Creative Ownership in the Age of AI

1. Problem Statement

The paper addresses a critical gap in current copyright law regarding Generative AI. Existing legal frameworks rely on the doctrine of "substantial similarity," which determines infringement based on whether an AI output closely resembles a protected work. However, generative AI can imitate an author's style, tone, or structure without copying specific content, thereby evading traditional infringement claims while still critically depending on the original work for its training.

The authors argue that current definitions are ill-suited for this setting. They propose a new, counterfactual criterion: an AI output infringes on an existing work if and only if that output could not have been generated without that specific work in the training corpus. The paper seeks to mathematically characterize the set of "permissible" (non-infringing) generations under this new definition and analyze how this set evolves as the corpus of training data grows.

2. Methodology and Framework

2.1 Mathematical Modeling

The authors model the creative space and generative process using tools from convex geometry and set theory:

• Creations: Modeled as points $c \in \mathbb{R}^d$ (vectors representing novels, images, etc.).
• Corpus: A subset $C \subset \mathbb{R}^d$ of existing works.
• Generator ($g$): A map $g: \mathcal{C} \to \mathcal{C}$ that takes a corpus and returns the set of generable outputs. The generator is defined as a closure operator satisfying three axioms:
  1. Preservation: $C \subseteq g(C)$ (Existing works remain generable).
  2. Monotonicity: $C \subseteq D \implies g(C) \subseteq g(D)$ (Larger corpora expand generable outputs).
  3. Idempotence: $g(g(C)) = g(C)$ (Applying the generator to its own output yields no new creations).
• Examples of Generators:
  • Convex Hull Generator ($g_{conv}$): Generates all convex combinations of inputs.
  • Splice Generator ($g_{splice}$): Generates outputs by combining features (dimensions) from different inputs.
  • Box Generator: A composition of convex hull and splice operations.

2.2 Definitions of Infringement and Permissibility

• Counterfactual Dependence: A creation $x$ is a violation of a specific work $c \in C$ if $x \in g(C)$ but $x \notin g(C \setminus \{c\})$.
• Violation Set ($v_g(C)$): The union of all violations across all works in the corpus.
• Permissible Set ($p_g(C)$): The intersection of all "c-permissible" sets. Formally:
  $$p_g(C) = \bigcap_{c \in C} g(C \setminus \{c\})$$
  These are outputs that remain generable even if any single existing work is removed from the training data.

3. Key Contributions and Results

3.1 Structural Properties of the Permissible Set

The paper establishes several fundamental properties of the permissible set:

• Monotonicity: As the corpus $C$ grows, the permissible set $p_g(C)$ weakly expands ($C \subseteq D \implies p_g(C) \subseteq p_g(D)$). Adding more works relaxes constraints because it provides alternative generative paths.
• Stability (Closure): The permissible set is closed under the generator: $g(p_g(C)) = p_g(C)$. Combining non-infringing works cannot produce an infringement.
• Non-Emptiness Condition: The authors derive a sufficient condition for the permissible set to be non-empty based on the Radon number ($R(g)$) from convex geometry. If the corpus size $|C| \ge R(g)$, then $p_g(C) \neq \emptyset$. For the convex hull generator in $\mathbb{R}^d$, this threshold is $d+2$.
• Comparative Statics:
  • Adding a permissible work to the corpus does not change the permissible set.
  • Adding a violation work to the corpus strictly expands the permissible set (because the violation becomes a primary source, breaking dependence on the original work).

3.2 Asymptotic Analysis: Light-Tailed vs. Heavy-Tailed Innovation

The paper's central result (Theorem 1) characterizes the long-run behavior of the permissible ratio ($r_g(C_n)$), defined as the volume of the permissible set divided by the volume of the generable set, as the corpus size $n \to \infty$.

• Light-Tailed Distributions (Gradual Innovation):
  • If new works are drawn from a distribution with uniform light tails (e.g., Multivariate Normal), the permissible ratio converges to 1 almost surely.
  • Implication: In markets with incremental innovation, the corpus eventually becomes so rich that every generable output has multiple paths. No single work is essential, and regulation imposes no limits on AI generation in the long run.
• Heavy-Tailed Distributions (Breakthrough Innovation):
  • If new works follow a heavy-tailed distribution (e.g., Pareto, where "superstar" breakthroughs occur), the permissible ratio does not converge to 1.
  • Implication: A positive measure of outputs remains violations indefinitely. The most recent "breakthrough" work often defines the frontier of the creative space, making it essential for generating nearby outputs. In these fields (e.g., avant-garde art), individual infringement claims remain potent even in large markets.

3.3 Generalization to Group Violations

The authors extend the model to allow for collections of protected works (e.g., a class action or a portfolio).

• Richness: A collection of protected works is "richer" if it contains larger groups or more works.
• Superadditivity: The violation set of a coalition is strictly larger than the union of individual violation sets. This suggests that collective bargaining by creators strengthens their position against AI generators.

4. Significance and Implications

• Legal Reform: The paper proposes a shift from "substantial similarity" (resemblance) to "counterfactual dependence" (necessity). This aligns legal liability with the economic reality of how AI models function.
• Market Differentiation: The results suggest that copyright policy should not be one-size-fits-all.
  • Incremental Markets (e.g., genre fiction, commercial movies): Regulation may become obsolete as the corpus grows, as individual works lose their "essential" status.
  • Breakthrough Markets (e.g., avant-garde art, paradigm-shifting literature): Regulation remains necessary to protect the "frontier" works that define the creative space.
• Strategic Behavior: The paper hints at a potential equilibrium where creators might strategically position themselves at the "frontier" (heavy tails) to maintain essentiality and avoid being rendered dispensable by the AI's growing corpus.
• Implementation: While retraining models to test counterfactuals is computationally expensive, the authors suggest using machine unlearning techniques or influence functions as practical proxies for legal adjudication.

Conclusion

Liang and Lu provide a rigorous mathematical framework demonstrating that the viability of copyright protection for AI training data depends fundamentally on the statistical nature of human creativity. If creativity is incremental (light-tailed), AI will eventually generate everything without needing specific inputs, rendering individual infringement claims moot. If creativity is driven by rare breakthroughs (heavy-tailed), specific works remain essential, and infringement risks persist indefinitely. This dichotomy offers a new lens for policymakers to evaluate the future of intellectual property in the age of generative AI.