A human-centric attribution framework would let creators, users and platforms negotiate when and how LLM outputs should be traced to training data, clarifying ownership and helping prevent unknowingly copied content; its success, however, hinges on technical feasibility and cross-stakeholder incentives.

Main Finding

The paper proposes a human-centric framework for data attribution in large language models (LLMs). Attribution should be specified and implemented case-by-case via negotiated parameters that reflect the heterogeneous objectives of stakeholders (creators, publishers, platforms, AI companies, users). This approach connects NLP methods for attribution with governance and economic analysis, aiming to support sustainable incentives in the LLM data economy rather than a single technical solution.

Key Points

• Problem framing

  • LLMs reshape information flows: creators’ works can be used without consent or disclosure; LLMs may compete with original sources and enable large-scale low-cost derivative production.
  • Current lack of disclosure produces ethical, legal, and economic frictions: creators lose agency and revenue; users face plagiarism and trust issues.
  • Attribution is a multi-level problem: instance-level (which source supported a specific output), dataset-level (which datasets influenced a model), and functional-level (which data improved abilities on tasks).

• Stakeholders & incentives

  • Primary: creators and readers/users.
  • Intermediaries: publishers, platforms, AI industry.
  • Incentives are heterogeneous (financial, social, intrinsic) and often conflicting; market power asymmetries favor large intermediaries.
  • Attribution matters differently depending on incentive: e.g., social recognition requires fine-grained attribution; some financial arrangements might work with coarse dataset-level attribution.

• State of methods & limits

  • Existing technical approaches include influence functions, Shapley-value–based valuation, similarity searches, data extraction tests, RAG (retrieval-augmented generation), and training-time interventions (weighting, modular architectures).
  • Many methods are post hoc, incomplete, and require access to training data or model internals; commercial opacity of training corpora is a major barrier.
  • There is no one-size-fits-all attribution metric because different use cases demand different trade-offs (precision, recall, granularity, privacy, computation cost).

• The proposed human-centric framework

  • Attribution should be grounded in negotiated, domain-specific parameterizations: define stakeholder objectives, what outputs require attribution, acceptable technical guarantees, privacy and IP constraints, and enforcement/verification mechanisms.
  • Stakeholders (creators, users, intermediaries) negotiate criteria; outcomes are implemented and empirically tested against the stated goals.
  • The framework acts as a bridge: specifying practical goals enables targeted methodological NLP work and informs policy/economic analysis (e.g., compensation schemes, market design).

• Vision / moonshot

  • The authors sketch an “attribution-backed” LLM service model: systems that provide provenance information (instance- and dataset-level), configurable by negotiated parameters, enabling attribution-driven monetization, transparency, and trust.
  • Such services would need infrastructure, standards, and governance to operationalize negotiated attribution regimes.

Data & Methods

• Type: conceptual, theoretical, and synthesis paper — not an empirical/experimental study.
• Methods:
  • Literature review across NLP attribution methods, data governance, data valuation, and related legal/economic scholarship.
  • Stakeholder analysis identifying roles, incentives, and conflicts in the LLM data ecosystem.
  • Systematization of attribution use cases and articulation of negotiable parameters for implementation.
  • Mapping of existing technical approaches to practical goals and constraints; discussion of feasibility and limits.
• Evidence base: prior methodological work (influence functions, Shapley values, RAG, differential-privacy notions of attribution, data valuation), case examples (e.g., disputes over training on copyrighted corpora), and conceptual economic/legal arguments. No original quantitative datasets or experiments are presented.

Implications for AI Economics

• Incentive alignment and supply of creator data

  • Fine-grained attribution can restore social and economic incentives for creators (recognition, bargaining power, licensing fees), potentially increasing supply and quality of data available to model developers.
  • Different attribution granularities map to different pricing/compensation designs (per-instance micropayments vs. dataset licenses vs. reputation-based rewards).

• Market structure and bargaining

  • Attribution standards could rebalance power between individual creators and large intermediaries by enabling collective bargaining, transparency, and enforceable revenue-sharing mechanisms.
  • Implementable attribution lowers information asymmetries, which can change negotiation dynamics, licensing markets, and the competitive landscape among AI firms.

• Productization and new business models

  • Attribution-enabled LLM services could monetize provenance (premium access to source links, pay-to-view sources, creator opt-ins), create marketplaces for attributed data, and differentiate products on trust/transparency features.
  • Platforms might offer configurable attribution settings tied to subscriptions or revenue-sharing agreements.

• Regulatory and transaction-cost effects

  • Attribution frameworks reduce enforcement costs for rights-holders if provenance is verifiable, easing legal frictions and potentially lowering litigation.
  • Conversely, implementing negotiated attribution regimes imposes governance and compliance costs; the net welfare effect depends on who bears them and whether standards are interoperable.

• Welfare and equilibrium concerns

  • If properly designed, attribution can sustain a healthier long-run data economy (more creator participation, higher-quality training data), increasing overall welfare from AI services.
  • Risks: lobbying or platform non-cooperation could lock in suboptimal equilibria; overly burdensome attribution regimes could stifle innovation or raise costs passed to users.

• Research & policy agenda for economists

  • Quantify how different attribution granularities affect creators’ willingness to supply content and optimal pricing.
  • Model bargaining and market-power effects under varying transparency regimes.
  • Assess distributional impacts (who captures surplus) across creators, intermediaries, AI firms, and users.
  • Evaluate transaction costs and the welfare trade-offs between enforcement, privacy, and innovation.

Limitations and open questions - Technical feasibility: robust, scalable instance-level attribution remains unsolved, especially without transparent data access. - Strategic resistance: incumbents may resist disclosure that reduces their rent extraction. - Legal heterogeneity: attribution interacts with copyright law, fair-use doctrines, and cross-jurisdictional differences. - Need for empirical validation: the framework requires piloting in specific domains to assess whether negotiated parameters actually achieve stakeholder goals.

Overall, the paper provides a normative and operational blueprint for tying attribution design to stakeholder objectives, enabling coordinated technical, economic, and policy interventions rather than relying on a universal technical fix.