1 · Introduction: Why Society Needs a Measurable Notion of AI Influence

1.1 · A public cognitive gap

In 2026 the general public encounters a stream of names — GPT, DeepSeek, Claude, Gemini, Kimi, Qwen, Codex, Trae, Cursor — and a stream of benchmarks that purport to rank their “capability.” None of this tells a non-specialist where the field stands. Capability leaderboards describe models; product announcements describe vendors; neither describes the structure of influence — which communities are changing how society works, how fast, and in what direction. The vision behind this system is that AI concept and information equality is part of the public-good promise of AI itself: the same way a market index lets a layperson read the state of an economy, a well-built influence index should let a layperson read the state of the AI landscape. To the best of our knowledge from searchable public information, no public-facing index of AI influence existed for a general audience at the time of writing (Section 2.5); the index family presented here is an attempt to fill that gap with a measurement that remains defensible to scrutiny. The gap is not merely aesthetic. A public that cannot distinguish “who is responsible for the AI changing my work” from “which vendor sent me the most emails” cannot form preferences over the direction of the field, cannot hold the relevant communities to account, and is structurally vulnerable to whichever narrative is loudest. Information equality about the AI landscape is therefore not a luxury layered on top of the technology; it is part of the social contract that the technology implies. An influence index is one instrument for providing it. It is not the only instrument, and this paper is careful (Sections 1.6 and 12.4) not to claim it is.

1.2 · Capability is not influence

The central premise of this system is that capability and influence are different quantities, and that confusing them corrupts public reasoning about AI. We give three arguments, each with its most serious counterargument. Argument 1: capability can be gamed. Whenever a quantity is used as a target, the statistical regularities behind it tend to collapse (Goodhart, 1984; Strathern, 1997; Campbell, 1979); the phenomenon has been taxonomized and formalized for machine learning under labels such as Goodhart's law variants (Manheim and Garrabrant, 2018), reward over-optimization (Gao et al., 2023), reward hacking (Skalse et al., 2022; Amodei and others, 2016; Pan et al., 2024), and benchmark data contamination (Balloccu et al., 2024; Koch et al., 2021). A capability score is precisely such a target: it can be optimized by engineering effort in a way that need not track whether the models are used, discussed, and built upon by society. Measuring “capability” therefore measures a necessary input to influence, not influence itself. Section 2.2 assembles the empirical record of this failure mode at length, because it is the strongest argument for why a benchmark composite cannot substitute for an influence measure. Counterargument and reply. One might reply that capability is the only “objective” quantity available, and that a noisy capability score beats an influence score whose very definition is contestable. The reply is twofold. First, the objectivity of a capability score is overstated: the empirical literature on benchmark contamination (Balloccu et al., 2024; Koch et al., 2021; Hendrycks et al., 2021) and on machine-assessor bias (Zheng and others, 2023; Wang and others, 2023) shows that “objective” scores are themselves artifacts of an evaluation pipeline that can be and has been gamed. Second, the counterargument concedes the premise we deny — that a univariate capability ranking is the right object — and merely disagrees about noise. An influence measure is not a noisier capability score; it is a different quantity built from different inputs (use, discussion, iteration, diffusion), and it can be designed to be harder to game, as Section 7.5 argues. Argument 2: omniscient and value-based assessments are value-contaminated. Any attempt to judge which models are “better for humanity” is forced to adopt a standpoint on what society should value, and different societies reach different standpoints; the resulting judgments inherit the values of their assessors (Winner, 1980; Douglas, 2004; Douglas, 2009; Elliott, 2022). Even a well-intentioned composite “quality” score quietly embeds a value choice. An influence measure should therefore minimize the amount of evaluation it performs and instead record observable social fact: who is used, by how much, by whom, and with what consequence. The empirical literature on LLM-as-a-judge confirms the practical bite of this worry: machine assessors exhibit systematic self-preference, position, and value biases (Zheng and others, 2023; Wang and others, 2023; Santurkar and others, 2023). (We return to this theme in Section 7.2.) Counterargument and reply. One might reply that “influence” is itself a value-laden construct — that choosing to weight usage over, say, safety is a value choice, so the system merely moves the problem. The reply is that the claim is narrower and testable. The system does not claim to be value-free; it claims that the set of value choices it must make is smaller and more transparent than the set a holistic quality judgment requires, and that each choice is designed to be auditable (Section 5.2 maps every weight to the influence definition). This is the distinction Douglas (Douglas, 2009) draws between values that enter a measurement as inescapable background and values that silently do the measuring: the design goal is to move as many choices as possible from the latter to the former. Whether it succeeds is an empirical claim about the audit trail, not a verbal claim. Argument 3: value arises through use, discussion, iteration, and diffusion. Capability becomes social value only through a chain: a model is released; it is used in products and workflows; it is discussed publicly; it is iterated upon by developers and researchers; and it diffuses into edge devices, small models, and new applications. We therefore define the object of measurement as a composite of these four channels rather than as a single score of technical ability. This definition is operationalized in Section 5, given a construct-validity treatment in Section 4.2, and formalized in Section 6. Counterargument and reply. One might reply that use without capability is impossible, so capability is a sufficient statistic after all. The reply is empirical: usage, discussion, and ecosystem stock are visibly not monotone functions of capability. A model that is adequate and free, that is embedded in a large developer ecosystem, or that is diffused through edge devices can generate more social value than a slightly more capable model locked behind a high price and a closed walled garden. Section 11 shows this structure in the first FA snapshot: capability leaders are not ecosystem leaders. The premise of “capability $\Rightarrow$ influence” fails as a measurement identity.

1.3 · A measurable object: AI Originators

Because capability alone is not influence, and because influence flows from communities that produce and propagate models rather than from models in isolation, the system takes as its measurement unit the AI Originator: the community that continuously creates, releases, and iterates AI models and thereby supplies AI productive capacity to society. This is a class-level object — deliberately not a company, a brand, or a legal person (Section 3). The measurement question becomes: how much influence does each AI Originator exert on society, measured relative to the others, on a stable scale, in something close to real time? The choice of unit deserves scrutiny, because every index is hostage to its unit. Measure influence at the level of the model and you lose the organization that funds, hosts, and propagates it; measure at the level of the company and you miss open-source collectives, joint initiatives, and the boundary where a company and its research lab are not the same actor; measure at the level of the country and you average away the very communities whose behavior you want to observe (Section 2.1). The Originator class is written to admit all of these forms. We are honest about the current roster (Section 3.6): at present the included set is company-dominated, because the verifiable public sources we rely on are organized that way. The class definition is the more general object; the roster is a time-indexed instantiation of it.

1.4 · Why an index

A single number, a single line, and a single consistent standard of measurement are not simplifications of the science; they are a design choice about the presentation layer. The design premise is that publics reason well with quantities they can track over time and compare across names. The index family presented here (Section 6) therefore behaves like a market-semantics quote machine for AI influence: a real-time sequence of values, periodic review-and-release of the slow anchor, and a suspend–resume release mechanism when measurement integrity is at risk. The semantics of a price quote are adopted; the financial products themselves are not, and the disclaimer in Section 13.6 makes this boundary explicit. The “why an index” question has a scientific as well as a communication answer. A scale that can be summed (the flagship's conservation law), compared across time (the settlement calendar), and disaggregated (the index family) is a scale designed to be auditable. A single published number is not a simplification that hides the science; it is the smallest object on which the science's invariants can in principle be checked by an outsider. This is the sense in which the index form is load-bearing for trustworthiness, not merely for communication.

1.5 · Contributions

The contribution is best stated at the level of the method system rather than of any single component: a measurement unit (the AI Originator, defined as a class), a measurement scale (relative, centered, real-time, with a conservation law at its center), and a trust architecture (asymmetric disclosure: open what is hard to manipulate, closed what is easy to manipulate). Individual methodological elements have precursors — model-level relative ratings, organizational capability scores, country-level AI indexes, citation-style influence measures — but we found no system combining an originator-set measurement unit with a centered, real-time, conservation-anchored scale (Section 2.5). Two properties deserve the modest framing of design claims rather than “firsts”: the system is, to our knowledge, the first public influence measure for AI Originators as contributors, and the first to give that influence a relative, centered, real-time scale. A third contribution is methodological and runs through the whole paper: the system is presented as a construct-validity case rather than as a “dashboard.” Section 4 treats “influence on society” as a theoretical construct in the sense of Cronbach and Meehl (Cronbach and Meehl, 1955) and Messick (Messick, 1995), derives a content-valid operationalization from the definition, and states the convergent and discriminant evidence an observer would need before trusting the index. This is unusual for index engineering, and it is deliberate: an influence index that cannot say what would falsify it is a publication instrument, not a measurement.

1.6 · Scope and non-goals

This system describes an assessment-method layer; it does not describe a data product, does not produce investment information, and does not constitute financial information, investment advice, a credit assessment, a reputation judgment, or a basis for individual decisions (Section 13.6). The system's current measure is bounded by the sources it can verify; it reflects what has already happened, not a prediction of what will happen (Sections 8.9 and 12.4). Non-goals are stated as carefully as goals: the system does not adjudicate model quality, does not rank “best models for X,” does not predict future winners, and does not allocate responsibility in the governance sense (Section 2.3). What it does is narrower and, we argue, still hard: it assigns a defensible, auditable-by-design, real-time number to the question “how much social influence does each AI Originator currently exert, relative to the reference set?”

Back to contents  |  2 · Prior Measurement Systems and the Positioning of This Work →