Generative AI models rely on vast datasets scraped from the public web, often bypassing the explicit preferences of original content creators. As Étienne Hubert et al. note in Exploring the mutations of society in the era of generative AI, this practice creates an acute tension between large-scale data collection and the rights of creators to control their intellectual property.
This friction highlights a fundamental confusion between technological capability—what a machine can produce—and artistic authorship—who receives credit for the work, a distinction explored by Aaron Hertzmann in Can Computers Create Art?. When models absorb copyrighted human art or code without explicit permission, they create significant economic and ethical challenges for creators (Sanchaita Hazra et al. in Position: AI Safety should prioritize the Future of Work; Qifeng Chen 0001 et al. in Diffusion-Based Visual Art Creation: A Survey and New Perspectives).
To resolve this, researchers are developing technical mechanisms to inject creator agency directly into the AI lifecycle, spanning proactive consent registries, post-hoc data attribution, and forensic tracking.
Proactive Agency through Consent Registries
The first line of defense for creator agency is controlling whether data enters the training pipeline at all. Traditionally, data collection has been a "black box" operation. However, initiatives like The Stack by Denis Kocetkov et al. (The Stack: 3 TB of permissively licensed source code) demonstrate that prioritizing developer agency through "opt-out" mechanisms does not inherently degrade model performance.
To scale this across the web, researchers have proposed decentralized architectures. For instance, Kar Balan et al. introduced DECORAIT (DECORAIT - DECentralized Opt-in/out Registry for AI Training), a blockchain-based registry that pairs visual fingerprinting with the C2PA provenance standard.
Instead of relying on centralized platforms to respect standard web crawlers, a decentralized registry allows creators to cryptographically sign their work and declare their training preferences. When a dataset scraper processes an image, it computes its visual fingerprint and queries the tamper-proof ledger. If the fingerprint matches an opt-out record, the scraper automatically excludes it. If it matches an opt-in record, the system can trigger automated micropayments to compensate the creator.
Quantifying Causal Influence via Training Data Attribution
Even when data is included in a model, current legal frameworks struggle with binary outcomes: either an output is deemed a fair-use transformation, or it is a copyright infringement requiring a total takedown. This binary approach fails to provide sustainable, recurring compensation for creators, as argued by Junwei Deng et al. in Computational Copyright: Towards A Royalty Model for AI Music Generation Platforms.
To bridge this gap, the authors of that study propose using Training Data Attribution (TDA) to trace the "causal influence" of specific training samples on individual model outputs. TDA mathematically measures how much a specific piece of training data (such as a musical motif or a specific painting style) contributed to the model's final generated distribution.
By quantifying this influence, platforms can move away from flat, one-time licensing fees and instead implement a dynamic royalty model. If a generative model outputs a track or an image that heavily relies on the statistical patterns of a specific artist's work, TDA calculates that exact dependency, allowing the platform to distribute royalties proportionally to the creators who actually drove that specific output.
Forensic Tracking and Self-Reporting Watermarks
Enforcing creator agency requires robust auditing tools to prove whether a model utilized a specific creator's work. While attributing a single generated snippet to a specific training source can be theoretically difficult due to the stochastic nature of generation, Wanlun Ma et al. demonstrate in The "code" of Ethics: A Holistic Audit of AI Code Generators that sets of outputs can be reliably traced using statistical "fingerprints."
Alternatively, models can be trained to actively "self-report" their influences. The WASA framework, introduced by Jingtan Wang et al. in WASA: WAtermark-based Source Attribution for Large Language Model-Generated Data, embeds invisible Unicode watermarks during a second stage of training.
The intuition here is that the model is fine-tuned to map specific content styles to corresponding data-provider watermarks. When the model generates text influenced by a particular source, it automatically embeds the corresponding invisible watermark into the output text. This allows downstream users or automated systems to instantly verify the data provider without sacrificing text quality or readability, making copyright enforcement verifiable at scale.
Reconfiguring the Legal Paradigm
These technical mechanisms are evolving alongside a shifting legal landscape that suffers from severe institutional lag, as noted by Ning Li in Research on Cultural Industry Innovation from the Perspective of New Quality Productivity. Traditional copyright law is strictly anthropocentric, requiring an author's "own intellectual creation" to grant protection, a challenge highlighted by Christin Kirchhübel et al. in Intellectual property rights at the training, development and generation stages of Large Language Models. This creates a legal vacuum for fully autonomous or synthetic outputs (Li Qiwei in A Law of One's Own: The Inefficacy of the DMCA for Non-Consensual Intimate Media).
To resolve this ambiguity, Tianyu Pang argues in Paradigm Reconfiguration of Artificial Intelligence Generated Content (AIGC) Copyright Protection for a "three-stage reconstruction" of copyright law that moves toward a technologically inclusive framework. Instead of treating all AI outputs equally, this paradigm uses a tiered system based on the complexity of human-AI interaction:
- High-Effort Human-AI Collaboration: Works involving significant human curation, prompting, and iterative editing receive full copyright protection.
- Low-Effort Automated Outputs: Purely automated outputs receive limited exclusivity to prevent the public domain from being overwhelmed by machine-generated content.
By combining these tiered legal rights with technical attribution mechanisms like TDA and watermarking, the creative ecosystem can transition from unconsented exploitation to an equitable, traceable model of digital co-authorship.
Key Foundational Work
Artificial Intelligence and Music: Open Questions of Copyright Law and Engineering Praxis (2019) — contributed 6 concept elements to this direction.
Is Photographic Image Represented the Reality? (2019) — contributed 1 concept elements to this direction.
Computational Analysis of the Historical Changes in Poetry and Prose (2019) — contributed 1 concept elements to this direction.