.
\end{minipage}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]
{primary_cpc_subclass_false_green_silent_green_counts_expanded_v2.png}
\caption{Technological concentration of
consensus-attributed administrative errors.
Leftward bars report False Green patents and rightward bars report
Silent Green patents. The upper block contains the ten primary
non-Y CPC subclasses with the largest False Green counts. The lower
block contains the fifteen subclasses with the largest Silent Green
counts. Because a displayed subclass can contain both types of
error, both counts are reported for every subclass. Administrative
over-inclusion is especially prominent in digital data processing,
wireless networks, digital communications, semiconductor devices,
and selected medical and chemical technologies. Administrative
omission is concentrated in separation processes, catalysis,
engine-exhaust control, heat pumps and refrigeration, power
systems, batteries and fuel cells, water treatment, and other
engineering technologies. Technology descriptions are abbreviated
for readability. The figure describes the scale and technological
location of the errors; the conditional tests in
Table~\ref{tab:resolved_bias_tests} account for differences in field
size.}
\label{fig:primary_cpc_subclass_errors}
\end{figure}
Figure~\ref{fig:primary_cpc_subclass_errors} reports the technological
location of the attributed administrative errors.
The subclass-level results reveal two contrasting technological error
profiles. False Green patents are concentrated in digital and
communications technologies. Digital data processing
(G06F) contains \num{18206} cases, followed by wireless networks
(H04W; \num{11199}) and digital communications
(H04L; \num{7526}). Semiconductor devices
(H01L; \num{6145}) and control systems
(G05B; \num{3895}) are also prominent. The concentration of
\num{7996} cases in medical preparations (A61K), together with large
numbers in organic chemistry and peptide subclasses, shows that
administrative over-inclusion extends beyond ICT.
Silent Green patents display a different technological profile. The
largest concentration occurs in separation processes
(B01D; \num{2587}), followed by catalysis and process chemistry
(B01J; \num{1175}), engine-exhaust control
(F01N; \num{1055}), heat pumps and refrigeration
(F25B; \num{905}), and electrical power distribution and storage
(H02J; \num{894}). Further concentrations occur in heating and
ventilation, batteries and fuel cells, engine fuel supply, engine
control, heat and steam power plants, water treatment, electric
machines, inorganic chemistry, and vehicle technologies. These
technologies typically achieve climate benefits through physical,
chemical, thermal, or systems-engineering mechanisms rather than
through explicit environmental terminology.
The displayed counts show where the largest errors occur, but counts
alone cannot establish bias because technological fields differ greatly
in size. The conditional tests provide that evidence. Error attribution
is substantially more strongly associated with detailed CPC subclasses
than with broad CPC sections. Cram\'er's $V$ is approximately 0.40 in
both disagreement directions at the subclass level, compared with
0.22--0.26 at the section level. Associations with filing period are
much smaller ($V\leq0.072$). Thus, the source and direction of
classification error depend primarily on the specific technology rather
than on filing period or broad differences in field size.
The technological associations remain similar when the analysis is
restricted to patents for which both independent LLMs report high
confidence. Cram\'er's $V$ remains 0.392 for the administrative Type I
subclass test and 0.408 for the administrative Type II subclass test;
the corresponding section-level values are 0.246 and 0.208. The
technological structure of the errors is therefore not driven by
low-confidence adjudications. The sharply contrasting error profiles across independently defined
non-Y CPC subclasses provide clear evidence that administrative
misclassification is technology-specific rather than random.
Appendix~\ref{sec:lexical_error_structure} complements this technological
evidence by showing that the attributed source of error is also systematically
related to how each invention's function is expressed in patent language.
\subsection{The Complexity Penalty: Why Atypical Green Technologies Go Unrecognized}
\label{sec:atypicality_silent_green}
Silent Green exhibits a pronounced nonlinear relationship with technological atypicality. The observed rate rises from 2.12\% in the lowest atypicality quartile to 9.04\% in the third quartile, before declining to 4.93\% in the highest. Patents in the third quartile are therefore over four times as likely to be Silent Green as those in the lowest quartile. This inverted-U pattern indicates that administrative omission is most prevalent among moderately-to-highly atypical technologies, rather than among conventional patents or extreme outliers.
\begin{table}[H]
\centering
\caption{Silent Green rates across technological atypicality}
\label{tab:atypicality_rates}
\begin{tabular}{@{}lrrrr@{}}
\toprule
\multicolumn{5}{@{}l}{\textit{Panel A. Observed rates by atypicality quartile}}\\
\midrule
Atypicality group & Patents & Silent Green & Rate & Rel.\ to Q1 \\
\midrule
Q1: Lowest & 87,475 & 1,852 & 2.12\% & 1.00 \\
Q2 & 87,474 & 5,671 & 6.48\% & 3.06 \\
Q3 & 87,474 & 7,912 & 9.04\% & 4.26 \\
Q4: Highest & 87,475 & 4,312 & 4.93\% & 2.33 \\
\addlinespace
\multicolumn{5}{@{}l}{\textit{Panel B. Filing-year-adjusted predicted rates}}\\
\midrule
Atypicality level & Std.\ value & Adjusted rate & Rel.\ to mean & Per 1,000 \\
\midrule
Low atypicality & $-1.000$ & 4.19\% & 0.643 & 42 \\
Mean atypicality & $0.000$ & 6.52\% & 1.000 & 65 \\
Estimated peak & $0.638$ & 7.04\% & 1.081 & 70 \\
High atypicality & $1.000$ & 6.87\% & 1.054 & 69 \\
Very high atypicality & $2.000$ & 4.93\% & 0.757 & 49 \\
\bottomrule
\end{tabular}
\vspace{0.5em}
\begin{minipage}{0.95\textwidth}
\footnotesize
\textit{Notes:} Panel~A reports unadjusted Silent Green rates within equal-sized quartiles of the complete-case sample; ``Rel.\ to Q1'' is each quartile's rate divided by the Q1 rate. Panel~B reports average predicted probabilities from the logistic model with filing-year fixed effects (Table~\ref{tab:atypicality_regressions}, column~1), evaluated at the indicated standardized atypicality value while retaining the observed filing-year distribution; ``Rel.\ to mean'' indexes each predicted rate to the mean-atypicality prediction. Because atypicality is right-skewed, the fitted peak at $Z=0.638$ falls within the upper part of the Q3 range, so the model peak and the highest observed quartile rate coincide in location while differing in level, the latter being unadjusted and pooled over a range of $Z$. All rates refer to the probability of Silent Green within the consensus-corrected green estimation sample, not within the full patent corpus.
\end{minipage}
\end{table}
Multivariate analyses accounting for filing year confirm this nonlinear relationship (linear $\beta = 0.262$, $p < 0.001$; quadratic $\beta = -0.205$, $p < 0.001$), with the implied within-sample maximum occurring at $0.638$ standard deviations above mean atypicality. The effect is substantively meaningful: the predicted probability of Silent Green rises from 4.19\% (one standard deviation below the mean) to 7.04\% at the peak. This 68\% relative increase corresponds to approximately 28 additional omitted patents per 1,000 otherwise comparable inventions in the corrected-green portfolio. Crucially, the relationship persists when comparing patents within the same firm and filing year. In a firm- and year-fixed-effects specification, the nonlinear association remains robust (linear $\beta = 0.277$, $p < 0.001$; quadratic $\beta = -0.146$, $p < 0.001$), peaking at $0.945$ standard deviations above the mean. Stable firm-level differences in patenting strategy therefore cannot explain the omission gap.
\begin{table}[H]
\centering
\caption{Logistic estimates of Silent Green and technological atypicality}
\label{tab:atypicality_regressions}
\begin{tabular}{@{}lcc@{}}
\toprule
& \textbf{(1)} & \textbf{(2)} \\
\midrule
Atypicality ($Z$) & $0.262^{***}$ & $0.277^{***}$ \\
& $(0.011)$ & $(0.027)$ \\
\addlinespace
Atypicality$^2$ ($Z^2$) & $-0.205^{***}$ & $-0.146^{***}$ \\
& $(0.009)$ & $(0.019)$ \\
\midrule
Implied peak (SD) & $0.638$ & $0.945$ \\
Filing-year FE & Yes & Yes \\
Firm FE & No & Yes \\
Patents & 349,898 & 136,030 \\
Silent Green patents & 19,747 & 6,909 \\
Firms & --- & 734 \\
\bottomrule
\end{tabular}
\vspace{0.5em}
\begin{minipage}{0.85\textwidth}
\footnotesize
\textit{Notes:} The dependent variable equals one for Silent Green patents and zero for patents retained as green under the consensus correction. Atypicality is standardized within the complete-case sample (mean $0$, SD $1$). The table reports logistic coefficients with standard errors in parentheses. Column~(1) includes filing-year fixed effects with heteroskedasticity-robust standard errors; column~(2) includes filing-year and firm (GVKEY) fixed effects with standard errors clustered by firm. Because each model contains a quadratic term, the implied within-sample maximum occurs at $-\beta_{1}/2\beta_{2}$. Significance: $^{*}\,p<0.05$; $^{**}\,p<0.01$; $^{***}\,p<0.001$.
\end{minipage}
\end{table}
Together, these results are consistent with bounded classification capacity. Routine administrative tagging works well for familiar technological configurations, but omission rates rise when climate relevance is embedded in less familiar combinations. The subsequent decline in omissions among the most atypical patents should be interpreted cautiously; it may reflect greater administrative scrutiny for extreme outliers, though our analysis establishes only the nonlinear association rather than the underlying mechanism.
\subsection{Did Firms Game the Green Patent System?}
\label{sec:greenwashing_event}
The capacity explanation has a strategic rival. Rather than the taxonomy
failing to recognize atypical inventions, firms may have learned to write
patents in the language the classifier rewards, so that over-inclusion
reflects strategic green framing rather than classification error. The two
accounts imply different responses to the moments when green classification
became salient and consequential, the introduction of Y02 tagging in 2010
and the launch of the CPC in 2013. Strategic framing makes two testable
predictions: after these events, (i) False Green patents should adopt explicit
green language at higher rates than genuinely green patents, and (ii) the
prevalence of False Green should rise discretely. We test both.
To enable firm-level controls, the language test filters the \num{200102} resolved Y02 patents to include only those that can be successfully linked to a specific patenting firm and grant year. This requirement reduces the sample to \num{85541} patents (\num{79225} False Green and \num{6316} LLM-supported genuine green) from \num{3422} firms over 1964--2024. False Green status is defined here by the Qwen--Gemma consensus
rather than by the screening model, and all specifications include CPC-section
fixed effects with firm-clustered standard errors.\footnote{Formally, for each cutoff $c\in\{2010,2013\}$ we estimate $Y_{ift}=\alpha+\beta_1 FG_i+\beta_2(FG_i\times Post_{tc})+\mu_f+\delta_t+\theta_s+\varepsilon_{ift}$, where $FG_i$ identifies a consensus-attributed False Green patent, $Post_{tc}$ equals one for filing years at or after cutoff $c$, and $\mu_f$, $\delta_t$, and $\theta_s$ denote firm, filing-year, and CPC-section fixed effects. The filing-year effects absorb the main post indicator, so $\beta_2$ captures the differential change for False Green patents relative to LLM-supported Y02 patents. Standard errors are clustered by firm; the rare binary outcome uses a linear probability model to retain the high-dimensional fixed effects.}
Explicit green language is
rare throughout the corpus: only 312 of the \num{85541} patents (0.36\%)
contain any explicit green phrase, so the test asks whether that rarity shifts
differentially for False Green patents after each event. The phrase dictionary,
the matching rule, and the two derived framing outcomes are detailed in
Appendix~\ref{app:green_dictionary}.
\begin{table}[H]
\centering
\caption{Explicit green-language use by patent type and period, around the
2010 Y02 introduction and the 2013 CPC launch.}
\label{tab:greenwashing_descriptives}
\begin{tabular}{llrrr}
\toprule
Patent type & Period & $N$ & Any green language (\%) & Mean green rate \\
\midrule
\multicolumn{5}{l}{\textit{Panel A: cutoff 2010}}\\
LLM-supported genuine green & Before & \num{3089} & 1.748 & 0.239 \\
LLM-supported genuine green & After & \num{3227} & 1.611 & 0.329 \\
False Green & Before & \num{34024} & 0.194 & 0.037 \\
False Green & After & \num{45201} & 0.310 & 0.046 \\
\midrule
\multicolumn{5}{l}{\textit{Panel B: cutoff 2013}}\\
LLM-supported genuine green & Before & \num{3879} & 1.908 & 0.290 \\
LLM-supported genuine green & After & \num{2437} & 1.313 & 0.277 \\
False Green & Before & \num{42878} & 0.238 & 0.042 \\
False Green & After & \num{36347} & 0.286 & 0.042 \\
\bottomrule
\end{tabular}
\begin{minipage}{0.92\textwidth}
\vspace{4pt}
{\footnotesize \textit{Notes:} ``Any green language'' is the share of patents
containing at least one explicit green phrase; ``mean green rate'' is the
average intensity of green-phrase use. Sample: \num{85541} resolved Y02
patents with valid firm and grant year.}
\end{minipage}
\end{table}
The difference-in-differences estimates (Table~\ref{tab:greenwashing_did}) show no significant shift following the 2010 introduction of the Y02 classification, as the relative change for False Green patents was negligible for both the probability of using explicit green language ($0.373$ percentage points; SE $0.389$, $p=0.338$) and language intensity ($0.0065$; SE $0.0101$, $p=0.518$). While the 2013 CPC launch was followed by a small but statistically significant relative increase in the probability of using explicit green language ($0.835$ percentage points; SE $0.402$, $p=0.038$; not significant after a Bonferroni correction for the four related tests, which sets the 5\% threshold at $0.0125$) and a marginal increase in intensity ($0.0204$; SE $0.0110$, $p=0.065$), two factors caution against interpreting this as strategic greenwashing.
First, the absolute effect size is economically trivial, less than one percentage point compared against a baseline rate of roughly $0.25$\% among False Green patents. Second, this relative increase is essentially a statistical artifact driven by a decline in the comparison group rather than a spike among False Green patents. Specifically, after 2013, the use of explicit green language among genuinely green patents dropped from $1.91$\% to $1.31$\%, whereas it barely changed among False Green patents ($0.24$\% to $0.29$\%; Table~\ref{tab:greenwashing_descriptives}). Consequently, the observed differential reflects changes in the control group and does not indicate that False Green patents began systematically describing themselves as green in absolute terms.
\begin{table}[H]
\centering
\caption{Difference-in-differences: the differential change in explicit green
language for False Green patents after each event
(False Green $\times$ post interaction).}
\label{tab:greenwashing_did}
\begin{tabular}{llrrr}
\toprule
Cutoff & Outcome & Interaction & SE & $p$ \\
\midrule
2010 & Any green language (pp) & 0.373 & 0.389 & 0.338 \\
2010 & Language intensity & 0.0065 & 0.0101 & 0.518 \\
2013 & Any green language (pp) & 0.835 & 0.402 & 0.038 \\
2013 & Language intensity & 0.0204 & 0.0110 & 0.065 \\
\bottomrule
\end{tabular}
\begin{minipage}{0.92\textwidth}
\vspace{4pt}
{\footnotesize \textit{Notes:} $N=\num{85541}$ patents across \num{3422} firms.
Estimates are the interaction between False Green status (Qwen--Gemma consensus)
and an indicator for grants after the cutoff, with firm, filing-year, and CPC-section fixed effects and firm-clustered standard errors. ``pp'' denotes percentage points.}
\end{minipage}
\end{table}
The second test asks whether the prevalence of False Green patents spiked when green classifications were introduced. An interrupted-time-series analysis of all \num{200102} resolved administratively green patents (\num{180384} False Green and \num{19718} LLM-supported genuine green) reveals no discrete break at either event (Table~\ref{tab:false_green_rate_its}). For the 2010 Y02 introduction, the immediate level shift is actually downward (odds ratio $0.649$, $p<0.001$), with the annual False Green rate dipping from roughly 90\% in the mid-2000s to 86.76\% in 2010, before beginning a gradual upward drift (annual odds ratio $1.060$, trend $p<0.001$). Similarly, the 2013 CPC launch shows no significant immediate jump (odds ratio $0.879$, $p=0.144$), again followed by a gradual drift (annual odds ratio $1.058$, trend $p<0.001$). Both cutoffs therefore indicate a gradual post-event drift rather than a sudden, policy-induced response. As shown in the underlying annual series (Figure~\ref{fig:false_green_rate_annual}), the False Green rate fluctuates within a narrow 87--93\% band across two decades, drifting upward gradually with no visible discontinuity at either policy date.
\begin{figure}[H]
\centering
\includegraphics[width=0.82\textwidth]{false_green_rate_annual_bw_600dpi.png}
\caption{Annual False Green rate among resolved administratively green
patents, 2005--2024. Dashed vertical lines mark the 2010 introduction of
Y02 tagging and the 2013 CPC launch. The rate exhibits a dip around
2009--2010 and a gradual upward drift, with no discrete break at either
event.}
\label{fig:false_green_rate_annual}
\end{figure}
\begin{table}[H]
\centering
\caption{Interrupted-time-series comparison of the False Green rate before and
after each event (Qwen--Gemma outcome).}
\label{tab:false_green_rate_its}
{\small
\begin{tabular}{lrrrrrccl}
\toprule
Cutoff & Pre $N$ & Post $N$ & Pre rate (\%) & Post rate (\%) &
Raw diff (pp) & Immed.\ level OR & Annual OR\\
\midrule
2010 & \num{80127} & \num{119975} & 88.90 & 90.98 & 2.09 & $0.649^{***}$ & $1.060^{***}$ \\
2013 & \num{100467} & \num{99635} & 88.76 & 91.54 & 2.78 & $0.879$ & $1.058^{***}$ \\
\bottomrule
\end{tabular}}
\begin{minipage}{0.98\textwidth}
\vspace{4pt}
{\footnotesize \textit{Notes:} All \num{200102} resolved administratively green
patents. ``Immed.\ level OR'' is the discontinuous shift at the cutoff;
``Annual OR'' is the post-cutoff annual trend. Immediate-level $p<0.001$ (2010)
and $p=0.144$ (2013); annual-trend $p<0.001$ and joint level/slope $p<0.001$ for
both cutoffs. $^{***}p<0.001$.}
\end{minipage}
\end{table}
Taken together, the strategic explanation receives at most weak and mixed
support. The 2010 introduction of Y02 tagging produced no differential increase
in green language and no discrete rise in misclassification. The 2013 CPC launch
produced a small, one-margin, statistically significant relative increase in
explicit green language, but no discrete jump in the False Green rate, and an
effect that is economically small and partly an artifact of declining framing
among genuinely green patents. Against the strong, consistent,
technology-specific error structure documented in Sections~\ref{sec:technological_bias} and~\ref{sec:atypicality_silent_green}, we read the event evidence as leaving
classification capacity, not applicant strategy, as the dominant mechanism,
while acknowledging that a modest strategic contribution after 2013 cannot be
excluded.
\subsection{Redrawing the Map of Green Inventions Landscape}
\label{sec:corrected_landscape}
The systematic error structure produces a materially different map of
green invention. The administrative Y02 indicator identifies
\num{592387} green patents, corresponding to 6.53\% of the
$\sim$9 million granted patents. Removing the \num{180384}
consensus-attributed False Green patents and adding the \num{29465}
Silent Green patents produces a consensus-corrected estimate of
\num{441468}. The correction therefore removes a net \num{150919} patents, reducing
the measured green-patent population by 25.5\% and its corpus share
from 6.53\% to 4.86\%. Unresolved downward disagreements retain their
administrative green status, whereas unresolved upward disagreements
remain administratively non-green. Table~\ref{tab:corrected_y02_landscape}
reports the resulting family-level changes.
\begin{table}[H]
\centering
\caption{Changes in the measured green-patent landscape by Y02 family}
\label{tab:corrected_y02_landscape}
\small
\setlength{\tabcolsep}{4pt}
\begin{tabularx}{\textwidth}{
@{}
>{\raggedright\arraybackslash}X
r
r
r
r
r
@{}
}
\toprule
\textbf{Y02 family}
&
\textbf{Admin.}
&
\textbf{Corrected}
&
\textbf{Change}
&
\textbf{Admin. share}
&
\textbf{Corrected share}
\\
\midrule
Y02A --- Adaptation
& 58,818 & 30,004 & -49.0\% & 9.93\% & 6.80\% \\
Y02B --- Buildings
& 58,362 & 46,224 & -20.8\% & 9.85\% & 10.47\% \\
Y02C --- Carbon capture
& 6,611 & 7,160 & +8.3\% & 1.12\% & 1.62\% \\
Y02D --- ICT energy efficiency
& 62,373 & 20,237 & -67.6\% & 10.53\% & 4.58\% \\
Y02E --- Energy
& 194,201 & 177,916 & -8.4\% & 32.78\% & 40.30\% \\
Y02P --- Industrial production
& 80,199 & 44,017 & -45.1\% & 13.54\% & 9.97\% \\
Y02T --- Transportation
& 112,506 & 98,908 & -12.1\% & 18.99\% & 22.40\% \\
Y02W --- Waste and wastewater
& 19,317 & 17,002 & -12.0\% & 3.26\% & 3.85\% \\
\midrule
\textbf{Total}
&
\textbf{592,387}
&
\textbf{441,468}
&
\textbf{-25.5\%}
&
\textbf{100.00\%}
&
\textbf{100.00\%}
\\
\bottomrule
\end{tabularx}
\vspace{0.35em}
\begin{minipage}{0.97\textwidth}
\footnotesize
\textit{Notes:}
The corrected measure removes consensus-attributed False Green patents
and adds consensus-attributed Silent Green patents. Because Silent
Green patents have no observed administrative Y02 family, their
candidate family is assigned using their audited technical domain or
primary non-Y CPC position. Family-level results are therefore
descriptive; the non-Y CPC tests in
Table~\ref{tab:resolved_bias_tests} provide the primary statistical
evidence of technological bias.
\end{minipage}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]
{full_green_patent_counts_by_y02_family_corrected_logic.png}
\caption{Green-patent counts under the administrative Y02
classification, the TF model, and the consensus-corrected measure
by Y02 family or inferred candidate family. The administrative
measure contains \num{592387} patents, the TF model identifies
\num{608099}, and the consensus-corrected measure contains
\num{441468}. For TF-positive patents without an administrative Y02
label, candidate-family assignment is based on the audit domain or
a primary non-Y CPC crosswalk. Candidate-family assignment affects
the presentation of the TF distribution but does not cause a patent
to enter the corrected measure; only cross-model consensus
identifying a direct climate mechanism does so.}
\label{fig:corrected_green_landscape}
\end{figure}
As Figure~\ref{fig:corrected_green_landscape} illustrates, correction
changes both the scale and composition of measured green
innovation.
ICT energy efficiency declines from \num{62373} patents to
\num{20237}, a reduction of 67.6\%, while Industrial production
declines by 45.1\% and Adaptation by 49.0\%. By contrast, Energy
declines by only 8.4\%, and Carbon capture records a small net increase
after recovered Silent Green patents are added. The relative technological landscape consequently shifts. Energy rises
from 32.8\% of administrative green patents to 40.3\% of the corrected
population, and Transportation rises from 19.0\% to 22.4\%. ICT energy
efficiency falls from 10.5\% to 4.6\%, while Industrial production
falls from 13.5\% to 10.0\%. Thus, even where the broad ranking of large
technology families remains stable, their measured importance changes
substantially.
Taken together, the non-Y CPC and corrected-landscape analyses establish two
distinct aspects of measurement bias. The non-Y CPC tests show that the
probability and direction of administrative error depend on the
underlying technology. The corrected landscape shows the substantive
consequence: the raw Y02 indicator overstates digitally visible and
broadly framed technologies while under-recognizing climate mechanisms
embedded in energy, transport, chemical, and industrial technologies.
\section{Discussion and Policy Implications}\label{sec:discussion}
Our findings reframe a debate that has been treated as a puzzle about
firm behavior as, in substantial part, a problem of measurement. The
ESG innovation disconnect looks less like a paradox of virtuous firms
failing to innovate and more like the predictable output of an
indicator whose classification capacity is unevenly distributed across
the technological landscape. This has a direct methodological
implication: studies that use raw Y02 counts as a dependent or
explanatory variable inherit a direction-specific bias that will
attenuate or invert estimated relationships between environmental
performance and green inventions, particularly for heavy-industry
samples.
\subsection{From Classification Error to Measurement Bias}
The findings demonstrate why individual classification errors and
systematic measurement bias should not be treated as equivalent. A
False Green, Silent Green, model false positive, or model false
negative is an instance-level classification error. Such an error does
not, by itself, establish a failure of construct reliability, which
concerns whether similar inputs to a measurement model yield similar
outputs \parencite{JacobsWallach2021}. Bias arises only when the
probability or direction of those errors varies systematically across
relevant groups. The present results instead provide evidence concerning construct
validity, particularly content and substantive validity. These forms of
validity concern whether an operationalization captures the substantive
nature of the intended construct and incorporates the relevant, and
only the relevant, properties \parencite{JacobsWallach2021}. False
Green patents indicate that Y02 includes inventions without a direct
climate mechanism, whereas Silent Green patents indicate that it omits
inventions whose technical mechanisms are directly climate-relevant.
More importantly, these errors are not evenly distributed across
technologies. Within the resolved disagreement pools, administrative
Type I attribution varies substantially across CPC sections
($V=0.256$) and subclasses ($V=0.397$), while administrative Type II
attribution exhibits similarly strong technological variation
($V=0.216$ and $V=0.401$, respectively, see Table \ref{tab:resolved_bias_tests}). False Green errors are
concentrated in ICT and communications technologies, whereas Silent
Green errors are concentrated in engineering-intensive technologies
such as energy conversion, emissions control, batteries, heat pumps,
industrial separation, and water treatment. The problem is therefore
not simply that Y02 contains classification errors, but that its
operationalization of green invention exhibits technology-specific
threats to construct validity: Y02 does not simply make mistakes,
it makes different mistakes in different parts of the technological
landscape.
The atypicality analysis adds a second form of evidence that the under-recognition is systematic. After controlling for filing year, the predicted probability of Silent Green increases from 4.19\% among low-atypicality patents to 7.04\% at the estimated peak, a difference of 2.85 percentage points. In practical terms, this corresponds to approximately 28 additional omissions per 1,000 patents in the corrected-green portfolio. The persistence of the inverted-U relationship in the firm- and year-fixed-effects model indicates that it is not explained solely by stable differences in firms' industries, patenting strategies, or environmental orientation. The magnitude is therefore substantively meaningful as a classification effect, although it should not be interpreted as a monetary economic effect because the analysis does not estimate consequences for patent value, investment, subsidies, or realized environmental impact.
An exploratory temporal-language analysis examines whether resolved False Green patents became more likely than LLM-supported Y02 patents with a direct climate mechanism to use explicit environmental phrases after candidate cutoffs in 2010 and 2013 (see Appendix~\ref{app:green_dictionary} for the predefined phrase dictionary). The analysis contains \num{85541} resolved Y02 patents with valid firm and filing-year information, including \num{79225} False Green patents, \num{6316} LLM-supported Y02 patents, and \num{3422} firms. Explicit green language is rare: only 312 patents (0.36\%) contain at least one phrase from the predefined dictionary. Models with firm, filing-year, and CPC-section fixed effects and firm-clustered standard errors show no differential increase after 2010 in either phrase incidence ($+0.373$ percentage points, $p=0.338$) or framing intensity ($\beta=0.0065$, $p=0.518$). For 2013, the probability of using a green phrase shows a small, nominally significant relative increase ($+0.835$ percentage points, $p=0.038$), though the intensity of that language does not ($\beta=0.0204$, $p=0.065$). Crucially, this slight shift in probability is misleading. It is driven primarily by a decline in explicit green language among genuinely green patents, rather than a meaningful absolute spike among False Green patents. The evidence therefore provides no robust support for the claim that firms deliberately increased strategic green framing to exploit the classification system.
This distinction also clarifies the objective of the audit. We do not
assume that all error can or should be eliminated. Some residual error
is unavoidable in any classification of millions of heterogeneous
technical documents. The central objective is to reduce the systematic
component of error that alters sectoral comparisons and substantive
conclusions. Random measurement error primarily reduces precision and
weakens construct reliability. By contrast, systematic over- or
under-inclusion across technological groups indicates measurement bias
and threatens construct validity \parencite{JacobsWallach2021}.
The Error-as-Signal framework changes the interpretation of model
failure. When administrative labels are imperfect, disagreement with
those labels can contain information about weaknesses in the underlying
indicator. PatentSBERTa is therefore valuable not because it provides a
new unquestioned truth, but because it generates a scalable competing
measurement that identifies cases requiring deeper evaluation. The
subsequent LLM audit turns disagreement into attributed evidence about
administrative and model failure.
\subsection{Implications for Algorithmic Auditing}
The results reveal both the value and the limitations of
dual-measurement disagreement auditing. Error correction becomes observable when
PatentSBERTa challenges an incorrect administrative label and the
substantive audit supports the model prediction. By contrast, bias
preservation remains hidden when the administrative system and the
model produce the same incorrect classification, because no
disagreement signal is generated. This limitation is particularly
important when a model is trained on the administrative labels that it
is subsequently used to audit. Future auditing designs should therefore
combine disagreement-based sampling with targeted substantive reviews
of representative agreement cases, especially high-confidence
agreements, emerging technologies, and domains underrepresented in the
training data.
The correction also shows why auditing should examine differences
between administrative categories rather than only aggregate accuracy.
Consensus correction disproportionately reduces measured adaptation:
the number of adaptation patents falls from \num{58818} to
\num{30004} ($-49.0\%$), compared with a decline from \num{533569} to
\num{411464} ($-22.9\%$) across mitigation-oriented Y02 families.
Adaptation's share of measured green invention consequently decreases
from 9.93\% to 6.80\%. This difference is driven by substantial
over-inclusion in Y02A: \num{29347} administratively classified
adaptation patents are identified as False Green, whereas only
\num{533} omitted patents are added as Silent Green. Algorithmic audits
should therefore test not only whether classification errors occur, but
also whether their frequency and direction differ across substantive
categories.
The bagging-inspired multi-model design separates scalable screening
from substantive adjudication. PatentSBERTa screens the full patent
corpus, while Qwen and Gemma independently evaluate whether each
flagged invention contains a direct climate-mitigation or adaptation
mechanism. Cross-model consensus permits scalable attribution of the
\num{400186} resolved disagreements, whereas the \num{117586}
mechanism disagreements form a transparent escalation set for
evaluation by a stronger adjudicating model or by domain experts. This
design reduces dependence on the reasoning and error tendencies of any
single model while preserving uncertainty rather than forcing
unresolved cases into definitive error categories.
\subsection{Implications for Research and Policy}
For research, the immediate implication is that raw Y02 counts are not
a neutral dependent variable. Analyses of induced innovation, green
directed technical change, and the returns to environmental performance
should either use corrected counts, bound their estimates using the
sensitivity range reported here, or demonstrate robustness to the
direction-specific error structure we document. For policy, indicators
built on Y02, innovation scoreboards, green industrial-strategy
dashboards, and sustainable-finance screens, currently over-reward
sectors whose inventions are described in digital language and
under-credit the mechanical, chemical, and process technologies that
carry much of the burden of decarbonization. This interpretation extends \textcite{Cohen2026}, who document that traditional energy firms make substantial contributions to green innovation despite weak ESG recognition. Our results suggest that measurement bias in Y02 may cause existing indicators to underestimate, and potentially widen, the documented ESG innovation disconnect. Reallocating measurement attention is cheaper
than reallocating capital wrongly.
The Error-as-Signal Framework is not specific to patents. Every
administrative indicator that governs the net-zero transition,
ESG ratings, the EU Taxonomy of sustainable activities, national R\&D
and green-jobs statistics, and green-bond and green-fund labels, is
a classification layered by finite human or algorithmic capacity onto a
heterogeneous underlying reality, and every one can be audited by
confronting its labels with an independent functional assessment and
treating the disagreements as signal. Just as \textcite{Berg2022}
decomposed ESG-rating divergence into scope, measurement, and weight,
and \textcite{LernerSeru2022} catalogued the biases of patent-count
data, we offer a portable diagnostic for the taxonomies of directed
technical change. We therefore propose a research agenda: (i) apply the
framework to ENV-TECH, the IPC Green Inventory, and the EU Taxonomy to
establish whether the bias toward information technology and against
heavy industry is scheme-specific or general; (ii) extend the audit to
non-US patent offices to separate classification-capacity effects from
CPC-diffusion effects; (iii) test whether correcting the indicator
changes the estimated returns to green inventions and the measured
direction of induced innovation; and (iv) develop capacity-aware
classification protocols that allocate examiner and algorithmic
attention toward the atypical, mechanically complex inventions the
current system renders silent.
\section{Conclusion and Limitations}\label{sec:conclusion}
We audited the construct validity of the most widely used green-patent
indicator against the full corpus of \num{9075421} USPTO granted
patents. Treating disagreement between the administrative Y02 label and
an independent model as a diagnostic signal, and attributing errors
only where two independent large language models agree on an
invention's primary technical function, we identified \num{180384}
False Green patents and \num{29465} Silent Green patents. Correcting
the consensus-attributed administrative errors reduces the green-patent
population by 25.5\%, from \num{592387} to \num{441468} patents, with
sensitivity bounds of \num{390540} to \num{508126}. The errors are
structured, not random: over-inclusion concentrates in digital
technologies and omission in mechanical, chemical, and industrial
process technologies; omission rises with technological atypicality in
an inverted-U pattern; and event analyses show no discrete jump in
misclassification when green classification became salient, with at most
a small relative increase in explicit green framing after the 2013 CPC
launch. The bias originates primarily in the classification system's
finite capacity rather than in applicant strategy, and it supplies a
measurement-based explanation for the ESG innovation disconnect.
The instruments we use to measure the green transition are not neutral
mirrors of innovation; they are lenses that magnify some technologies
and dim others. When an indicator systematically over-credits the
visible language of digital efficiency and under-recognizes the hard,
unglamorous engineering of decarbonizing heavy industry, it does not
merely mismeasure the transition --- it misdirects it. Administrative
indicators do not simply count innovation; they redistribute it.
Auditing these instruments is therefore not a technical afterthought
but a first-order task of climate-innovation policy, and the
Error-as-Signal Framework offers a practical way to begin.
The contribution is therefore not simply a more accurate count. The study distinguishes individual error from systematic bias and shows that opposing errors can cancel numerically while continuing to distort the sectoral map of green inventions. Random or non-differential error mainly reduces precision; systematic error changes who appears to innovate, which technologies appear strategically important, and where policy and finance are directed. The findings reinforce recent work on patent classification and the ESG innovation disconnect. Text-based studies have shown that conventional green-patent measures can diverge substantially from narrower functional definitions \parencite{Santarlasci2023}. Our analysis extends that work by attributing resolved disagreement to administrative or model error and by showing that the consensus-attributed administrative errors follow distinct technological patterns. It also provides a patent-side measurement mechanism that may contribute to the limited recognition of industrial firms central to decarbonization under commonly used ESG and patent indicators \parencite{Cohen2026}. The exploratory temporal-language test does not alter this measurement-based conclusion: it yields no robust evidence that firms deliberately increased green patent framing after either candidate Y02 cutoff.
Several limitations remain. First, disagreement-based auditing cannot directly identify error convergence, in which Y02 and PatentSBERTa agree but are both wrong. Targeted sampling of agreement cases is therefore necessary to estimate residual shared error. Second, LLM-bagging approach consensus is an audit-based reference judgment rather than independently observed ground truth. The \num{117586} unresolved cases, including \num{84333} high-confidence disagreements, show that model confidence is not equivalent to accuracy and that additional frontier-model, expert, or human adjudication is needed. Third, the reported error rates are conditional on entry into the PatentSBERTa--Y02 disagreement pool and should not be interpreted as corpus-wide precision or recall for either system. Fourth, the substantive standard identifies direct and credible climate contributions but does not estimate realized environmental impact after commercialization. Future work should extend the framework to agreement-case audits, multilingual and older patent corpora, additional patent offices, and other administrative indicators used in ESG and innovation research.
The broader implication is that errors should not be viewed only as observations to remove. Once independently validated and examined across domains, they become evidence about how a measurement infrastructure sees, misses, and redistributes the phenomenon it is intended to represent. For green inventions, reducing systematic measurement bias is essential if policy and capital are to recognize the technologies most relevant to a net-zero transition.
\printbibliography