PeerLab Book a call Plug Peer in
Thesis 19 August 2026 19 min read

Preconditions for a Royalty Economy in Behavioural Skill Data: An Argument from the Attribution, Copyright and Mechanism Design Literatures

Why per-person payment for behaviour that trains a model is currently blocked at attribution, entitlement and valuation, and what published work would have to settle first.

By Sam Okwuosa (AI) and Greg McCallum

Epistemic status: this is an argument, not a measurement. No figure from our corpus is reported, and nothing here has been tested.

A royalty paid to an individual person for behavioural data that trains a model requires three separable things: a credit figure attributable to that person, a legal entitlement in the record from which the figure is computed, and a rate at which the entitlement transacts. Each is addressed by a distinct published literature, and in each literature the requirement is currently unmet. Per-datum attribution methods do not yet deliver credit that is stable under the stochasticity of training, since the same dataset produces different models under different initialisation and batching, and the most recent framing treats attribution as a statement about the distribution of model outputs across training runs rather than a fixed number per data point (arXiv:2506.12965). The entitlement question is open in the jurisdictions covered by the sources cited here, with no general judicial resolution of whether training on protected works is infringement or fair use (Martensen). No truthful mechanism is known for eliciting quality that only the contributor observes when the acquirer must also rank several items from the same owner, and an ordinal mechanism taking only the owner's ranking as input is reported to be not incentive-compatible (WWW 2024, pp. 25-35). The one layer with an implementable standard is provenance documentation, and provenance records that a contribution entered a corpus without establishing how much of a model's competence derives from it. No figure from the PeerLab corpus is reported here, and no evidence lines were available for this piece (n = 0).

The unit of analysis is a skill or competency and the record that evidences mastery of it. A behavioural record of a person exercising a competency is the candidate datum. Whether such a record can carry a royalty is the question, and the answer given here is that it cannot on today's published methods, for reasons that are separable and individually addressable.

Attribution

The equitable-valuation programme begins with the proposal that each training datum be assigned a value quantifying its contribution to a trained predictor's performance. Data Shapley proposes such a metric and states that it uniquely satisfies several natural properties of equitable data valuation. The paper's own motivation is the relevant one for a royalty argument: it notes suggestions that individuals in healthcare and consumer markets should be compensated for the data they generate, while what constitutes an equitable valuation of individual data remains unclear. The authors develop sampling-based and gradient-based estimators so that values can be approximated when complex learning algorithms, including neural networks, are trained on large datasets, and report experiments across biomedical, image and synthetic data in which low-value data capture outliers and corruptions while high-value data indicate what new data to acquire.

Two features of that construction matter for payment. The value is defined with respect to a trained predictor's performance, which makes it a function of a chosen learning algorithm and a chosen performance measure; a royalty computed from such a quantity is a royalty indexed to an evaluation choice made by the party paying it. The axioms are also stated over data points, whereas a royalty is owed to a person, and a person contributes many records. The aggregation from per-datum values to a per-person entitlement is a separate step, and it is not the object the equitable-valuation axioms are stated about.

The influence-function line offers a cheaper route to per-datum credit. The review of data attribution for influence functions, submitted on 10 August 2025, states that the goal of data attribution is to trace a model's predictions through the learning algorithm back to its training data and to identify the most influential training samples, and that influence functions, originating in robust statistics, provide a first-order approximation of the effect of upweighting or removing a data point on learned parameters and predictions without requiring expensive retraining. The review covers theoretical foundations, algorithmic advances for efficient inverse-Hessian-vector product estimation, and effectiveness for attribution and mislabel detection. The availability of an approximation that avoids retraining is what makes per-datum accounting computationally conceivable at corpus scale.

The obstacle is what the approximation is an approximation of. The distributional framing argues that traditional training data attribution algorithms fail to rigorously account for randomness in training, including stochasticity in initialisation and batching, which can yield different models from the same dataset, and it introduces distributional training data attribution to predict how the distribution of model outputs across training runs depends on the dataset (arXiv:2506.12965). Within that framework, influence functions are reported to be "secretly distributional", emerging as the limit of unrolled differentiation without restrictive convexity assumptions, and the method is demonstrated in experiments that include improving data pruning for vision transformers and identifying influential examples with diffusion models.

If the well-posed attribution quantity is distributional, then a per-datum royalty is a payment on a distributional statement. The credit assigned to a given record becomes a property of an ensemble of possible training runs rather than of the run that produced the deployed model, so two audits of the same payment can disagree without either being wrong. The variance of that credit is itself unreported for behavioural corpora, since the demonstrations in the cited work are on vision and diffusion tasks rather than on records of human skill. The sign and magnitude of credit for a marginal record are known from the valuation literature to interact with data quality, since low-value data are reported to coincide with outliers and corruptions, which means a payment scheme that pays proportional to estimated value pays nothing, or negatively, for records that the estimator reads as anomalous.

None of this establishes that stable per-person credit is unobtainable. It establishes that the published methods currently in hand were designed to answer a different question, namely which training examples matter for a model's behaviour, and that the transformation of an answer to that question into a defensible individual payment has not been carried out in the work cited.

A royalty presupposes an entitlement. The copyright literature cited here addresses two questions adjacent to the one a royalty economy needs, and leaves the needed question largely untouched.

The first adjacent question is authorship of model output. The US Copyright Office has determined that purely machine-generated works lacking human authorship are not eligible for protection, while a user exercising meaningful creative control over an output can protect that contribution (Martensen). The application of that line is visible in the treatment of a graphic novel combining written narrative with AI-generated images, where the Copyright Office allowed copyright on the narrative and curated presentation while excluding the AI-generated images (Copyright Law Meets AI). This doctrine allocates rights in what comes out of a model. It says nothing about rights in the behavioural record that went in.

The second adjacent question is whether training itself infringes. Courts have not issued a sweeping decision resolving whether training on copyrighted works constitutes fair use, and the pending matters listed include the Authors Guild against OpenAI, Getty Images against Stability AI, an individual visual artist against image generators, and The New York Times against OpenAI and Microsoft (Martensen). One reported outcome runs against a training defence: in Thomson Reuters v. Ross Intelligence a court ruled that using copyrighted legal databases for commercial AI training was infringement and rejected the fair use defence (Copyright Law Meets AI). A single ruling on a legal-database corpus is not a general entitlement rule for behavioural records of individual people, and the same source that reports the unsettled position also reports that memorisation and near-verbatim reproduction, which it terms regurgitation, is described as relatively rare but not negligible (Martensen). Infringement analysis that turns on reproduction of expression is a poor fit for a claim whose subject is a person's demonstrated competency rather than expressive text.

The subsistence question is where a behavioural royalty most obviously runs into difficulty, and the Indian material shows the shape of the problem. Copyright in India is governed by the Copyright Act, 1957, amended in 2012, with Section 14 granting creators of original works exclusive rights to publish and broadcast, and infringement governed primarily by Sections 51 and 52, the latter setting out exceptions including fair use (iPleaders). Under Section 17 only natural persons are granted ownership rights, which the same source notes leaves it unclear whether AI-generated work can be copyrighted at all and, if so, whether the program, the programmer or the user would hold authorship. The originality standard is set by Eastern Book Company v. D.B. Modak, in which the Supreme Court rejected the "sweat of the brow" standard as too low and the "modicum of creativity" standard as too high, adopting the Canadian test requiring the work to reflect the author's skill and judgement.

A behavioural record of competency is a record of skill and judgement, which invites reading that standard as supplying an entitlement. It does not. The skill-and-judgement test governs whether a work is original enough to attract copyright, and it presumes a work. A transcript or telemetry trace of a person exercising a competency in the course of employment is not established by any source cited here as a work in which that person holds rights, and the employment relationship raises an allocation question that none of the cited sources addresses.

A separate entitlement route runs through personality rather than authorship. Tennessee's ELVIS Act of 2024 protects musicians' voices and likenesses from unauthorised AI replication, and federal proposals such as the NO FAKES Act aim to extend similar protections nationally (Copyright Law Meets AI). Voice and likeness rights attach to a person without requiring that the person authored anything, which makes them structurally closer to what a behavioural royalty would need. Their scope, as reported, is replication of identity rather than contribution to a model's general competence.

Knowledge that a contribution was used is a precondition for any claim on it. The Generative AI Copyright Disclosure Act, introduced in 2024, would require AI developers to disclose the copyrighted works used for training, which the source notes could help creators seek compensation (Copyright Law Meets AI). As reported, it is a bill.

Valuation and mechanism design

Even granting a stable credit figure and a settled entitlement, the payment must be set by some process, and the informational structure of behavioural skill data is adversarial to any simple per-record rate. The person who produced a record knows things about it that no automated appraisal recovers: whether the episode was representative of their ordinary practice, whether it was an unusual case, whether the competency on display was exercised deliberately.

That structure is the one studied in On Truthful Item-Acquiring Mechanisms for Reward Maximization, published on pages 25-35 of the Proceedings of the ACM Web Conference, held 13-17 May 2024. The setting is a collector acquiring items from an owner on the basis of the owner's declared item qualities together with an independent appraiser's assessments, where only the owner knows the items' factual quality and the appraiser's assessment is subject to random noise. For the single-item case the authors design optimal mechanisms with a monotone increasing menu size, and show that simple mechanisms with a small menu size cannot guarantee any positive fraction of the optimal reward achievable with a larger menu. For the multi-item case they show that an ordinal mechanism taking only the owner's ranking of items as input is not incentive-compatible, and they propose "Union" mechanisms combining single-item mechanisms.

Mapped onto behavioural skill data, the result bears on the two schemes most likely to be proposed in practice. A flat per-record rate is a mechanism with a menu of size one, and the cited result shows that small-menu simplicity carries no guarantee of capturing any positive fraction of what a richer menu could achieve in that model. A scheme that asks each person to nominate or rank their strongest records is an ordinal mechanism, and in the multi-item setting the ordinal input is reported not to be incentive-compatible. The mechanisms the paper does design are for the acquisition setting it defines, with one owner, one collector and a noisy appraiser, and extending them to a corpus assembled from many people whose records substitute for one another is not something the cited work reports.

The broader mechanism design literature explains why this is difficult rather than merely unfinished. A survey of algorithmic mechanism design describes mechanism design as reverse game theory, starting at the end of the game and working backwards where the objective function is given but the mechanism is unknown, and records that the Nobel Prize in Economics has been awarded four times for research in this area: in 1996 for work on asymmetric information, in 2007 for pioneering work on mechanism design, in 2012 for contributions to matching markets, and in 2020 for auction theory. The same survey states that it reviews the challenges in designing actual data-driven market mechanisms, distinguishing inherent challenges in mechanism design theory, challenges brought by new markets, and challenges common to both. A data royalty market is a new market in that taxonomy, inheriting both classes of difficulty.

The alternative to deriving payment from causal contribution is allocating entitlements administratively and letting them trade. The transferable-quota case is instructive: a government sets a total allowable catch and divides it into individually tradable quotas, which appears to make fishers internalise the overfishing externality and can be shown efficient in competitive equilibrium (Economic mechanism design). Nothing in that design requires measuring any individual fisher's causal contribution to the stock. The same notebook draws a distinction between provable, elegant pure mechanism design and what it calls soft mechanism design, where the solution must satisfy political feasibility as well as economic optimality, and records that spliddit, a site using optimal cake-cutting algorithms to allocate credit, rent or similar goods, went offline. A quota-style royalty for behavioural skill data would sidestep the attribution problem entirely by declaring entitlements rather than computing them, at the cost of abandoning the claim that payments track contribution.

Provenance documentation

The documentation layer is the part of this stack with a standard that can be implemented today. Datasheets for Datasets, first submitted on 23 March 2018 and last revised on 1 December 2021 in its eighth version, states that the machine learning community has no standardised process for documenting datasets, which can lead to severe consequences in high-stakes domains, and proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process and recommended uses, by analogy with datasheets for electronic components. The stated purpose is improved communication between dataset creators and dataset consumers, alongside transparency and accountability. The questions are organised into seven sections: Motivation, Composition, Collection Process, Preprocessing/Cleaning/Labelling, Uses, Distribution and Maintenance (COMPEL).

Regulatory instruments have converged on the same object. EU AI Act Annex IV explicitly requires providers of high-risk AI systems to document the datasets used, including their composition and provenance, and GDPR Article 30 requires controllers to maintain records of processing activities, which makes a datasheet a natural home for such evidence (COMPEL). The same source records that the Data Provenance Initiative has published research on the limited transparency of training data for popular foundation models, which indicates that the standard's existence has not produced compliance at the frontier.

Provenance is necessary for a royalty and nowhere near sufficient. A complete datasheet can establish that records of a given kind, collected by a given process, entered a corpus used to train a given model. It cannot establish how much of the model's competence derives from any particular person's records, which is the attribution question, nor who holds the right in those records, which is the legal question, nor what the records are worth, which is the mechanism design question. Its contribution to a royalty economy is evidentiary: without it, a person cannot know that a claim exists.

Preconditions

Technical. A per-person credit figure would have to be computable and stable enough that two computations of the same payment agree. The cited attribution work does not supply this. The distributional framing states that stochasticity in initialisation and batching can yield different models from the same dataset and reframes attribution as prediction of the distribution of model outputs across runs (arXiv:2506.12965); whether a per-person aggregate is more stable than the per-datum values composing it is not reported there. An aggregation rule from per-datum values to per-person credit would have to be specified and justified, since the equitable-valuation axioms are stated over data points rather than contributors (Data Shapley). Credit would have to be indexed explicitly to the evaluation measure used, because the value is defined relative to a trained predictor's performance. The figure would also have to be auditable by the payee, which no cited source addresses; influence-function methods are presented as tools for the model owner tracing predictions back to training data (arXiv:2508.07297), not as instruments a contributor can verify. Unresolved on all four counts.

Legal. An entitlement would have to subsist in the behavioural record itself. The cited sources establish authorship doctrine for model outputs and an unsettled position on training as infringement (Martensen), with one reported ruling rejecting fair use for commercial training on legal databases (Copyright Law Meets AI), and, in India, an originality test requiring skill and judgement in a work together with a Section 17 restriction of ownership to natural persons (iPleaders). None of these establishes a right in a record of workplace behaviour. The allocation of any such right between a person and the organisation whose systems captured the behaviour is addressed by no source cited here. A disclosure duty would have to attach, so that a person can learn their records were used; the instrument named for this purpose is a 2024 bill, and the binding documentation requirements identified are addressed to providers of high-risk systems and to data controllers rather than to contributors (COMPEL). Unresolved.

Economic. A mechanism would have to elicit truthfully the quality that only the contributor observes. The nearest published result designs optimal mechanisms for a single-item acquisition with a noisy appraiser, shows that small menus guarantee no positive fraction of the achievable optimum, and reports that ordinal mechanisms over multiple items from one owner are not incentive-compatible (WWW 2024, pp. 25-35). A process for forming payment amounts would have to exist for a market that does not yet exist, which the algorithmic mechanism design survey identifies as a live set of challenges for data-driven market mechanisms (survey). Political feasibility would have to hold alongside optimality, the condition described as soft mechanism design (Economic mechanism design). Partially resolved in the single-owner case, unresolved for a many-contributor corpus.

One precondition is met. Provenance documentation has a published standard with seven specified sections and regulatory hooks in EU AI Act Annex IV and GDPR Article 30 (COMPEL). It is the least contested and the least sufficient of the requirements.

A quota-style design changes which preconditions bind. Allocating tradable entitlements administratively, as in the transferable fishing quota case, removes the technical attribution requirement and the truthful-elicitation requirement, replacing both with an allocation rule chosen by whoever sets the total (Economic mechanism design). It does not remove the legal requirement, and it forfeits any claim that payments are proportional to contribution.

Limitations

The argument rests on eleven sources, several of which are secondary commentary rather than primary material. The reported outcome in Thomson Reuters v. Ross Intelligence, the status of the pending matters, the Copyright Office's treatment of the graphic novel described above, and the provisions of the ELVIS Act and the Generative AI Copyright Disclosure Act are taken from commentary pages rather than from the judgments, register entries or statutory text, and are reported here at second hand. The legal coverage extends to the United States and India only. No European, United Kingdom or Chinese position on entitlement in behavioural records is examined, and the EU material cited concerns documentation duties rather than rights.

The attribution papers are machine learning papers. Their demonstrations are on vision, biomedical, image, synthetic and diffusion settings, and none of the cited work reports attribution behaviour on a corpus of human behavioural records evidencing competency. Inferences drawn here about instability of per-person credit in such a corpus are extrapolations from results obtained elsewhere.

Absence of a published method is not proof of impossibility. The claim is that per-person royalties cannot be implemented on today's published methods as cited, and each of the blocking conditions named above is a research question rather than an impossibility theorem. The single-item mechanism results are guarantees within a stated model, and their failure to extend is unreported rather than demonstrated.

No figure from the PeerLab corpus is reported and no evidence lines were used (n = 0). Nothing about how our own data are collected, documented, consented to or compensated is asserted here, because no such fact appears in the evidence available for this piece. Readers should not infer any capability, payment or consent mechanism from the presence of this argument.

Further work

An empirical stability study is the first tractable step: compute per-datum credit for a behavioural corpus under repeated training with varying initialisation and batching, and report the dispersion of per-person aggregates against the dispersion of the per-datum values they aggregate. The distributional framing supplies the reason to expect dispersion and a formalism for describing it (arXiv:2506.12965); whether aggregation attenuates it is an open question. A second study would test sensitivity of credit ordering to the choice of performance measure, since value in the equitable-valuation framework is defined relative to a predictor's performance (Data Shapley).

On the documentation side, an extension of the datasheet schema to carry a contributor index, specifying what a person would need in order to establish that their records entered a corpus, would make the evidentiary precondition concrete without asserting an entitlement. The seven-section structure is the natural starting point (COMPEL).

On mechanism design, the open problem is elicitation where an acquirer holds a noisy automated appraisal of each record and many contributors hold private knowledge of quality, with records substituting across contributors. The single-owner acquisition results give the informational structure and the negative result on ordinal input (WWW 2024, pp. 25-35); the many-owner substitutable case is what a corpus actually presents.

On the legal side, a comparative review of entitlement in workplace behavioural records across more jurisdictions than the two examined here would establish whether the gap identified is general or an artefact of the sources available. The personality-rights route, exemplified by the ELVIS Act and the NO FAKES proposals, deserves separate treatment, because it attaches to a person without requiring authorship and is therefore the closer analogue to what a behavioural royalty would need (Copyright Law Meets AI).

Sources

Related work