Generative AI outputs belong in the public domain: granting copyright to raw algorithmic output would enclose the digital commons and chill innovation, while only discernible human creative contributions should qualify for protection.

Main Finding

Generative AI (GenAI) outputs should not receive copyright protection. Elmahjub argues from first principles that GenAI’s probabilistic, high-dimensional generation severs the necessary causal link between human creative intent and expressive form; traditional justifications for copyright (utilitarian incentives, labor-desert, and personality theories) therefore cannot coherently support exclusive rights in raw algorithmic outputs. The public domain is the default; raw GenAI outputs belong there, while only genuine, identifiable human creative contributions to AI-assisted works may warrant protection.

Key Points

• Technical basis for the claim

  • Modern GenAI (Transformers, diffusion models, multimodal latent-space synthesis) generates outputs via probabilistic inference rather than deterministic execution; identical prompts can yield materially different outputs.
  • Prompts act as initial conditions in a stochastic process; slight parameter/seed changes can produce unforeseen expressive elements. Human input often represents a tiny fraction of the final product’s tokens or pixels.
  • Models internalize statistical relationships rather than store retrievable copies of training works.

• Philosophical critique of copyright justifications

  • Utilitarian/incentive theory: awarding exclusivity to raw outputs would not clearly improve socially valuable incentives and risks concentrating control of creative commons in incumbents (model owners), harming downstream innovation.
  • Labor-desert (Lockean) theory: attributing ownership to a human prompter or to the model misaligns with the notion of ownership tied to meaningful labor—prompts are often steering, not full expressive labor.
  • Personality (Hegelian) theory: copyright that tracks authorial personality cannot sensibly apply to non-mortal, non-agentic algorithms.

• Doctrinal and comparative law evidence

  • International instruments (Berne Convention, TRIPS) presuppose human authorship (moral rights, life-based terms).
  • Recent U.S. practice and case law (U.S. Copyright Office rejections; Thaler v. Perlmutter; Naruto v. Slater precedent) emphasize human authorship as a bedrock requirement.
  • Some jurisdictions (UK statute on computer-generated works) offer workarounds, but these are legally and philosophically contested and arguably inadequate for modern GenAI.

• Policy conclusion and recommendations

  • The default legal baseline should remain the public domain for raw GenAI outputs to prevent "unprecedented enclosure" of the digital commons.
  • Those who seek copyright for GenAI outputs bear the burden of proof to demonstrate coherent human authorship.
  • A bifurcated approach is appropriate: protect genuine human creative contributions to AI-assisted works, but not raw algorithmic outputs.

Data & Methods

• Methodological approach: normative/legal-philosophical analysis from first principles rather than purely doctrinal patchwork.
• Technical grounding: explanation of GenAI architectures and generation processes (Transformers, self-attention, token prediction, diffusion models, latent-space synthesis) and the stochasticity inherent in sampling and seed/temperature settings.
• Doctrinal survey: review of international treaties (Berne, TRIPS), national statutes (e.g., UK Copyright, Designs and Patents Act §9), administrative practice (U.S. Copyright Office decisions and guidance), and case law (Thaler v. Perlmutter; Naruto v. Slater).
• Philosophical testing: systematic evaluation of the three canonical normative bases for copyright (utilitarian/incentive, Lockean labor/desert, Hegelian personality) against the technical realities of GenAI.
• Comparative policy analysis: survey of scholarly positions (protectionist, AI-personhood, pro-user-arranger) and showing their weaknesses relative to the broken causal chain thesis.
• Limitations: not an exhaustive doctrinal map of all downstream issues (joint authorship allocation, moral rights mechanics), instead setting foundational boundaries for those debates.

Implications for AI Economics

• Market structure and concentration

  • Granting copyright in raw GenAI outputs risks creating proprietary monopolies over large swathes of machine-generated content, reinforcing incumbent model providers’ market power and raising barriers to entry for startups and downstream creators.
  • Enclosure would enable incumbents to extract rents from downstream firms (platforms, aggregators, content repurposers) and could incentivize data- and compute-centric scale economies that further entrench leaders.

• Innovation incentives and social welfare

  • Copyright is traditionally justified as an incentive mechanism. For GenAI outputs, the primary investment incentives for model builders derive from data access, compute scale, and platform monetization—not exclusively from output-level copyrights—so output exclusivity is a poor-targeted subsidy.
  • Placing raw outputs in the public domain lowers transaction costs for downstream experimentation, reuse, and productization, likely accelerating complementary innovation (apps, tools, derivative works) and network effects that benefit consumers and smaller firms.

• Creative labor and distributional effects

  • Denying copyright in raw outputs clarifies rights and reduces litigation risk for users and intermediaries, but it does not eliminate labor-displacement pressures; economic policy should consider alternative support (retraining, new bargaining structures, platform-level compensation schemes) for affected creators.
  • Granting narrow protection only to identifiable human contributions helps preserve incentives for meaningful human creativity and preserves bargaining leverage for artists who genuinely contribute expressive elements.

• Licensing, transaction costs, and legal uncertainty

  • If outputs were copyrightable, complex multi-party ownership (trainers, dataset providers, model builders, prompt engineers, end-users) would create high transaction costs, hinder market transactions, and increase legal frictions—reducing productive dynamic experimentation.
  • Clear public-domain default reduces frictions and fosters a more fluid market for AI-mediated creative goods.

• Alternative policy levers

  • Because copyright is ill-suited to align incentives for GenAI, policymakers should consider other instruments: data-access regulation, mandatory model-interoperability standards, competition policy to limit vertical integration, subsidies or prizes for human-centric creations, and clearer contract/consumer-protection rules for model outputs.
  • Empirical research agendas: quantify effects of public-domain vs. enclosure regimes on venture formation, downstream product variety, pricing, and wages in creative industries to inform optimal policy.

Overall economic takeaway: treating raw GenAI outputs as commons (public domain) likely better supports dynamic innovation, competition, and lower transaction costs, while narrowly protecting genuine human-authored contributions preserves incentives for meaningful human creativity. Granting broad copyright in raw outputs would more likely concentrate rents, raise entry barriers, and reduce social welfare than produce plausible incentive benefits.