A Human-Centric Framework for Data Attribution in Large Language Models

Amelie Wührl, Mattes Ruckdeschel, Kyle Lo, Anna RogersarXiv 2026 · arXiv

TL;DR

The current Large Language Model (LLM) ecosystem exploits web data without giving creators agency or compensation, leading to an unsustainable data economy. This paper introduces a human-centric framework that redefines data attribution as a domain-specific negotiation between creators, users, and AI companies. By bridging technical NLP methods with socio-economic goals, the framework provides a blueprint for fair credit and revenue distribution in the AI era.

What Problem Does This Paper Address?

The generative AI boom is built on a shaky foundation: scraping massive amounts of high-quality data without the consent, control, or economic benefit of its original creators. This creates a double-sided crisis. Creators lose the incentive to produce new content, while LLM users face legal and reputational risks from accidental plagiarism or untrustworthy outputs.

While NLP researchers have developed various technical tools for "data attribution," they often treat it as a purely algorithmic challenge. The authors argue that a single technical solution cannot fix this issue because the problem is fundamentally socio-economic, characterized by massive power asymmetries and deeply conflicting stakeholder goals.

Fig. 1. Major changes in information flow from the creators to readers/users when LLMs started serving as providers of content, including for assisted production of new materials. In addition to stakeholders discussed in §3.1, a relevant group is non-profit organizations who do not serve as intermediaries, but whose outputs (including datasets) may be used by AI industry.

Methods

Instead of introducing a new algorithm, this paper contributes a conceptual, three-step framework to operationalize data attribution based on case-by-case stakeholder negotiations:

  1. Identify Relevant Stakeholders and the LLM Use Case (e.g., creative writing assistance vs. fact-checking).
  2. Negotiate Intended Outcomes (balancing financial, social, and epistemic goals).
  3. Specify What is Attributed and How It's Implemented.

The paper maps these steps to three existing technical approaches in NLP:

  • Content Similarity: Checking if the output matches existing text structurally or conceptually (corroborative attribution).
  • Causal Influence: Using counterfactuals (like influence functions or Shapley values) to measure how much a specific training data point changed the model's behavior (contributive attribution).
  • Data Usage/Documentation: Publicly acknowledging that a dataset was used during development, regardless of its specific impact on a single output.

Additionally, they propose four ranking principles to keep user-facing citations manageable: uniqueness, chronological precedence, extent of shared context, and random sampling as a last resort.

Datasets

As this is a conceptual framework and position paper, no empirical datasets were trained or evaluated. The authors reference existing open-source LLM projects (such as OLMo, Pythia, and BLOOM) and web-scale corpora (like Common Crawl) to contextualize the technical feasibility of data tracking.

Key Insights and Ideas

  • Attribution is a Negotiation, Not Just an Algorithm: A single mathematical definition of attribution cannot satisfy everyone. A fiction writer might care about stylistic similarity and financial compensation, while an academic cares about exact claim verification and citation metrics.
  • Heterogeneous and Conflicting Incentives: Stakeholders have diverse extrinsic (financial, social) and intrinsic motivations. AI companies want to maximize revenue and minimize liability, while users want free, convenient, and legally safe information.
Examples of stakeholder incentives in the data economy.
  • The "Human Learning" Analogy is Flawed: The paper highlights that the common legal defense—"models learn just like humans do"—ignores material differences in precision, speed, and scale that allow AI to exponentially multiply and dilute creative markets.
  • Synthetic Data Won't Save Us: Even if AI companies shift to training on synthetic data, that data is ultimately derived from models trained on human data. Attribution and compensation loops are still required to avoid "data laundering."

Key Results

Because this is a conceptual framework, the "results" are demonstrated through qualitative analysis and real-world case studies:

  • Real-World Validation: The authors demonstrate that their framework accurately tracks emerging industry practices, such as the revenue-sharing partnership between AI startup ProRata and the Danish Press Publications' Collective Management Organisation (DPCMO).
  • The Creative Writing Moonshot: They outline a hypothetical workflow for an LLM-assisted creative writing tool, illustrating how instance-level attribution could trigger micro-payments and social credit for authors directly within a user interface.

Limitations and Future Work

  • Computational Cost: Causal attribution methods (like Shapley values) are currently prohibitively expensive at LLM scale, requiring models to be retrained multiple times to isolate the value of specific data.
  • Human Interpretability: Causal influence metrics reflect internal model states but are rarely intuitive or interpretable for non-expert human creators.
  • Framework Granularity: The framework needs a more fine-grained classification of motivations and explicit mechanisms to protect creators' moral rights, such as the right to object to misrepresentation.
  • Implementation Blueprints: Future work requires collaboration with experts in mediation and participatory design to build concrete negotiation protocols for specific industries.

Summary updated

Abstract

In the current Large Language Model (LLM) ecosystem, creators have little agency over how their data is used, and LLM users may find themselves unknowingly plagiarizing existing sources. Attribution of LLM-generated text to LLM input data could help with these challenges, but so far we have more questions than answers: what elements of LLM outputs require attribution, what goals should it serve, how should it be implemented? We contribute a human-centric data attribution framework, which situates the attribution problem within the broader data economy. Specific use cases for attribution, such as creative writing assistance or fact-checking, can be specified via a set of parameters (including stakeholder objectives and implementation criteria). These criteria are up for negotiation by the relevant stakeholder groups: creators, LLM users, and their intermediaries (publishers, platforms, AI companies). The outcome of domain-specific negotiations can be implemented and tested for whether the stakeholder goals are achieved. The proposed approach provides a bridge between methodological NLP work on data attribution, governance work on policy interventions, and economic analysis of creator incentives for a sustainable equilibrium in the data economy.
BibTeX
@misc{wuhrl2026human,
  author = {Amelie Wührl and Mattes Ruckdeschel and Kyle Lo and Anna Rogers},
  title = {{A Human-Centric Framework for Data Attribution in Large Language Models}},
  year = {2026},
  eprint = {2602.10995},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.10995},
}
Figure