Digital thesis
An index and a coefficient for measuring economic and cultural value in the age of AI
Emre Arıcı / ARC Studios - Tam Bağımsız Fikir Stüdyosu
Digital thesis · Conceptual model and research design
A Proposed Index and Coefficient for Measuring Economic Value and Science, Art and Culture Value in the Age of Artificial Intelligence
The author owns an independent studio that uses generative AI tools in its production. This creates a direct relationship between the subject of the study and its author. In the proposed research design, the author's own outputs will not be included in the sample; if they are included, they will be reported as a separate analysis.
The idea that economic value and cultural value are two distinct but connected concepts has become established in the cultural economics literature through the work of Throsby. This study does not claim to put forward a new theory of value; its contribution is to turn that distinction into a scale that can be scored at the level of the unit of production, and to define a coefficient that summarizes the volume–value relationship. The full text of Throsby's 1999 article and Mazzucato's (2018) book could not be accessed directly, and statements about these two sources rest on publisher summaries and later commentary. Some of the method sources in the bibliography were not reopened and compared page by page during this work; it is recommended that the bibliographic details be verified against the relevant databases before any printed use.
Arıcı, E. (2026). The value-producing economy: A proposed index and coefficient for measuring economic value and science, art and culture value in the age of artificial intelligence [Digital thesis, version 1.0]. ARC Studios.
Generative AI tools have lowered the unit cost of production and increased the output volume of solo and micro-scale producers. The measurement literature treats this output mainly through quantity and price; it does not measure the quality of the output, its durability, or the accumulation it leaves for the future. This study conceptualizes the value of AI-assisted micro-production in two layers: economic value (E) and science, art and culture value (K). Each layer is represented by a five-criterion scale scored from 0 to 10. The two scores are not summed; they are read jointly as a two-dimensional position, the Value Creation Index, which places each output in one of four quadrants. The interaction between the layers is analyzed through four mechanisms (funding, short-termism, accumulation, and erosion of trust), and their theoretical foundations are discussed.
The second contribution is the Value Production Coefficient (VPC), defined as the elasticity of a producer's value mass with respect to output volume. The coefficient is shown to equal the ratio of marginal to average yield, to be invariant to multiplicative rescaling of scores, to be decomposable by criterion, and to extend to the ecosystem level. An AI-mediated variant, defined through the effects of AI adoption on volume and on average quality, offers an estimation route that needs no exclusion restriction. The study reports no empirical findings. It proposes a two-phase research design for testing the reliability and validity of the scale, estimating the coefficient with panel data, and testing the interaction mechanisms, together with nine hypotheses (one precondition and eight substantive). The dual framework of economic and cultural value is due to Throsby; the contribution of this study lies in making that framework scorable at the level of the individual output and in summarizing the volume–value relationship in a single parameter.
Keywords: value measurement, cultural capital, generative AI, micro-production, value creation index, value production coefficient
JEL classification: E01, O33, Z10
Generative AI tools have markedly lowered the unit cost of producing software, text, images and audio content. The cost of reproducing information and cultural goods had already approached zero with digitization; what is new is the falling cost of the first copy, that is, of bringing the initial version of a software release, a song or a book into existence. Production activities that once required a team, a studio or a publishing house can now be carried out by a single person, which expands the productive capacity of micro-scale producers. Recent work that conceptualizes this shift treats AI agents as a kind of "digital co-founder" of the solo venture (Rezazadeh and Bonehgazy, 2025). These studies, however, describe the process; they do not separately discuss what the value of the output is or how it should be measured.
The expansion of productive capacity also changes what is scarce. The binding constraint on production is increasingly not the ability to produce but the ability of what is produced to be noticed, trusted and used. Simon (1971) seems to anticipate this shift when he argues that an abundance of information creates a poverty of attention: the abundance of a resource makes the scarcity of the attention that can be spent on it more visible. Within this framework, the extent to which a rise in production volume raises the value produced remains an open question. If average quality falls as volume rises, the value of total output does not grow in proportion to volume and may even decline.
The problem addressed in this study has two layers. The first concerns measurement. Official statistics cannot see a substantial part of AI output, because this output is either not priced in the market or remains an intermediate input, consumer surplus or unrecorded production (Makridis and Brynjolfsson, 2026; Corrado, 2025; Spittler and Patel, 2026). Even when output is made visible, its qualitative dimension, namely its originality, durability and the accumulation it leaves for the future, lies outside these frameworks. The visibility metrics used by digital platforms, such as likes, followers and view counts, likewise do not reflect the cultural or social value of an output (Hutchinson, 2024).
The second layer is conceptual. The idea that economic value and cultural value are distinct but connected is well established in cultural economics (Throsby, 1999, 2001), yet this distinction has not been turned into a scale that can be scored at the level of the unit of production, nor has it been tested in the context of the productive capacity that generative AI gives to micro-producers. Nor does the literature contain a parameter that summarizes the relationship between production volume and value mass. The core problem can therefore be stated as follows: as cost approaches zero, does the increase in volume correspond to an increase of the same proportion in the economic and cultural value masses; if it does not, in which quality dimensions does this dilution appear, and how do the two value layers affect one another in the process?
The purpose of the study is to conceptualize the value of AI-assisted micro-production within a two-layer framework, to turn this framework into a scorable scale, to translate the interaction between the layers into testable hypotheses, and to define a coefficient that summarizes the relationship between production volume and value mass. Four research questions are set for this purpose.
The contribution of the study is gathered at four points. First, the dual framework of economic and cultural value is brought down to the level of the unit of production and converted into two scales of five criteria each. Second, the interaction between the two layers is formulated as four mechanisms, with their theoretical foundations and the strength of those foundations stated. Third, the Value Production Coefficient is defined, and its relation to average and marginal yield, its scale invariance, its decomposability by criterion and its extension to the ecosystem level are shown. Fourth, a research design is presented for testing the reliability and validity of the scale and for estimating the coefficient with a strategy that moves it closer to a causal interpretation.
The position of these contributions in the literature must be carefully delimited. The study does not claim to have developed a new theory of value. The distinction belongs to Throsby; this study focuses on making the distinction operational at the production level and on measuring its relationship with volume. The value of the contribution depends on whether the scale can be applied reliably to real data and on whether the coefficient shows informative variation across producers, fields and periods.
The study is limited to micro-producers who have no employees or work with at most two people and who use generative AI tools in their production. The unit of analysis is the separately assessable unit of production that a producer puts out in a given period, referred to as an output. No normative judgment is intended about which cultural product is "more valuable"; measurement is restricted to observable and discussable properties such as originality, durability and accumulation.
The study contains no empirical findings. What is presented is a conceptual model and a research design prepared for testing it. The hypotheses are accordingly stated not as expectations but as propositions to be tested and capable of being falsified. The literature review rests on openly accessible sources and is not a systematic review; this limit is noted again in the relevant sections and in the declarations.
The key concepts used in the study are summarized in Table 1. The definitions are functional for the purposes of the study and do not aim to cover the full range of definitions in the relevant literature.
| Concept | Definition in this study |
|---|---|
| Output | A separately assessable unit of production that a producer puts out in a given period (a software release, a song, a book, a video, an application). |
| Production volume (Q) | The number of outputs of a producer in a period. |
| Micro-producer | A producer who has no employees or works with at most two people and who uses generative AI tools in production. |
| Economic value (E) | The degree to which an output finds a real counterpart in the market and does so by creating new value; scored from 0 to 10 on five criteria. |
| Cultural value (K) | The degree to which an output joins the stock of knowledge, art and culture left to future generations; scored from 0 to 10 on five criteria. |
| Value mass (V) | The sum of the normalized scores of a producer's outputs in a period; computed separately for each layer. |
| Value Creation Index (VCI) | The two-dimensional position formed by the (E, K) score pair of an output, and the reading of that position as one of four quadrants. |
| Value Production Coefficient (VPC) | The elasticity of value mass with respect to production volume. |
| Dilution | Value mass rising more slowly than volume, that is, a VPC below 1. |
The second chapter sets out the conceptual ground of the study: the development of the concept of value in the literature, the distinction between creating and extracting, the dual value system and cultural capital, the economics of information and creativity, attention scarcity and quality uncertainty. The third chapter reviews the measurement literature, the work on platform metrics, the micro-production literature and the findings on model collapse, and identifies the gap in the literature. The fourth chapter defines the economic and cultural layers, the interaction mechanisms and the Value Creation Index. The fifth chapter presents the Value Production Coefficient formally. The sixth chapter contains the research design and the seventh the hypotheses. The eighth chapter discusses the possible meanings of the findings, counter-arguments, limitations and future work, and the ninth concludes the study. The scoring rubric, the notation table and the scoring form are given in the appendices.
This chapter gathers the theoretical material on which the measurement proposal rests. It begins with the place of value in economics and the limits of the national accounts, and then turns to the distinction between creating and extracting value, the dual system of economic and cultural value, and cultural capital. The second half of the chapter brings together theoretical material on the economic nature of information, attention scarcity, quality uncertainty and the functioning of the field of cultural production, in order to supply the rationale for the model built in later chapters.
In the history of economics, value has been discussed along a spectrum running from the labour theory of value to marginalist theories of utility and price. Contemporary national accounts represent a practical settlement of this debate: value is read as the monetary sum of transactions realized in the market. The limits of this settlement have been known since the founders of the accounts. Kuznets (1934) warned that the welfare of a nation can scarcely be inferred from a measurement of national income. Coyle (2014) showed that gross domestic product is a collection of historically formed answers to a measurement problem, and that the boundaries of what counts as production are drawn for political and practical reasons as well as scientific ones. Stiglitz, Sen and Fitoussi (2009) presented a commission report arguing that the emphasis in measuring economic performance should shift from production to well-being.
The shared lesson of this literature is that production volume is not an adequate indicator of value. Activities outside the market do not enter the accounts, and every priced transaction is deemed to have created value. The value of goods that are unpriced or whose price deviates from their real benefit, such as free digital products and new variety, is either understated or entirely invisible in conventional measurement (Brynjolfsson, Collis, Diewert, Eggers and Fox, 2019). This study shares the same concern but at a different point: in AI output the problem is not only invisibility; the visible quantity says nothing about quality.
Mazzucato (2018) argues that economic theory has not drawn the distinction between value creation and value extraction clearly, and that this ambiguity allows actors who merely redistribute or erode existing value to present themselves as value creators. In her view, this ambiguity has led rent and profit to be confused and has distorted measures of growth and output. Mazzucato's examinations consist of cases from finance, pharmaceuticals and Silicon Valley; micro-scale producers and generative AI lie outside them. As reported in a review of the book, creators are those who invest for the long term and act with a social vision, parties who create value and do not merely take it, while extractors are parties who seek rent through monopoly advantage, short-term share support or financial manoeuvres.
This distinction forms the theoretical basis of the second criterion of the economic layer (E2). The question is not what the price of an output is but whether that price corresponds to a new benefit or to a redistribution of existing value. In the context of generative AI the distinction takes on particular importance, because part of the content that can be produced in large volumes at low cost provides no new benefit beyond redirecting existing attention or demand.
Throsby (1999, 2001) argued that economics, which traditionally speaks of physical, human and natural capital, should add cultural capital to this classification as a fourth type. Cultural capital is an asset that embodies, stores or gives rise to cultural value in addition to whatever economic value it may possess; the asset may be tangible, such as a building or a painting, or intangible, such as a tradition, a work or a craft skill (Throsby, 2003). Like other capital assets, it produces economic value as both a stock and a flow: a historic building can be sold, and visitors can pay to come and see it. But unlike ordinary capital assets, cultural capital carries an additional and different kind of value, namely cultural value (Throsby, n.d.).
In Throsby's framework economic value can be expressed in monetary terms and rests on the amount people are willing to pay for something. This value has three parts: use value arising from direct consumption; non-use value, composed of existence, option and bequest values that arise from public-good-like attributes and have no market counterpart; and positive externalities. Cultural value, by contrast, is the name given to a broad set of non-monetary values; it has no single unit of account, it is multi-dimensional and it changes over time. Hidalgo-Villota (2025) reports that Throsby decomposes cultural value into aesthetic, spiritual, social, historical, symbolic and authenticity components, and also mentions emotional value as an element that goes beyond functional rationality.
The relationship between the two values is not one-directional in this framework. The two values may be complementary, so that increasing returns emerge for both; they may also trade off within the same decision, so that increasing one requires giving up some of the other. According to Throsby, the resolution of trade-offs is a deliberative and political process based on a full assessment of both values. The measurement proposal is consistent with this logic. The use component of economic value can be observed from transactions, while non-use value is usually estimated with stated-preference methods. Cultural value is broken into components, the components are scored according to degrees of agreement with qualitative statements, the scores are aggregated with or without weights, and the assessment is carried out by teams in which economists and field experts work together. Subjective judgments are accepted as unavoidable; the aim is to establish transparent and shared standards (Throsby, n.d.). The scoring rubric and the two-rater design of this study are a direct extension of that proposal.
As reported by Hidalgo-Villota (2025), the literature is divided on the direction of effect: some argue that cultural value largely explains market price, whereas Throsby's reading is that culture shapes the economy and not the reverse. Consumption feeds cultural value, accumulates cultural capital and supports economic development. Formulating the interaction mechanisms in both directions (Section 4.6) turns this disagreement into a problem of hypotheses.
Bourdieu (1986) argues that the social world cannot be reduced to capital in the only form economics recognizes, and treats capital in three fundamental forms: economic, cultural and social. Economic capital is directly convertible into money, whereas cultural and social capital become economic capital only under certain conditions. Cultural capital exists in three states: embodied (durable dispositions of mind and body), objectified (cultural goods such as books, paintings and instruments) and institutionalized (educational qualifications). Embodied capital is acquired through personal time and labour, cannot be transferred to another instantly, and perishes with its bearer. The ownership of objectified capital can be transferred, but the ability to use these goods properly depends on embodied capital.
The fact that conversion requires time, labour and loss is the most useful aspect of Bourdieu's framework for this study. Economic capital is the ground that frees families from pressing needs and allows them to spend time acquiring cultural capital; the intergenerational transmission of cultural capital, in turn, often works invisibly. Economic success therefore does not convert into cultural accumulation automatically; conversion is lossy and conditional. Throsby (1999) deliberately uses the concept of cultural capital apart from its Bourdieusian sense in sociology. Abduljalil, Komosany and Bahkali (2026), however, regard the two approaches as complementary rather than opposed: Bourdieu's sociological view explains how cultural value acquires meaning within social positions, taste and symbolic power, while Throsby's economic view makes visible the cultural assets that carry value beyond price. In this study, Bourdieu's contribution is the idea that conversion between forms of capital is conditional and lossy; this idea forms one of the supports of the claim that the two value layers are not substitutes for one another (Section 4.5).
The economic nature of knowledge is directly related to the production-level criteria of cultural value. Arrow (1962) argued that, because knowledge is non-rival and partly non-excludable, its producer cannot capture the full return and investment in knowledge is under-supplied by the market. Romer (1990) showed that the non-rivalry of ideas, that is, the fact that one person's use of an idea does not prevent another's, is fundamental to endogenous growth. Weitzman (1998) proposed a recombinant growth model, arguing that new ideas arise largely from new combinations of old ideas; in this model the number of combinations grows rapidly with the stock of ideas, and processing capacity becomes the limiting factor.
The implication of this literature in the generative AI context is as follows: once the act of recombination becomes cheap and fast with AI, the combination itself ceases to be a scarce resource. What becomes scarce is the ability to sift out which of the combinations carries a new piece of knowledge or form, and the original input that goes into the combination. The originality criterion (K1) therefore aims to measure not recombination but the element of novelty that makes a contribution. The accumulation-effect criterion (K3) is the practical counterpart of the non-rivalry of knowledge: an output carries the public character of knowledge when it is offered in a form on which others can build, and loses that character when it is offered in a closed and non-replicable form. Benjamin (1935/2008) argued that technical reproducibility weakens the "aura" of the work of art, that is, its unique presence and authenticity. Applying this view to generative AI is contested; even so, the kinship with Throsby's authenticity component points to the theoretical lineage of the K1 criterion.
Simon (1971) noted that what information consumes is the attention of its recipients, and that an abundance of information creates a poverty of attention. In environments where production volume grows at low cost, this observation has two consequences. The first is that a limited resource such as attention has the character of a common pool: each producer who adds to the pool increases their own gain while possibly reducing the total yield of the pool. Hardin's (1968) classic argument on the tragedy of the commons supplies the general logic of this structure; what is used here is the similarity of structure and not a direct application of the argument to output. The second is that if value is proportional to attention and total attention is fixed, an increase in volume does not raise total value.
This second consequence defines a limiting case for the coefficient presented in Chapter 5. If total attention is A and the number of outputs is Q, the attention share of each output is A/Q; when value is taken to be proportional to attention, total value remains equal to A and does not change when volume changes. In this limiting case the elasticity of value mass with respect to volume is zero. Real situations will lie somewhere between this extreme and proportional growth, and this position is what the coefficient aims to measure. The limiting case is not an empirical claim but a theoretical reference that bounds the interpretation of the coefficient.
Akerlof (1970) showed that in markets where quality is uncertain for buyers, low quality can drive high quality out of the market. Nelson (1970) distinguished between search goods and experience goods, noting that the quality of experience goods can be learned only after consumption. Cultural and information products are mostly experience goods. Spence (1973) showed that where quality cannot be observed directly, only costly signals that the high-quality party can send at low cost are informative.
Generative AI affects this structure in three respects. First, when the cost of production falls, the act of producing no longer carries the signal of effort and care it once did; the existence of a work is no longer evidence that its producer exerted effort. Second, as output volume grows, the cost to the buyer of trying each item rises; learning quality through experience becomes more expensive. Third, the greater difficulty of telling quality apart prepares the ground for an Akerlof-type selection mechanism to operate: indistinguishable low-quality abundance reduces the value of quality signals and erodes the returns of original producers. This argument forms the theoretical basis of the fourth mechanism (M4, erosion of trust). The limit of the argument is that whether generative AI actually makes quality harder to discern is an empirical question; this study does not answer it but proposes a measurement design with which it can be answered.
Bourdieu (1993) treated cultural production as a field, with its own rules and distribution of capital, in which works are produced and evaluated. Position within the field is determined not only by economic success; symbolic recognition is also a determinant of position. In listing the economic properties of the creative industries, Caves (2000) emphasizes the unpredictability of demand (the nobody-knows property), the infinite variety of products, and the fact that creative workers care about the quality of the work enough to sacrifice income (the art-for-art's-sake property). Brynjolfsson, Hu and Smith (2003) showed that the increased product variety in online book sales made a substantial contribution to consumer surplus. The conclusion drawn from these studies is that variety can carry value, but that this value depends on the variety being genuinely differentiating.
Micro-producers work in this field in an environment where demand is uncertain and where many different outputs are tried to learn which will catch on. A fall in the cost of production increases the number of trials. This gives rise to two opposing effects at once: as trials multiply, the probability that a successful output emerges rises; on the other hand, average quality in total output may fall. The Value Production Coefficient aims to gather the net of these two effects under a single parameter.
The review in this chapter covers the measurement literature, recent work on the measurement of cultural value, the micro-production literature and findings on model collapse. Sources were selected for their relevance to the study's four research questions, and openly accessible texts were given priority. The review is not a systematic one; the search strategy, search dates and exclusion criteria were not fixed in a protocol. The full text of Throsby's 1999 article and of Mazzucato's (2018) book could not be accessed, and these two sources were assessed through publisher summaries and later commentary. Forecasts of one-person billion-dollar companies circulating in industry publications were not used as evidence in this review, because the examples examined contained neither data nor method.
The literature on measuring the economic impact of AI is growing rapidly, and its common concern is that official statistics capture this impact incompletely. This concern has a historical precedent: Solow's (1987) observation that the computer age did not show up in the productivity statistics established the idea that the reflection of general-purpose technologies in productivity is delayed and hard to measure. Brynjolfsson and colleagues (2019) argued that the value of new and free goods does not enter the accounts and proposed GDP-B as a complementary measure.
Makridis and Brynjolfsson (2026) list three reasons why the economic impact of AI is under-measured: investments in models, datasets and in-house tools are recorded as operating expenses and not as capital; quality change and service flows are inadequately measured; and part of the value arises outside the market, as consumer surplus and cross-spillovers in free or bundled tools. For the short term, the authors propose an index that tracks AI usage intensity by sector and region; for the medium term, satellite accounts that separately track AI infrastructure investment, services as intermediate inputs, task-level changes in labour and household production. The most notable aspect for the present study is their argument that intangible liabilities such as data leakage, loss of privacy and erosion of intellectual property must also be taken into account; this approach is consistent with the design of the E5 and K5 criteria, in which external costs are deducted from net contribution.
Highfill, Wasshausen and Prunchak (2025) discuss how the U.S. Bureau of Economic Analysis could track AI production through the existing supply–use tables and how a thematic satellite account could be built. AI is not treated as a separate type of asset; sales of custom software are tracked as software investment, software developed for own use as own-account software and research and development investment, and use as an intermediate input within the value added of the producing sector. The study lists difficulties such as the absence of an AI-specific product category, the lack of consensus on a definition of AI, the limits of reliable source data and the absence of AI-specific price indices; it leaves second-order effects and non-market value outside its scope and does not address cultural value.
In assessing Coyle and Poquiz's "Making AI Count", Corrado (2025) accepts that the national accounts respond slowly to structural change: the fact that software publishing is still classified together with newspaper and book publishing is an example. She largely agrees with Coyle and Poquiz's proposals to treat time as a dimension of measurement, to develop outcome-oriented measures and to move to more detailed data, but diverges by arguing that total factor productivity still captures AI-driven cost reductions and that growth accounting retains its conceptual primacy. Spittler and Patel (2026) call the AI output that the national accounts cannot see "dark output" and define it in two forms: spending that disappears from the accounts as work formerly done by people moves to AI, and new work that would not be done at all without AI because it would be too expensive. The authors note that token counts are not a measure of value, since the same number of tokens can correspond to an empty output or to a decision that changes strategy; the absence of a standard quality unit in service production is the core measurement problem in which volume and value come apart. The same authors mark part of their proposed monitoring framework as speculative and describe the size of dark output as "opaque".
The debate reported by Nolan (2026) is a visible example of this uncertainty. A policy note by Korinek and McKelvey estimates that AI produced about 250 billion dollars of economic activity in 2025 and that U.S. growth would look about four points higher if official statistics reflected the gains in AI capabilities, and presents this as an upper bound and not as a forecast. Coyle, for her part, finds the scale and speed overstated; she says AI is mostly used in producing other products, that GDP counts final goods, and that firm-level systematic evidence of productivity is limited, and she notes that the real problem is the absence of consensus on what should be measured.
The common feature of this literature is that it treats AI output through quantity, price and effects spread across the community. The quality of the unit of production, that is, its originality, durability and the accumulation it leaves for future generations, does not appear in these frameworks as a dimension of measurement. The measurement proposed in this study is designed to complement the macro frameworks at the micro level and not to compete with them.
Hutchinson (2024) argues that the practice of digital platforms measuring the contribution of content creators with proprietary metrics such as likes, followers, views and shares is problematic because of the limited relation of these metrics to cultural and social value. The author cites an influencer index in Australia as an example: creators with high follower counts were found unsuitable by agencies running real campaigns. Because metrics also determine which content becomes visible, they push creators towards visibility and mainstream appeal and can narrow cultural production. Hutchinson treats the social processes that seek value through the concept of "value mediation", draws on Schwartz's theory of basic values, and proposes developing qualitative measures for authenticity, representation and social impact, and judging success by long-term cultural impact and not by short-term popularity. The article is largely conceptual: the proposed measures are not operationalized, and the section on generative AI is limited. The author also accepts that it is unclear how such measures, for example "truthfulness" or "morality", could be reliably captured.
Deriya and Huynh (2026) reach a similar point starting from the music industry. Structural measurement systems track countable signals such as listens, saves and skips; they do not see relational capital based on community, personal story and cultural resonance. The authors present the example of a synthetic AI-generated band reaching one million monthly listeners as an indication of the weakness of visibility metrics in this context. The recommendations rest on artist-centric royalty models, super-fan tiers and the use of relational labour as a signal in hiring and contracting decisions. The study is a research report that has not been peer reviewed, extends its findings in music to fields such as journalism and consulting by analogy, and offers no workable method for measuring relational value; the authors also accept that judgment should be tied to a view of value only once the data are sufficient.
The problem of metrics becoming targets was expressed as a general law in the measurement literature by Campbell (1979): the more a quantitative indicator is used in social decision-making, the more exposed it becomes to corruption pressures and the more it distorts the process it is meant to monitor. This warning applies to the index of this study as well: if the VCI is used as a target, producers may shape production in ways that game the scoring criteria. The research design (Section 6.8) aims to limit this risk by having raters evaluate without knowing the producer or the level of AI use, and by using the index as a research instrument and not as a reward.
The component-based scoring proposed by Throsby (n.d.) is the basic model of expert-based measurement. Abduljalil, Komosany and Bahkali (2026) propose an interpretive framework, which they call "joint expert evaluation", that combines Bourdieu's and Throsby's understandings of cultural capital. In the framework, a group of qualified experts jointly evaluates a cultural asset within its own time and place; value arises not from individual opinion or market price but from discussion and consensus. In the case study on the Yemeni dagger, a three-stage consultation was held with twelve experts, and for a criterion to count as a core consensus element it had to be endorsed by at least eight experts, that is, two thirds. The authors present the framework as an exploratory conceptual tool, not a validated or generalizable model, and note that the findings depend on the local expertise of the panel.
The two-rater design of this study is close to expert-based measurement but not identical to it. In joint evaluation the experts discuss and reach consensus; here two raters score independently and agreement is reported statistically. The two approaches are not alternatives to each other, and comparing them in future work would be useful (Section 8.6).
Rezazadeh and Bonehgazy (2025) propose a conceptual framework that treats AI agents as a "digital co-founder" and shows the process of solo ventures in three recursive stages: shaping the vision, testing reality and scaling reality. Value creation is accepted here implicitly as solving a problem for a defined audience, and no explicit definition is given. The study reports no empirical measurement; measures to be tied to the stages, and hypotheses to be tied to outcomes such as income stability, growth, well-being and identity coherence, are left to future research. The authors also record difficulties of the process such as uncertainty, cognitive load, decision fatigue and dependence on AI. The literature on the AI-enabled transformation of micro-production thus describes the process; it does not measure the quality or the quantity of the value produced. This study comes into play at precisely this point.
Shumailov and colleagues (2024) showed that training generative models on data produced by the previous generation of models leads to a degenerative process they call "model collapse". In this process the data generated by one generation pollutes the training set of the next, and later models misperceive reality. The authors distinguish early and late collapse: in early collapse the model begins to lose information at the tails of the distribution, and in late collapse the distribution converges to one very different from the original, often with much lower variance. Three sources of error compound: statistical approximation error arising from finite sampling, the main source; and two secondary sources arising from the limited expressive power of the model and from the learning procedure. In the language-model experiment, the OPT-125m model was fine-tuned on wikitext2, and by the ninth generation the output had turned into repetitive and incoherent text. When a random ten percent of the original human data was kept at each generation, the degradation remained only mild. From this the authors conclude that human-produced data will become more valuable as model-generated content increases online, and that preserving the tails of the distribution, that is, low-probability events, is also important for fairness, since these events mostly affect minorities. Limitations are also noted: the experiments cover a small number of generations, the case the theory actually approximates is infinitely many generations, the speed of collapse depends on the problem, and the LLM experiments use fine-tuning instead of training from scratch.
Seddik and colleagues (2024) model iterative training with a linear softmax classifier with single-context, one-hot representations, and separate two regimes: training only on synthetic data and training on a mixture of original and synthetic data. They show that collapse is unavoidable in the fully synthetic regime, while in the partially synthetic regime collapse can be avoided if the share of synthetic data remains considerably smaller than the share of original data. The study's representation assumptions are simplified and it covers only statistical approximation error.
These findings were demonstrated for the training of machine-learning models. They are not direct evidence for cultural production; they are used as an analogy supporting the intuition that, if cultural production takes earlier synthetic production as input, original and rare elements at the tails of the distribution may be lost. The functional limit of the analogy is that it motivates the K5 criterion and the dilution reading of the coefficient; it is not treated as the justification of any empirical claim.
The studies reviewed are summarized in Table 2. The macro-measurement literature is oriented to quantity and price; the cultural economics literature offers the dual value framework but does not apply it to micro-producers who produce with AI; recent work on measuring cultural value identifies the problem but offers no scorable scale; and micro-production frameworks do not measure value. A fourth gap is the absence of an approach that summarizes the relationship between production volume and value mass in a single parameter.
| Study | Focus | What it measures or addresses | What is missing for this study |
|---|---|---|---|
| Mazzucato (2018) | Theory of value | The creation–extraction distinction | Micro-producers and AI are outside its scope |
| Throsby (1999, 2001) | Dual value | Economic and cultural value, cultural capital | No scoring at the level of the unit of production or in the AI context |
| Makridis and Brynjolfsson (2026) | Macro measurement | AI investment and usage intensity | The quality of output is not measured |
| Highfill et al. (2025) | National accounts | Reflection of AI production in GDP | Cultural and non-market value are outside its scope |
| Spittler and Patel (2026) | Invisible output | Output the national accounts do not see | The lack of a quality unit is identified; no solution is offered |
| Hutchinson (2024) | Cultural analytics | Limits of platform metrics | Metrics are not operationalized |
| Deriya and Huynh (2026) | Relational value | What countable signals fail to see | No workable measurement method; not peer reviewed |
| Abduljalil et al. (2026) | Expert-based measurement | Joint expert evaluation | Single case, not validated, not at the level of the unit of production |
| Rezazadeh and Bonehgazy (2025) | Solo venture | AI-assisted process | Value is neither defined nor measured |
| Shumailov et al. (2024) | Model collapse | Loss of diversity | Model training only; not evidence for culture |
These four gaps determine the position of this study. First, the dual value framework must be converted into a scorable scale. Second, the interaction between the layers must be turned into testable mechanisms. Third, a coefficient must be defined that summarizes the relationship between volume and value. Fourth, a research design must be set out to test the reliability and validity of these instruments. The next three chapters respond to these gaps in turn.
In this chapter value is defined in two layers, the scale of each layer is built, the interaction mechanisms are presented together with their theoretical foundations, and the Value Creation Index is defined as a two-dimensional position. Chapter 5 builds the Value Production Coefficient, which summarizes the volume–value relationship, on top of this structure.
The model rests on five assumptions. First, production activity consists of discrete outputs that can be assessed separately. This is an approximation for services produced as a continuous flow, but it is reasonable for the typical outputs of micro-producers, such as software releases, songs, books and videos. Second, each output can be scored on two separate dimensions through defined criteria. This assumes not that cultural value is measurable, but that measurement can be carried out with transparent standards and without hiding subjective judgment (Throsby, n.d.). Third, economic and cultural value are separate dimensions; one does not stand in for the other (Section 4.5). Fourth, production volume is observable, and outputs can be counted from public records. Fifth, the scoring definitions are fixed within a given research period; comparison across periods relies on this assumption.
The set of outputs that producer i puts out in period t is denoted Jit, and the number of elements in this set, that is, production volume, is denoted Qit. Each output j receives two scores: an economic score Ej and a cultural score Kj. Each score is the sum of five criteria, and each criterion takes the value 0, 1 or 2.
Thus Ej and Kj take integer values between 0 and 10. For use at the aggregate level, the scores are normalized to the 0–1 range.
Three points about the structure of the scale must be stated explicitly. First, the criterion scores form an ordinal scale; according to Stevens's (1946) distinction between levels of measurement, operations such as sums and means are meaningful on an ordinal scale only under the assumption that the intervals are equal. Treating the total score as an interval scale is an assumption in this study and will be tested with ordinal methods in Chapter 6. Second, the criteria that make up E and K are not reflective indicators of one another; they are formative indicators that bring together distinct aspects. Bollen and Lennox (1991) argue that reflective and formative indicator models must be distinguished and that conventional notions of reliability and measurement cannot be applied to formative indicators. Diamantopoulos and Winklhofer (2001) state that in building an index from formative indicators the content must be fully specified, the indicators identified, collinearity checked and external validity tested. This study aims to meet these four conditions through the criterion rationales in Sections 4.3 and 4.4, the rubric in Appendix A, the dependence checks in Section 6.6 and the comparisons with external criteria. Third, equal weighting is an initial choice; weighting will be addressed in the sensitivity analysis.
The economic layer measures whether an output finds a real counterpart in the market and whether it obtains this by creating new value or by redistributing existing value. The five criteria of the layer ask the following questions.
E1, real demand. Does someone demand the output by paying for it or by using it regularly? In Throsby's (n.d.) framework economic value can be expressed in the amount people are willing to pay for something; the observable evidence of demand is payment or regular use. Since non-use values have no counterpart in the market, this criterion targets the observable component of use value.
E2, creating or extracting. Does the output create a new benefit or a cost reduction, or does it redistribute existing value through arbitrage, intermediation or attention-seeking? This criterion brings Mazzucato's (2018) creation–extraction distinction down to the unit of production. The distinction rests on the need to separate price from value; the same price may correspond to a new benefit or to a redistribution of existing value.
E3, non-substitutability. Can anyone with access to AI produce the same output, or does the output have an original input? The rationale rests on economic reasoning: when the cost of production approaches zero and there are no barriers to entry, the return on an output that anyone can produce is eroded by competition. Under these conditions, economic value that persists comes mainly from an original input: domain knowledge, proprietary data, community trust. The original input is also the element that becomes scarce as the cost of recombination falls, as discussed in Section 2.5.
E4, durability. Is the income or use a one-time flow, or does it accumulate into a stock, that is, an asset? As in the discussion of cultural capital (Throsby, 1999), the distinction between flow and stock is fundamental to understanding the continuity of value.
E5, sharing of value and external costs. How is the value created divided among the producer, the user and society, and is the net contribution positive once external costs such as copyright and privacy are deducted? Makridis and Brynjolfsson (2026) argue that the net contribution of AI cannot be understood without accounting for intangible liabilities such as data leakage, loss of privacy and intellectual property disputes; this criterion carries the same logic to the level of the unit of production.
The cultural layer measures whether an output joins the stock of knowledge, art and culture left to future generations. Throsby (2001) breaks cultural value into six components and proposes assessment by component; K1–K5 are inspired by these components but adapted to the production level and do not coincide with them one to one.
K1, originality. Does the output add something new, or does it merely reshuffle what exists? The criterion is the production-level counterpart of Throsby's authenticity component and observes the distinction discussed in Section 2.5: the act of recombination is no longer scarce in itself; what is scarce is the element of novelty within the combination. The criterion aims not to reward recombination but to tell apart the contribution of novelty.
K2, durability and transmissibility. Will the output be understood, found and used years later? According to Bourdieu (1986), the capacity to benefit from objectified cultural capital depends on embodied capital; transmissibility therefore covers not only accessibility but also intelligibility independent of context.
K3, accumulation effect. Does the output feed the production of others? It rests on the non-rival nature of knowledge (Arrow, 1962; Romer, 1990): an output offered in a form on which others can build carries the public character of knowledge, while an output offered in a closed form loses that character.
K4, cultural memory. Does the output enrich the language, identity and memory of a community? It is related to Throsby's social, historical and symbolic components. The observable evidence for the criterion is community feedback and expert evaluation.
K5, future cost. Does the output create synthetic noise or cultural pollution, and does it erode original producers? The criterion carries the liability logic of Makridis and Brynjolfsson (2026) into the cultural layer and is fed by analogy from the findings on model collapse (Shumailov et al., 2024); the limits of the analogy are noted in Section 3.6.
| Criterion | Property measured | Theoretical basis |
|---|---|---|
| E1 | Real demand | Throsby (n.d.): economic value, willingness to pay |
| E2 | Creating and extracting | Mazzucato (2018) |
| E3 | Non-substitutability | Reasoning on competition and free entry (this study); Weitzman (1998) |
| E4 | Durability | Stock–flow distinction; Throsby (1999) |
| E5 | Sharing of value | Makridis and Brynjolfsson (2026) |
| K1 | Originality | Throsby (2001), authenticity component; Weitzman (1998); Benjamin (1935/2008) |
| K2 | Durability and transmissibility | Throsby (1999); Bourdieu (1986) |
| K3 | Accumulation effect | Arrow (1962); Romer (1990); Throsby (2003) |
| K4 | Cultural memory | Throsby (2001): social, historical and symbolic components |
| K5 | Future cost | Makridis and Brynjolfsson (2026); Shumailov et al. (2024), by analogy |
Not adding the two scores into a single composite score is a deliberate choice. The literature on composite indicators notes that summing components rests on an assumption of compensability: weakness in one component can be offset by strength in another (Nardo et al., 2008). A linear aggregation takes the form S = aE + bK, in which case the rate of substitution between E and K is constant and equal to a/b; that is, a loss of a fixed amount of K can always be compensated by a gain of a/b points of E.
This substitution structure conflicts with the theoretical sources of the study. According to Throsby, cultural value has no single unit of account and covers non-monetary values; the two values may trade off within the same decision, but the resolution of this trade-off is a negotiation and not a calculation (Throsby, n.d.). According to Bourdieu, conversion between forms of capital is conditional, time-consuming and lossy (Bourdieu, 1986); economic capital therefore cannot be converted into cultural capital at a fixed rate and without loss. Neither view supports the assumption that a high economic score can offset a low cultural score. For this reason outputs are not summed; they are read jointly, without compensation, in a two-dimensional position.
The two layers do not operate independently of one another. According to Throsby, the two values may complement one another or trade off within the same production decisions; according to Bourdieu, conversion is conditional and lossy. Combining these two views, four mechanisms can be anticipated. The mechanisms are hypotheses whose theoretical foundations differ in strength, and they are detailed in Table 4.
M1, funding (E → K, positive). Economic gain finances science and art production. According to Bourdieu (1986), economic capital is the ground that makes it possible to set aside time for acquiring cultural capital. At the level of the micro-producer this means reinvesting part of the earnings in longer-lasting, more original or more open production. The mechanism requires a condition such as the reinvestment share of income; the effect will therefore not be the same for every producer.
M2, short-termism (E → K, negative). If demand rewards only what is clicked, originality and long-term work are suppressed. Hutchinson (2024) argues that visibility metrics steer production towards mainstream appeal; Campbell (1979) notes that the more an indicator is used in decision-making, the more it is distorted. The mechanism predicts that as the share of production with a high economic score but low originality and durability scores rises, cultural value declines in the following period.
M3, accumulation (K → E, positive). Cultural and scientific capital gives rise to new markets, talent and brands. In Throsby's (2001) reading, cultural value feeds consumption, consumption feeds cultural capital, and cultural capital feeds economic development (Hidalgo-Villota, 2025). The non-rivalry of knowledge (Arrow, 1962; Romer, 1990) supports the economic rationale of this mechanism: accumulation lowers the cost and raises the return of subsequent production.
M4, erosion of trust (K → E, negative). As synthetic noise grows, trust and attention become more expensive, which lowers economic value. The basis is a combination of the three sources discussed in Section 2.7: the tendency of low quality to drive out high quality under quality uncertainty (Akerlof, 1970), the fact that the quality of experience goods is learned only through consumption (Nelson, 1970), and the loss of effort's character as an informative signal as the cost of production falls (Spence, 1973). Although the mechanism is theoretically coherent, whether generative AI actually makes quality harder to discern is an empirical question.
| Code | Direction | Mechanism | Sign | Theoretical basis | Indicator to monitor |
|---|---|---|---|---|---|
| M1 | E → K | Funding | Positive | Bourdieu (1986): economic capital buys time | Reinvestment share of income and next-period K |
| M2 | E → K | Short-termism | Negative | Hutchinson (2024); Campbell (1979) | Share of outputs with high E and low K1–K2, and next-period K |
| M3 | K → E | Accumulation | Positive | Throsby (2001); Arrow (1962); Romer (1990) | Next-period E score of high-K production |
| M4 | K → E | Erosion of trust | Negative | Akerlof (1970); Nelson (1970); Spence (1973) | Course of E3 and E4 in cells with a high share of low-K5 outputs |
The Value Creation Index (VCI) is the composition of the two scores of an output; it is not a single number.
Outputs are divided into four quadrants by a threshold value c (initial proposal c = 5). Quadrant position is a first indicator of which interaction mechanism may be at work (Table 5 and Figure 1). The threshold rests not on a theoretical justification but on an initial choice based on the midpoint of the scale, and it will be tested with the values 4, 5 and 6 in the sensitivity analysis.
Earns, but originality and durability are low.
Watch: short-termism (M2)
Earns and leaves accumulation for the future.
Watch: funding and accumulation (M1, M3)
Neither earnings nor accumulation.
Watch: erosion of trust (M4)
Accumulation is high, income is still low.
Watch: funding need (M1)
| Quadrant | E | K | Interpretation | Mechanism to monitor |
|---|---|---|---|---|
| Cumulative value | ≥ c | ≥ c | Earns and leaves accumulation for the future | M1 and M3 |
| Brilliant but transient | ≥ c | < c | High earnings, low originality and durability | M2 |
| Early value | < c | ≥ c | High accumulation, income still low | M1 (funding need) |
| Noise | < c | < c | Neither earnings nor accumulation | M4 |
In moving to the producer level, the share of each quadrant within the producer-period is used. The producer's quadrant profile πq,it is the ratio of the outputs falling in quadrant q to total volume.
Here I(·) is the indicator function. The quadrant shares form a profile that sums to 1 and shows in what balance the four quadrants appear. If, for example, the share of a high-volume producer in the "noise" quadrant is rising, this is read together with the dilution reading of the coefficient to be defined in Chapter 5; a producer's quality profile and its sensitivity to volume are thus monitored within the same framework.
The Value Creation Index gives the quality of an output. The core question of the study, however, is how a change in production volume is reflected in value mass. If volume grows rapidly as the cost of production falls and the average quality of outputs stays constant, value mass rises in proportion to volume. If average quality falls, the mass grows more slowly than volume or does not grow at all; if average quality rises, the mass grows faster than volume. The Value Production Coefficient (VPC) is defined to summarize these three cases in a single parameter. Because the parameter is built as an elasticity, it can be interpreted independently of units and scale and is expressed in a language familiar from the economics literature.
The economic and cultural value masses of producer i in period t are defined as the sum of the normalized output scores. Average value yield is this mass divided by volume (s ∈ {e, k}).
The Value Production Coefficient is the elasticity of value mass with respect to volume and is computed separately for each layer.
The coefficient shows by what percentage value mass changes when volume changes by one percent. Taking logarithms of both sides of V = Q·ȳ, the coefficient can also be written in terms of the elasticity of average yield with respect to volume.
Proposition 1 (Decomposition). At the producer level, the VPC is obtained by adding the elasticity of average yield with respect to volume to one (Equation 7).
Proof. Since ln V = ln Q + ln ȳ, differentiating both sides with respect to ln Q gives the result directly. ∎
Proposition 2 (Marginal and average yield). If the scores of existing outputs do not change and volume rises by ΔQ, the coefficient computed as an arc elasticity equals the ratio of marginal yield to average yield.
Proof. The expression (ΔV/V)/(ΔQ/Q) can be rewritten as (ΔV/ΔQ)/(V/Q); the numerator is marginal yield and the denominator is average yield. ∎ Hence if λ < 1, outputs added later are worth less than the average output.
Proposition 3 (Scale invariance). The coefficient does not change when scores are rescaled multiplicatively, for example when moving from the 0–10 range to the 0–1 range; it does change under additive transformations.
Proof. Multiplying V by a positive constant adds a constant term to ln V, and this term vanishes in the derivative. An additive shift, by contrast, changes the dependence of V on Q. ∎ Normalizing scores by dividing by 10 therefore does not affect the value of the coefficient.
Proposition 4 (Criterion decomposition). Layer mass is the sum of criterion masses; the coefficient is therefore the average of the criterion coefficients weighted by the criterion mass shares.
Proof. The derivative of the sum Vs = Σm Vm with respect to ln Q equals the sum of Vm × (∂ ln Vm / ∂ ln Q); dividing both sides by Vs gives the result. ∎ The proposition makes it possible to see which criterion the dilution comes from: if, in the economic layer, the coefficient of the non-substitutability criterion (E3) is low, one can interpret that the increase in volume is diluting economic value because of a lack of original input.
The interpretation of the coefficient is given in Table 6. A negative coefficient is possible only when the scores of existing outputs fall, that is, through an external effect or erosion, since the score of an added output itself cannot be below zero.
| Range | Name | Meaning |
|---|---|---|
| λ < 0 | Eroding | The increase in volume reduces the existing value mass (external effect; K5-type erosion) |
| 0 ≤ λ < 1 | Diluting | Value mass rises but average yield falls |
| λ = 1 | Proportional | Average yield is constant |
| λ > 1 | Cumulative | Average yield rises (learning or complementarity) |
To show how the coefficient works, Table 7 gives four hypothetical producers. All move from ten outputs to twenty, and the average score of their old outputs starts at 0.5. The figures are for illustration only; they represent no real producer and rest on no empirical finding.
| Producer | Average score of old outputs (before → after) | Average score of the ten new outputs | V1 | λ (arc) | Interpretation |
|---|---|---|---|---|---|
| A | 0.5 → 0.5 | 0.2 | 7 | 0.4 | Diluting |
| B | 0.5 → 0.5 | 0.5 | 10 | 1.0 | Proportional |
| C | 0.5 → 0.5 | 0.8 | 13 | 1.6 | Cumulative |
| D | 0.5 → 0.3 | 0.1 | 4 | −0.2 | Eroding |
For producer A, value mass has risen by forty percent while volume has risen by one hundred percent; the coefficient is 0.4 and, by Proposition 2, equals the ratio of marginal yield (0.2) to average yield (0.5). For producer D, the low score of the new outputs is accompanied by a loss of score on the old outputs; this lowers value mass from 5 to 4 and makes the coefficient negative. This illustrates a case of erosion in which the increase in volume lowers the value of existing outputs, for example because the common pool of attention is eroded or because the loss of trust spills over to older works.
If AI usage intensity is denoted Ait, the effect of AI on volume and its effect on average yield can be estimated separately. Because of the identity ln V = ln Q + ln ȳ, the sum of these two effects gives the total effect on value mass. The AI-mediated coefficient is the ratio of the effect on value mass to the effect on volume.
One advantage of this definition is that, unlike common instrumental-variable estimates, it does not require an exclusion restriction. AI is not assumed to affect value only through volume; the volume channel and the quality channel are estimated separately, and by the identity the total effect is their sum. The ratio in the second equality is read as the ratio of the quality effect to the volume effect. When the volume effect is positive and the ratio lies between −1 and 0, the increase in volume raises value while eroding part of quality; if it is below −1, value mass falls. A causal interpretation of the estimated coefficient depends on AI adoption being regarded as sufficiently exogenous within the estimation strategy (Section 6.7); the definition does not remove this dependence, it only does not add a further exclusion restriction. The coefficient becomes unstable when the denominator of the ratio, that is, the volume effect, approaches zero; in that case the coefficient is not reported and only the estimates of the two channels are reported.
The interaction mechanisms (Table 4) can also be expressed as cross elasticities. The effect of economic value mass on the next period's cultural value mass is denoted θE→K, and the effect of cultural value mass on the next period's economic value mass θK→E.
Since M1 and M2 are mechanisms in the same direction but with opposite signs, a single θE→K reflects the net of these two effects. The sign of the net depends on which mechanism dominates; to separate the mechanisms, interaction terms are therefore used (the reinvestment share of income, the E3 score and the share of low-K1–K2 outputs). A similar decomposition applies to M3 and M4. The cross coefficients and the Value Production Coefficient are estimated independently of one another; reading them together, however, makes it possible to see the volume–value and layer-to-layer relationships of the same producer in a single framework.
The producer-level coefficient implicitly assumes that one producer's increase in volume does not affect the value of other producers. Where common resources such as attention and trust are at stake, however, one producer's increase in volume can also change the value mass of the others. A system coefficient that includes this effect can be defined. Let there be N producers, and let Vi be a function of producer i's own volume Qi and the total volume of the others Q−i. The own elasticity is denoted λii and the cross (external) elasticity λi,−i.
Proposition 5 (System coefficient). If the volumes of all producers change by the same proportion, the elasticity of total value mass with respect to total volume is the sum of own and cross elasticities weighted by producers' mass shares.
Proof. For each producer, write d ln Vi = λii d ln Qi + λi,−i d ln Q−i. When all volumes change by the same proportion, d ln Qi = d ln Q−i = d ln Q, and d ln Vi = (λii + λi,−i) d ln Q is obtained. Since the total change is dV = Σi Vi d ln Vi, dividing by V gives the result. ∎
The proposition shows that a situation that looks proportional at the producer level (λii = 1) can turn into dilution or erosion at the system level: if the cross elasticity is negative, that is, if one producer's increase in volume lowers the value of the others, the system coefficient falls below one. In the limiting case of Section 2.6, where total attention is fixed and value is proportional to attention, each producer's own elasticity is positive and the cross elasticity is negative in exactly the amount that offsets it, so that the system coefficient equals zero. Estimating the system coefficient requires a broader data structure that tracks multiple producers and the ecosystem; the empirical design of this study therefore focuses on the producer level, and the system coefficient is left to future work.
The coefficient rests on a constant-elasticity assumption; since the dependence of value mass on volume may be non-linear, flexible forms (piecewise or spline) will also be tried in estimation. Average yield contains measurement error when estimated from a scored sample. If the error is random and independent of volume it only inflates standard errors; if the share of the scored sample varies systematically with volume, for example if the sampling rate is lower for high-volume producers, bias may arise. The definition and counting of an output also matter: whether multiple versions of the same work are counted as separate outputs directly affects Q and therefore the coefficient; the counting rules are defined explicitly in Chapter 6. Finally, the coefficient is a descriptive parameter; its causal interpretation depends on the conditions provided by the estimation strategy, and in this study a causal claim will be made only when the design conditions are met.
This chapter presents the research design proposed for testing the model. The design has not yet been implemented; the sampling, measurement and estimation steps defined here are the plan for future implementation. Part of the plan is designed to be preregistered in a public registry (for example, the Open Science Framework) before any data are collected.
The research adopts a post-positivist approach that values the falsifiability of testable propositions and the reliability of measurement. However, since the measurement of cultural value inherently involves interpretation, it is not assumed that subjectivity can be removed from measurement; the aim is to tie subjectivity to transparent criteria and to report agreement (Throsby, n.d.). The design is a two-phase, quantitative and observational one, consisting of scale development and a reliability study followed by panel estimation (Table 8).
| Phase | Aim | Sample | Product | Related question and hypothesis |
|---|---|---|---|---|
| Phase 1 | Development of the rubric, rater training, reliability and content validity | 20–30 micro-producers; at most 10 outputs per producer | Revised rubric, reliability coefficients | RQ1, RQ4; H0 |
| Phase 2 | Panel estimation of the VPC and cross coefficients, testing of the interaction mechanisms | Panel of producers to be determined by power analysis; multi-period | Estimates of λ, λAI, θ | RQ2, RQ3; H1–H8 |
The population consists of producers who have no employees or work with at most two people and who use generative AI tools in their production. To ensure variety across fields, sampling is drawn from more than one field of production, such as software and mobile applications, music, books and publishing, and video content; this makes it possible to see whether results depend on the characteristics of a single field. In Phase 1, purposive sampling is used: the aim is to test the applicability of the rubric across different fields, not to represent the population. In Phase 2, probability sampling is preferred in fields where a public list of producers exists; where this is not possible, selection bias is reported and addressed through sensitivity analyses.
The sample size in Phase 2 will be determined by a simulation-based power analysis using variance estimates obtained from Phase 1. The power analysis must take into account the smallest effect size at which the coefficient of interest (for example, its deviation from 1) can be detected, the within-producer correlation and the expected attrition. No figure is put forward at this stage.
Because production volume (Q) directly determines the coefficient, the rules for counting outputs must be defined explicitly. An item counts as an output when it meets all of the following conditions: it has been publicly released, it carries a date stamp, and it is distinguishable in content from earlier items. Small corrections in the nature of patches, and copies of the same work on different platforms, are not counted as separate outputs. Variants produced in bulk whose content differs only by small parameter changes are counted as a single output when the raters cannot tell them apart. Counting is done independently by two raters and disagreements are resolved by an adjudicator. Recomputing the coefficient under an alternative counting rule (for example, counting only items that are distinguishable above a given threshold) will be carried out as a sensitivity analysis.
Production volume is obtained from public release records (version histories, upload dates, release lists). AI usage intensity (A) is measured through producer self-report and tool adoption dates; it is noted that self-report may carry response bias, and cross-verification against records (tool account dates, tool information stated in releases) is planned wherever possible. Economic and cultural scores are produced by two independent raters using the rubric in Appendix A. Control variables are the field of production, the producer's years of experience, the release platform and the pricing model. As external criteria, indicators of income and user continuity, and indicators of long-term use and archiving, are collected; these are used in testing the validity of the scale (Section 6.6).
The scoring protocol consists of five steps. First, the criteria and anchors in the rubric are presented to a panel of experts in cultural economics, scale development and the fields of production, and the necessity of each criterion is assessed for content validity. The content validity ratio can be computed with Lawshe's (1975) approach; under this approach, a sufficient majority of the panel finding a criterion necessary is the condition for retaining it. Second, the raters are trained; training includes discussing the anchors with examples and joint scoring on a calibration set. Third, scoring is conducted blind: raters evaluate without knowing the producer of the output or the level of AI use. It is known that blind scoring is not possible in every field; for example, if the producer's identity is evident from the work, blinding is partially broken, and this is reported. Fourth, each output is scored independently by two raters; disagreements are resolved by an adjudicator or a consensus meeting. Fifth, criteria for which no evidence can be found are marked "not assessable" and not zero, and are handled separately; to monitor rater drift over time, the calibration set is rescored at set intervals.
Because scoring all outputs for every producer is costly, a random subsample of outputs is scored for each producer-period; average yield is estimated from this subsample and volume is counted from public records. The subsample fraction should be held constant across producers and periods; otherwise measurement error becomes correlated with volume (Section 5.8).
Inter-rater reliability will be assessed with Krippendorff's alpha for the ordinal scale at the criterion level and with the intraclass correlation coefficient for total scores (Koo and Li, 2016). Krippendorff (2004) states that results with an alpha of 0.800 or above can be considered reliable and those between 0.667 and 0.800 acceptable for tentative conclusions; these thresholds will be adopted as a starting point. Internal consistency coefficients (such as Cronbach's alpha) will not be treated as the primary reliability indicator for the E and K scales, because the criteria are not reflective indicators of the same latent variable but formative indicators that bring together different aspects of value (Bollen and Lennox, 1991). In indices built from formative indicators the criteria are not expected to correlate highly with one another; on the contrary, excessively high correlation suggests that the criteria do not carry separate information (Diamantopoulos and Winklhofer, 2001). Correlations among criteria are therefore examined as a collinearity check, and criterion pairs showing high dependence are redefined or merged.
Validity will be examined along three dimensions. For content validity, the mapping of the criteria to Throsby's components of cultural value and to Mazzucato's distinction will be endorsed by an expert panel. Construct validity will be examined by testing theoretical expectations, in line with the frameworks of Cronbach and Meehl (1955) and Messick (1995): the K score is expected to show a weak relationship with platform visibility metrics and a stronger relationship with indicators of long-term use and archiving; this expectation rests on Hutchinson's (2024) argument about the limited relation of visibility metrics to cultural value. For criterion validity, the E score will be compared with indicators of real income and user continuity. These expectations are presented not as findings but as propositions to be questioned.
The Value Production Coefficient. The baseline estimate rests on a panel model with producer and period fixed effects.
Here αi is the producer fixed effect, τt the period fixed effect, Xit the control variables and λs the coefficient of interest. Because value mass can take the value zero, the Poisson pseudo-maximum likelihood estimator will be used instead of the linear log form of the model (Santos Silva and Tenreyro, 2006). Standard errors will be clustered at the producer level.
The AI-mediated coefficient. Because producers adopt AI tools at different dates, a difference-in-differences method suited to staggered adoption (Callaway and Sant'Anna, 2021) will be used. The method estimates the average effect on the treated by adoption cohort and time, and avoids problems that can arise in two-way fixed-effects estimates under staggered adoption. The effect on ln Q and on ln ȳ will be estimated separately, the coefficient in Equation 10 will be computed as the ratio of these two effects, and its confidence interval will be obtained by bootstrap. The key assumption in the staggered adoption design is the presence of parallel trends with the groups that are not treated, irrespective of the timing of adoption; this assumption will be examined with event-study plots for the pre-adoption periods.
Cross coefficients. For the interaction mechanisms, a dynamic panel model with a lagged dependent variable will be built.
To remove the bias arising from the relationship between the lagged dependent variable and the producer fixed effect in short panels, the generalized method of moments (Arellano and Bond, 1991) will be used. To separate M1 and M2, interaction terms with the reinvestment share of income and with E3 and the share of low-K1–K2 outputs will be added to the model. For M3 and M4, the reverse model (the effect of the lagged value of K on E) will be built in the same way.
Sensitivity analyses. Analyses will be repeated along three axes: setting the threshold c to 4, 5 and 6; setting criterion weights by the expert panel instead of equal weights; and using methods appropriate to an ordinal scale, such as ordered logistic regression, instead of total scores (Stevens, 1946). Because more than one hypothesis is tested on the same data, a multiple-testing correction (Holm, 1979) will be applied; in addition, the hypotheses with the weakest foundation will be handled as a separate family.
The main threat is selection. Producers who adopt AI may differ at the outset from those who do not, and this difference may affect both volume and quality. The staggered adoption design and producer fixed effects partly reduce this threat, but they do not remove unobserved factors that vary over time. The second threat is measurement error: rater subjectivity and sampling error may pull the coefficient towards zero; for criteria with low inter-rater agreement, results will be interpreted accordingly. The third threat is the metric becoming a target. Campbell (1979) argues that the more a social indicator is used in the decision process, the more it will be distorted; the Value Creation Index will be used only as a research instrument and not as a reward or ranking tool, and scores will not be presented to producers as rewards. The fourth threat is that the AI tools themselves change over time; advances in tool versions may distort the effect measured against the adoption date, and tool version information will be added as a control variable wherever possible. The fifth threat is external effects: if one producer's increase in volume affects the value of other producers, the producer-level coefficient does not reflect the system-level effect (Section 5.7).
The research will be conducted only on public outputs and on information that producers share with their explicit consent, and producer identities will be anonymized in the analysis results. Ethics committee approval will be obtained if there is an institutional requirement. The author is the owner of an independent studio that produces with AI; because of this interest, the author's own outputs will either be excluded from the sample or reported as a separate analysis. It will also be required that raters have no interest relationship with the producers they evaluate.
The matching of hypotheses, parameters of interest and estimation methods is summarized in Table 9; the content of the hypotheses is detailed in Chapter 7.
| Hypothesis | Parameter | Model | Decision criterion |
|---|---|---|---|
| H0 | Krippendorff's alpha (criterion level) | Reliability analysis (Phase 1) | α ≥ 0.667 for each criterion (preferably ≥ 0.800) |
| H1 | λeAI | Staggered-adoption difference-in-differences; ratio | λeAI < 1; confidence interval excludes 1 |
| H2 | λkAI − λeAI | Same model; bootstrap for the difference | Difference negative and significant |
| H3 | ln Q × E3 interaction | PPML, with interaction (13) | Interaction coefficient positive |
| H4 | θE→K (high reinvestment) | Dynamic panel (14); subgroup | Coefficient positive |
| H5 | θE→K × low K1–K2 share | Dynamic panel; interaction | Interaction coefficient negative |
| H6 | θK→E | Dynamic panel (reverse direction) | Coefficient positive |
| H7 | θK→E × low K5 share | Dynamic panel; interaction | Interaction coefficient negative |
| H8 | Quadrant profile and continuity | Survival analysis | Cumulative quadrant has longer continuity than the noise quadrant |
The hypotheses are derived from the model built in the previous chapters and are presented not as empirical findings but as propositions to be tested. There are nine hypotheses: one addresses the precondition of measurement and eight address the substantive questions. Their foundations differ in strength; those with the weakest foundations are marked as exploratory. All are written so as to be falsifiable: for each, the finding that would refute the hypothesis is stated in Table 10.
H0. The rubric provides sufficient reliability between two independent raters. This hypothesis is the precondition for the others: estimates made with a scale of insufficient reliability cannot be interpreted. As the criterion, Krippendorff's alpha is expected to be at least 0.667 for each criterion (preferably 0.800 and above) (Krippendorff, 2004). If a criterion does not meet the threshold, it is redefined and scoring is repeated; until the threshold is met, the results of H1–H8 are not interpreted.
H1 (economic dilution). After AI adoption, economic value dilutes among micro-producers; that is, λeAI < 1. The rationale draws on two sources. First, the E3 logic: when the cost of production approaches zero and entry is free, the return on an output that anyone can produce is eroded by competition. Second, attention scarcity (Simon, 1971): if the buyer's attention is limited, an added output is expected to attract less attention than the average output. The hypothesis is refuted if the confidence interval of the coefficient includes 1 or if the coefficient is found above 1. In that case it will be concluded that AI-assisted volume growth raises economic value mass at least proportionally, and the study's concern about dilution will be questioned in light of these data.
H2 (cultural dilution, exploratory). Dilution in the cultural layer is more pronounced than in the economic layer; that is, λkAI < λeAI. The rationale is the relative scarcity of originality as the cost of recombination falls (Weitzman, 1998) and the analogy drawn from findings on model collapse (Shumailov et al., 2024). This foundation is weak, and the opposite view can also be defended: market competition may be eroding economic value faster than cultural value. The hypothesis is therefore treated as exploratory and will be handled as a separate family in the multiple-testing correction. It is refuted if the difference is found positive or insignificant.
H3 (moderating role of the original input). The volume–value relationship is stronger among producers with a high original-input score (E3); that is, the interaction coefficient of ln Q and E3 is positive. The rationale is that the original input (domain knowledge, proprietary data, community trust) is the element that remains scarce, in contrast to the recombination that AI makes cheap (Weitzman, 1998). The hypothesis is refuted if the interaction coefficient is found to be zero or negative.
H4 (funding, M1). Among producers with a high reinvestment share of income, economic value mass increases next-period cultural value mass; that is, θE→K > 0 in this subgroup. The rationale is Bourdieu's (1986) finding that economic capital buys time for acquiring cultural capital. It is refuted if the coefficient is found to be zero or negative.
H5 (short-termism, M2). As the share of outputs with a high economic score but low originality and durability scores (K1–K2) rises, the effect of economic value mass on next-period cultural value mass weakens; that is, the interaction coefficient is negative. The rationale is that visibility metrics steer production towards mainstream appeal (Hutchinson, 2024) and that metrics are distorted as they are used (Campbell, 1979). It is refuted if the interaction coefficient is found to be zero or positive.
H6 (accumulation, M3). Cultural value mass increases next-period economic value mass; that is, θK→E > 0. The rationale is the Throsby reading that cultural value feeds consumption and economic development (Hidalgo-Villota, 2025), together with the cumulative and non-rival nature of knowledge (Arrow, 1962; Romer, 1990). It is refuted if the coefficient is found to be zero or negative.
H7 (erosion of trust, M4). Among producers with a high share of outputs with low K5 scores, the effect of cultural value mass on next-period economic value mass is lower; that is, the interaction coefficient is negative. The rationale is the theory of quality uncertainty, experience goods and signalling (Akerlof, 1970; Nelson, 1970; Spence, 1973). The foundation of the hypothesis is theoretical; whether generative AI actually makes quality harder to discern is an empirical question, and this hypothesis provides an indirect test of it. It is refuted if the interaction coefficient is found to be zero or positive.
H8 (quadrant and continuity). Producers whose quadrant profile is concentrated in the cumulative value quadrant sustain their activity longer than those concentrated in the noise quadrant. This hypothesis tests whether the index is not merely descriptive but related to an outcome variable. The basis concerns a general framework, the sustainability of production that carries both cultural and economic value; its theoretical foundation is more limited than that of the other hypotheses. It is refuted if no significant relationship is found between the cumulative quadrant and continuity.
| Code | Hypothesis | Foundation and strength | Refutation condition |
|---|---|---|---|
| H0 | The rubric provides sufficient reliability | Krippendorff (2004); measurement requirement | α < 0.667 for any criterion |
| H1 | Economic value dilutes with AI adoption (λeAI < 1) | E3 logic; Simon (1971); moderate | λeAI ≥ 1 |
| H2 | Cultural dilution is more pronounced than economic dilution (exploratory) | Weitzman (1998); analogy with Shumailov et al. (2024); weak | Difference positive or insignificant |
| H3 | The volume–value relationship is stronger when the original input is high | Weitzman (1998); E3; moderate | Interaction coefficient zero or negative |
| H4 | Under high reinvestment, E raises next-period K (M1) | Bourdieu (1986); moderate | θE→K ≤ 0 (in this subgroup) |
| H5 | As the share of high-E, low-K1–K2 outputs rises, the effect of E on K weakens (M2) | Hutchinson (2024); Campbell (1979); moderate | Interaction coefficient ≥ 0 |
| H6 | K raises next-period E (M3) | Throsby (2001); Arrow (1962); Romer (1990); moderate | θK→E ≤ 0 |
| H7 | When the share of low-K5 outputs is high, the effect of K on E weakens (M4) | Akerlof (1970); Nelson (1970); Spence (1973); theoretical, empirical basis weak | Interaction coefficient ≥ 0 |
| H8 | The cumulative quadrant carries longer continuity than the noise quadrant | General framework; limited | No significant relationship, or one in the opposite direction |
The causal interpretation of the hypotheses depends on the conditions of the estimation strategy defined in Section 6.7. Results obtained from observational data will be read as associations when the conditions are not met, and a causal claim will be made only when these conditions are satisfied and only for the relevant hypothesis.
Because the study contains no empirical findings, this chapter discusses what the findings might mean, the counter-arguments that can be raised against the proposed framework, the limitations and future work. The inferences made here are conditional: they will rest on a finding only once the relevant hypotheses have been tested.
The theoretical contribution of the study lies in redefining the object of measurement. Macro frameworks that measure AI output ask how much is produced or how much of it is reflected in the accounts. This study argues that in an environment where production is getting cheaper, the distinguishing question is how much of what is produced carries value and what it leaves for the future. The Value Production Coefficient expresses this question as an elasticity; its link to average and marginal yield (Proposition 2) makes the measure readable in a language familiar from the economics literature.
Two features of the model deserve particular discussion. The first is that the scores are not summed but read in a two-dimensional position (Section 4.5). This choice avoids the assumption that economic and cultural value can stand in for one another; its price is that producers cannot be placed in a single ranking. The second is that the coefficient is defined at the producer level. External effects at the ecosystem level (Proposition 5) are treated theoretically but left outside the empirical design. For common resources such as attention, a situation that looks proportional at the producer level may point to dilution at the system level (Section 5.7); this is one of the most important limits of the design.
Table 11 summarizes the four basic patterns the economic and cultural coefficients can take and the theoretical interpretation of each. The patterns are only an interpretive framework; none is presented as an expectation or a result.
| Pattern | Possible interpretation | Further examination |
|---|---|---|
| λe < 1 and λk < 1 | The increase in volume dilutes value in both layers; suggests a constraint rooted in attention and trust | Criterion decomposition (Proposition 4); system-level effects |
| λe < 1, λk ≥ 1 | Economic return erodes with volume while cultural accumulation grows with volume; clustering in the "early value" quadrant and a funding need | M1 and H4; quadrant profile |
| λe ≥ 1, λk < 1 | Gains from volume remain high while cultural accumulation declines; clustering in the "brilliant but transient" quadrant | M2 and H5; criteria K1–K2 |
| λe ≥ 1 and λk ≥ 1 | The increase in volume raises value at least proportionally; the concern about dilution is questioned in light of these data | Which subgroups sustain this pattern (H3) |
"Value is price; the market decides." A neoclassical view may argue that value cannot be measured by any criterion other than market price. The economic layer of this study already takes market counterpart (E1) as its basis. Moreover, Throsby argues that cultural value cannot be fully expressed in monetary terms and that economics misses this value; Mazzucato likewise argues that price must be separated from value. The two-layer structure rests on these views and leaves them open to debate.
"Cultural value is a subjective and elitist judgment." A Bourdieusian critique may say that criteria of "valuable culture" reflect the taste of privileged groups. This risk is real and cannot be entirely removed. The proposed criteria (originality, durability, accumulation, memory, future cost) look at observable properties instead of setting up a hierarchy of taste or formal quality; multiple and diverse raters and a transparent rubric aim to limit this risk. Adding the criteria of representation and social impact proposed by Hutchinson (2024) could be a development in this direction. Even so, the possibility that the scores produced by the rubric may reflect raters from a particular cultural position is an explicit limitation of this study.
"AI output is not culture," or conversely, "AI is a new cultural form." The rubric does not look at whether an output was produced with AI; raters are kept blind to this information as far as possible. An output produced with AI can therefore receive a high K score if it meets the criteria of originality, durability and accumulation. This choice aims to keep a value judgment about AI use itself outside the measurement; it does not resolve the debate on Benjamin's (1935/2008) view that the technique of reproduction weakens the "aura", it only ensures that measurement does not depend on that debate.
"The metric becomes a target." Campbell (1979) is the general formulation of this risk. Using the index only as a research instrument and not as a reward or ranking tool offers some protection; once the index is published, however, how it is used outside is beyond the author's control. For this reason the study does not recommend that the index be used as a grade or rating for producers.
"The coefficient says nothing the index does not." The index gives the quality of the output but does not summarize in a single parameter how quality changes with volume. Two producers can have the same average quality, yet one maintains its average as volume grows while the other loses it. The coefficient makes this difference visible. The two instruments are therefore complements and not substitutes.
The implications below acquire meaning only if the relevant hypotheses are supported. For platforms, if visibility metrics are found to have a weak relationship with the cultural layer (the construct validity expectation), adding part of the proposed criteria as auxiliary signals to recommendation and ranking systems could be discussed; Hutchinson (2024) proposes qualitative criteria in this direction. For funders and cultural policy, if producers clustered in the "early value" quadrant have low economic returns but high cultural accumulation (H4 and M1), their funding need is theoretically higher; this cannot be used directly as a rationale for a particular funding instrument, but it supplies input to a discussion of targeting. For producers, seeing their own quadrant profiles and coefficients could serve as a self-assessment tool; within the framework of Campbell's warning, however, this use must remain an observation tool and not a target.
The proposed framework is sensitive to context in terms of local language and cultural memory, particularly through the K4 criterion. For micro-producers producing Turkish-language content, this criterion directly concerns the measure of enriching the language, memory and identity of a community. However, data on this context have not been compiled in this study; given that cultural value changes over time and from community to community (Throsby, n.d.), the rubric needs to be tested separately in the local context and adapted if necessary. The introduction of the local context should not be regarded as empirical evidence of this study; it only indicates under what conditions future application should be carried out.
The first and fundamental limitation of the study is that it contains no empirical findings; no hypothesis has been tested, and the reliability and validity of the scale have not been demonstrated. The second limitation concerns the measurement instrument itself. Cultural value varies over time and from community to community (Throsby, n.d.); the five-criterion rubric captures only part of this variability, and equal-weighted totals and the summing of ordinal scores as if they were an interval scale are assumptions. The third limitation is the scope of the literature review: the review is not systematic, Throsby's 1999 article and Mazzucato's book could not be accessed directly, the literature on intergenerational justice and cultural sustainability has not been reviewed, and some of the method sources in the bibliography were not reopened and compared during this work. The fourth limitation is that the basis of the model collapse analogy consists only of findings demonstrated for machine-learning models (Shumailov et al., 2024; Seddik et al., 2024). The fifth limitation is that the debate on the macro impact of AI continues; there is no consensus on the speed and scale of the effect (Nolan, 2026). The sixth limitation is the possible conflict of interest arising from the author's relationship with an independent production studio (Section 6.9). The seventh limitation is that the study is single-authored and was prepared outside an institutional context; this creates an additional need for peer review and critical reading.
The direct continuation is carrying out the first phase defined in Chapter 6 and revising the rubric according to the reliability results. This should be followed by carrying out the panel estimation and reporting the Value Production Coefficient and the cross coefficients. In the medium term, estimating the system coefficient (Proposition 5) to capture ecosystem-level effects, and treating it as a micro-level complement to the intensity indicator proposed by Makridis and Brynjolfsson (2026), could be considered. In addition, comparative studies questioning whether the rubric is valid in different cultural contexts, and comparing the scoring with joint evaluation methods based on expert panels (Abduljalil et al., 2026), would be useful. Furthermore, adding dimensions such as representation and social impact proposed by Hutchinson (2024) to the cultural layer, and reviewing the K criteria by drawing on the literature on intergenerational justice (for example, debates on cultural sustainability), would strengthen the theoretical foundation of the model.
This study has conceptualized the value of AI-assisted micro-production as a framework in which the economic layer and the science, art and culture layer are scored separately and the interaction between them is monitored through four mechanisms. Outputs are read with two scores based on five criteria each, not summed, in a two-dimensional position and as one of four quadrants; producers are described by the vector of their quadrant shares. The second contribution of the study, the Value Production Coefficient, defines the elasticity of value mass with respect to production volume; it has been shown to equal the ratio of marginal to average yield, to be scale invariant, to be decomposable into criteria and to be extendable to the ecosystem level. The AI-mediated variant rests on separate estimation of the volume and quality effects and requires no exclusion restriction.
The claim of the study has been deliberately kept limited. The distinction between economic and cultural value belongs to Throsby; what is done here is to make this distinction scorable at the production level, to translate the interaction between the layers into testable mechanisms, and to express the volume–value relationship in a single parameter. What is presented is not a finding but a design: a measurement proposal tied to a two-phase research design and nine falsifiable hypotheses. The value of the proposal depends on whether the scale can be applied reliably to real data and on whether the coefficient varies meaningfully across producers, fields and periods; both are empirical questions and lie outside the scope of this study.
At the theoretical level, the study proposes that the debate on value in the age of AI can be moved from the question "how much was produced" to the question "how much of what is produced carries value and what does it leave for the future", and that this transition can be given a measurable form. Testing the proposal depends first of all on demonstrating the inter-rater reliability of the rubric. If this condition is met, the Value Creation Index and the Value Production Coefficient could offer a common language for researchers, platforms and producers who wish to monitor the relationship between economic return and cultural accumulation in an environment where production is getting cheaper. If it is not met, the study will nonetheless have set out the measurement problem clearly and shown at what point the solution stalls.
Abduljalil, M., Komosany, N., & Bahkali, I. (2026). Cultural capital between Bourdieu and Throsby: Toward an interpretative framework of collective expert appraisal. Interdisciplinary Description of Complex Systems, 24(4), 472–490. https://doi.org/10.7906/indecs.24.4.8
Akerlof, G. A. (1970). The market for "lemons": Quality uncertainty and the market mechanism. The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431
Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies, 58(2), 277–297. https://doi.org/10.2307/2297968
Arrow, K. J. (1962). Economic welfare and the allocation of resources for invention. In National Bureau of Economic Research (Ed.), The rate and direction of inventive activity: Economic and social factors (pp. 609–626). Princeton University Press.
Benjamin, W. (2008). The work of art in the age of its technological reproducibility, and other writings on media. Harvard University Press. (Original work published 1935)
Bollen, K., & Lennox, R. (1991). Conventional wisdom on measurement: A structural equation perspective. Psychological Bulletin, 110(2), 305–314. https://doi.org/10.1037/0033-2909.110.2.305
Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241–258). Greenwood. https://www.marxists.org/reference/subject/philosophy/works/fr/bourdieu-forms-capital.htm
Bourdieu, P. (1993). The field of cultural production: Essays on art and literature (R. Johnson, Ed.). Columbia University Press.
Brynjolfsson, E., Collis, A., Diewert, W. E., Eggers, F., & Fox, K. J. (2019). GDP-B: Accounting for the value of new and free goods in the digital economy (NBER Working Paper No. 25695). National Bureau of Economic Research.
Brynjolfsson, E., Hu, Y., & Smith, M. D. (2003). Consumer surplus in the digital economy: Estimating the value of increased product variety at online booksellers. Management Science, 49(11), 1580–1596.
Callaway, B., & Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200–230. https://doi.org/10.1016/j.jeconom.2020.12.001
Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning, 2(1), 67–90.
Caves, R. E. (2000). Creative industries: Contracts between art and commerce. Harvard University Press.
Corrado, C. (2025, September 29). Review of "Making AI Count: The Next Measurement Frontier" by D. Coyle and J. L. Poquiz. National Bureau of Economic Research. https://www.nber.org/system/files/chapters/c15308/revisions/c15308.rev0.pdf
Coyle, D. (2014). GDP: A brief but affectionate history. Princeton University Press.
Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
Deriya, S., & Huynh, C.-W. (2026). Counting what doesn't count: What AI reveals about the limits of algorithmic measurement (ESCP Impact Paper No. 2026-36-ENG). ESCP Business School. https://escp.eu/sites/default/files/academ_attachments/contrib_410413_attachment_409192.pdf
Diamantopoulos, A., & Winklhofer, H. M. (2001). Index construction with formative indicators: An alternative to scale development. Journal of Marketing Research, 38(2), 269–277.
Hardin, G. (1968). The tragedy of the commons. Science, 162(3859), 1243–1248.
Hidalgo-Villota, M. E. (2025). The cultural value in the thought of David Throsby. Pensamiento palabra y obra, (34), Article e22739. https://revistas.upn.edu.co/index.php/revistafba/article/download/22739/15954
Highfill, T., Wasshausen, D., & Prunchak, G. (2025). Concepts and challenges of measuring production of artificial intelligence in the U.S. economy (BEA Working Paper No. WP2025-1). U.S. Bureau of Economic Analysis. https://bea.gov/system/files/papers/BEA-WP2025-1.pdf
Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6(2), 65–70.
Hutchinson, J. (2024). What is the value of cultural analytics? Discerning value in digital environments. International Journal of Communication, 18. https://ijoc.org/index.php/ijoc/article/view/21777
Koo, T. K., & Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of Chiropractic Medicine, 15(2), 155–163. https://doi.org/10.1016/j.jcm.2016.02.012
Krippendorff, K. (2004). Content analysis: An introduction to its methodology (2nd ed.). Sage.
Kuznets, S. (1934). National income, 1929–1932 (73rd Congress, 2nd session, Senate Document No. 124). U.S. Government Printing Office.
Lawshe, C. H. (1975). A quantitative approach to content validity. Personnel Psychology, 28(4), 563–575.
Makridis, C. A., & Brynjolfsson, E. (2026, January 15). Counting AI: A blueprint to integrate AI investment and use data into US national statistics. Brookings Institution. https://www.brookings.edu/articles/counting-ai
Mazzucato, M. (2018). The value of everything: Making and taking in the global economy. PublicAffairs.
Messick, S. (1995). Validity of psychological assessment: Validation of inferences from persons' responses and performances as scientific inquiry into score meaning. American Psychologist, 50(9), 741–749. https://doi.org/10.1037/0003-066X.50.9.741
Nardo, M., Saisana, M., Saltelli, A., Tarantola, S., Hoffmann, A., & Giovannini, E. (2008). Handbook on constructing composite indicators: Methodology and user guide. OECD Publishing.
Nelson, P. (1970). Information and consumer behavior. Journal of Political Economy, 78(2), 311–329.
Nolan, B. (2026, June 2). AI may already be adding hundreds of billions to the economy—without showing up in the data. Fortune. https://fortune.com/2026/06/02/ai-may-already-be-adding-hundreds-of-billions-to-the-economywithout-showing-up-in-the-data/
Rezazadeh, F., & Bonehgazy, P. (2025). Digital co-founders: Transforming imagination into viable solo business via agentic AI (arXiv:2511.09533). arXiv. https://arxiv.org/pdf/2511.09533v1
Romer, P. M. (1990). Endogenous technological change. Journal of Political Economy, 98(5, Pt. 2), S71–S102.
Santos Silva, J. M. C., & Tenreyro, S. (2006). The log of gravity. The Review of Economics and Statistics, 88(4), 641–658. https://doi.org/10.1162/rest.88.4.641
Seddik, M. E. A., Chen, S.-W., Hayou, S., Youssef, P., & Debbah, M. (2024). How bad is training on synthetic data? A statistical analysis of language model collapse (arXiv:2404.05090). arXiv. https://arxiv.org/abs/2404.05090
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y
Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger (Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press.
Solow, R. M. (1987, July 12). We'd better watch out. New York Times Book Review, 36.
Spence, M. (1973). Job market signaling. The Quarterly Journal of Economics, 87(3), 355–374.
Spittler, M., & Patel, D. (2026, May 29). AI dark output: The visible cost of invisible output. SemiAnalysis. https://newsletter.semianalysis.com/p/ai-dark-output-the-visible-cost-of
Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103(2684), 677–680.
Stiglitz, J. E., Sen, A., & Fitoussi, J.-P. (2009). Report by the Commission on the Measurement of Economic Performance and Social Progress.
Throsby, D. (n.d.). Heritage economics: Coming to terms with value and valuation (Chapter 14). Getty Publications. https://www.getty.edu/publications/heritagemanagement/part-two/14/
Throsby, D. (1999). Cultural capital. Journal of Cultural Economics, 23(1–2), 3–12. https://doi.org/10.1023/A:1007543313370
Throsby, D. (2001). Economics and culture. Cambridge University Press.
Throsby, D. (2003). Cultural capital. In R. Towse (Ed.), A handbook of cultural economics. Edward Elgar. https://researchers.mq.edu.au/en/publications/cultural-capital-4/
Weitzman, M. L. (1998). Recombinant growth. The Quarterly Journal of Economics, 113(2), 331–360.
Each criterion is scored 0, 1 or 2. The intermediate level (1 point) is used for cases in which only part of the definition at either end is met. For each criterion the rater briefly records the evidence relied on. Where no evidence can be found, the score is marked "not assessable" and not 0, and these observations are handled separately in the analysis.
| Criterion | 0 points | 1 point | 2 points | Observable evidence |
|---|---|---|---|---|
| E1 Real demand | No evidence of payment or regular use | Limited evidence: a one-time purchase, short-lived use or a very narrow audience | Repeat payment or a regular user base | Sales and subscription records, number of active users, repeat use |
| E2 Creating or extracting | The output redistributes existing value or relies on intermediation, arbitrage and attention-seeking | Mixed: partly new benefit, partly redistribution | Measurable new benefit or cost reduction | User's gain (time, cost, new capability); comparison with customary alternatives |
| E3 Non-substitutability | An output that someone else could produce in the same form quickly with general-purpose tools | Partly original input | Original input (domain knowledge, proprietary data, community trust) is decisive | Documentation of the original input on which the output rests; existence of similar outputs |
| E4 Durability | Income or use is one-time | Continues for a limited period | Accumulates and gains continuity, turning into an asset | Income and user curve over time |
| E5 Sharing of value | External costs (copyright or privacy violation, misleading presentation) exceed the value created | External costs exist but are limited or compensated | External costs negligible; the share going to users and society is clear | Licensing status, privacy practices, complaint and objection records |
| Criterion | 0 points | 1 point | 2 points | Observable evidence |
|---|---|---|---|---|
| K1 Originality | Recombination of existing outputs; indistinguishable from similar ones | Partly original; a limited contribution to known forms | New knowledge, form or perspective | Comparison with similar outputs, expert assessment |
| K2 Durability and transmissibility | Short-lived; unintelligible when severed from its context | Partial: accessible but undocumented | Archivable, documented, intelligible independent of context | Openness of format, archiving, documentation |
| K3 Accumulation effect | Closed, ends within itself | Limited reuse (restricted licence or low accessibility) | Open in a form that others can build on | Licence, citations and number of derivative works |
| K4 Cultural memory | No contribution to a social or cultural context | Indirect contribution | A clear contribution that documents or enriches the language, memory or identity of a community | Community feedback, expert assessment, records of relevant institutions |
| K5 Future cost | Increases synthetic noise or erodes original producers (bulk production of similar works, misleading labelling) | Neutral | No erosion; cites sources, labels, or protects the original producer | Citing of sources, explicit disclosure of the mode of production, comparison with original producers |
The two examples below are constructed to show how the rubric is operated; they are not based on any real output or producer.
| Output (hypothetical) | E1–E5 | K1–K5 | E, K | Quadrant and rationale |
|---|---|---|---|---|
| A mobile guide documenting the techniques of a vanishing regional craft on the basis of recordings of master craftspeople | 1, 2, 2, 1, 2 | 2, 2, 1, 2, 2 | 8; 9 | Cumulative value. Evidence of income is limited (E1 = 1), but the content rests on an original input (E3 = 2) and makes a clear contribution to community memory (K4 = 2); the licence is partly closed (K3 = 1). |
| A series of information cards produced in bulk with general-purpose tools and indistinguishable from similar ones | 1, 0, 0, 0, 1 | 0, 1, 0, 0, 0 | 2; 1 | Noise. Evidence of demand is limited, there is no original input (E3 = 0), it recombines existing content (K1 = 0) and it cites no sources (K5 = 0). |
| Symbol | Meaning |
|---|---|
| i, t, j | Indices for producer, period and output |
| Jit, Qit | The set of outputs of producer i in period t and the number of elements in this set (production volume) |
| Ej, Kj | Economic and cultural score of output j (0–10) |
| Ejm, Kjm | Score of output j on the m-th criterion (0, 1 or 2) |
| ej, kj | Normalized scores (Ej/10 and Kj/10) |
| Vsit | Value mass in layer s (sum of normalized scores); s ∈ {e, k} |
| ȳsit | Average value yield (Vsit / Qit) |
| λs | Value Production Coefficient (elasticity of value mass with respect to volume) |
| λsAI | AI-mediated value production coefficient |
| Ait | AI usage intensity |
| θE→K, θK→E | Cross elasticities (next-period effect) |
| λii, λi,−i | Elasticity of producer i with respect to its own volume and to the total volume of other producers |
| Λ | System-level coefficient |
| πq,it | Share of outputs in quadrant q (quadrant profile) |
| c | Threshold for quadrant classification (initial proposal 5) |
| wm | Mass share of the m-th criterion (Proposition 4) |
| αi, τt | Producer and period fixed effects |
| Xit, Zit | Control variables |
The form below is completed for each output and each rater in Phase 1. The evidence note briefly records the observation on which the score rests; if no evidence can be found, the box is marked "NA" (not assessable).
| Field | Information to be entered |
|---|---|
| Output ID | Coded identifier (the producer's name is not entered) |
| Rater code and date | — |
| E1, E2, E3, E4, E5 | 0 / 1 / 2 / NA for each; evidence note |
| K1, K2, K3, K4, K5 | 0 / 1 / 2 / NA for each; evidence note |
| Total E, total K | Computed automatically (flagged separately if any criterion is NA) |
| Output counting note | Rationale for whether the output is counted separately (Section 6.3) |
| Blinding status | Was the producer's identity or AI use evident during assessment? (Y/N) |
Digital thesis
Generative AI is lowering the cost of production, and the real question is changing: how much of what is produced carries value, and what does it leave for the future? This thesis proposes an index that scores value separately in the economic layer and in the science, art and culture layer, and a coefficient that measures the relationship between production volume and value.
value measurement · cultural capital · generative AI · micro-production · value creation index · value production coefficient
Emre Arıcı / ARC Studios - Tam Bağımsız Fikir Stüdyosu
For information, evaluation and comments:
info@arcstudios.com.tr
November 2026 · Version 1.0 · Not peer reviewed