Critical ReviewArts & Design

Who Owns AI Art? The Copyright Crisis That Has No Good Answers

When a human types a prompt and an AI generates an image, who is the author? Current copyright law has no coherent answer. Recent work proposes normative frameworks for determining the 'minimum threshold of human authorship,' but jurisdictional fragmentation and conceptual confusion persist.

By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

In 2022, the US Copyright Office rejected a copyright application for "A Recent Entrance to Paradise," an image autonomously generated by the AI system DABUS. In August 2023, a federal judge upheld this rejection in Thaler v. Perlmutter, ruling that copyright requires human authorship. In February 2023, the Copyright Office issued a limited registration for "Zarya of the Dawn," a graphic novel that combined human-written text with AI-generated imagesโ€”granting copyright to the text and arrangement but not to the individual AI-generated images. Each decision was consistent with existing law; together, they exposed a conceptual vacuum at the center of intellectual property doctrine.

The core problem is that copyright law was built on an assumption that seemed self-evident for centuries: creative works are produced by humans. Generative AI disrupts this assumption without providing a clean alternative. The four papers reviewed here represent different attempts to navigate the resulting confusion.

The Research Landscape: Searching for Authorship Thresholds

Ramos-Zaga (2025), with 1 citation, provides the most analytically rigorous contribution. His normative framework begins by identifying the anthropocentric premise that copyright law takes for grantedโ€”"human intervention as a precondition for protection"โ€”and asks what minimum level of human involvement should be required for copyright protection of AI-assisted works.

The framework distinguishes three levels of human involvement:

  • Specification (lowest): The human provides a text prompt; the AI generates the work. Ramos-Zaga argues this is analogous to commissioning an artist with a briefโ€”the commissioner does not become the author of the resulting work. Under this analysis, a simple prompt should not confer authorship.
  • Direction (intermediate): The human makes iterative creative choicesโ€”selecting outputs, modifying prompts, combining elements, curating from alternatives. This involves genuine creative judgment, even though the human does not produce the visual or textual content directly. Ramos-Zaga argues this level could meet the threshold, depending on the degree of selection and arrangement.
  • Collaboration (highest): The human and AI engage in a sustained creative process where the human makes substantive creative contributions beyond selectionโ€”composing parts of the work, significantly modifying AI outputs, integrating AI and human elements into a coherent whole. This level, Ramos-Zaga argues, clearly meets the authorship threshold.
  • The framework is helpful for its analytical clarity, but it faces practical challenges. How do you distinguish "direction" from "specification" in practice? A sophisticated prompt engineer may exercise considerable creative judgment in crafting prompts, yet the output may look identical to what a naive prompt would produce. The framework requires a way to assess the process of creation, not just the productโ€”and copyright law has traditionally focused on the product.

    The Legislative Gap

    Zain and Kareem (2025) survey the legislative landscape across jurisdictions and find pervasive gaps. The EU, US, China, and UK take different approaches:

    • US: No copyright for works without human authorship. AI-generated elements are unprotectable; human elements may be protected.
    • EU: Similar to US, but complicated by the sui generis database right that could protect AI-generated compilations.
    • China: Some courts have granted copyright to AI-generated works under a "work made for hire" analogyโ€”treating the AI user as the "employer" and the AI as the "employee."
    • UK: The Copyright, Designs and Patents Act already provides for "computer-generated works" (section 9(3)), granting authorship to the person who makes the "arrangements necessary for the creation of the work." But what counts as "arrangements necessary" is undefined.
    The fragmentation matters because AI-generated content crosses borders effortlessly. An image generated by an American user using a European AI service, distributed on a Chinese platform, faces three different legal regimes. The authors argue for international harmonization, but the historical record of copyright harmonization (decades-long negotiations over relatively minor provisions) suggests this will not happen quickly.

    Computational Creativity: Can Machines Be Authors?

    Paksi (2025) approaches the question from the other direction: rather than asking what human contribution copyright requires, he asks whether AI systems can exhibit creativityโ€”the capacity that authorship is meant to protect.

    His review of the computational creativity literature identifies three criteria commonly used to evaluate machine creativity:

    • Novelty: The output is newโ€”not a direct copy of training data.
    • Value: The output is evaluated positively by some audience.
    • Process: The generation process involves something more than random combinationโ€”some form of evaluation, selection, or goal-directed behavior.
    Current generative AI systems satisfy the first criterion routinely and partially satisfy the second (many AI outputs are aesthetically compelling). The third criterion is where the debate concentrates. A diffusion model does not have goals; it minimizes a loss function. Whether loss minimization constitutes a form of goal-directed behavior is a philosophical question, not an empirical one.

    The Convergence Problem

    Hintze and Schossau (2025) contribute an empirical finding that complicates the creative-AI narrative. They constructed autonomous AI-to-AI creative systemsโ€”feedback loops where an image generator (Stable Diffusion XL) produces images, a vision-language model (LLaVA) describes them, and the descriptions are fed back as prompts for new images.

    The finding: these autonomous loops systematically converge toward generic visual motifs. Left to iterate without human intervention, the systems produce increasingly homogeneous outputsโ€”losing the diversity and specificity that characterizes creative work. The convergence is consistent across different initializations and model combinations.

    This is a significant empirical constraint on claims about autonomous AI creativity. While a single AI-generated image can be novel and aesthetically interesting, the autonomous creative processโ€”generation without human directionโ€”tends toward mediocrity. The implication for the authorship debate is that the human contribution in human-AI creative collaboration may be precisely the element that prevents this convergence: human taste, selectivity, and direction maintain the diversity and intentionality that distinguish creative work from statistical output.

    Critical Analysis: Claims and Evidence

    <
    ClaimEvidenceVerdict
    A normative threshold for human authorship in AI-assisted works can be definedRamos-Zaga's three-level frameworkโš ๏ธ Uncertain โ€” analytically clear but practically difficult to apply
    International copyright law is inadequately prepared for generative AIZain et al.'s multi-jurisdictional surveyโœ… Supported โ€” significant gaps and fragmentation
    Current AI systems do not exhibit genuine creativityPaksi's analysis against creativity criteriaโš ๏ธ Uncertain โ€” depends on how "goal-directed behavior" is defined
    Autonomous AI creative processes converge to generic outputsHintze et al.'s autonomous loop experimentsโœ… Supported โ€” consistent convergence across conditions

    Open Questions and Future Directions

  • The training data question: Even if AI outputs are not direct copies, they are derived from copyrighted training data. Does this constitute infringement? Several ongoing lawsuits (Getty v. Stability AI, NYT v. OpenAI) will shape the answer.
  • Moral rights: Copyright includes not just economic rights but moral rights (attribution, integrity). Can an AI have moral rights? Should the human prompter? The training data creators?
  • Proof of process: If authorship depends on the creative process (as Ramos-Zaga argues), how is the process documented and verified? AI art tools do not currently log the iterative creative decisions that distinguish "direction" from "specification."
  • Open-source implications: Many AI art tools are open-source. If the tool is freely available and the training data is publicly sourced, is the resulting artwork in the public domain?
  • Economic realities: Regardless of legal answers, AI-generated content is flooding creative markets. What economic protections should human artists receive, and through what mechanisms?
  • What This Means for Your Research

    For legal scholars, Ramos-Zaga's framework offers a starting point for jurisdiction-specific analysis. The "specification-direction-collaboration" taxonomy maps onto existing copyright concepts in ways that could inform legislative reform.

    For artists working with AI, documenting your creative process is increasingly importantโ€”both for potential copyright claims and for artistic credibility.

    For AI researchers, Hintze et al.'s convergence finding is a reminder that autonomy is not creativity. The aesthetic value of AI-generated art appears to depend on human direction more than the technology's proponents sometimes acknowledge.

    Explore related work through ORAA ResearchBrain.

    References (4)

    [1] Ramos-Zaga, F.A. (2025). Reconceptualizing Human Authorship in the Age of Generative AI: A Normative Framework for Copyright Thresholds. Laws, 14(6), 84.
    [2] Zain, A., Yousuf, A., & Kareem, A. (2025). The Legislative Gap for Copyright in the Era of Generative AI: Where do We Stand in Achieving Sustainable Development Goals? Journal of Lifestyle and SDGs Review, 5(4).
    [3] Paksi, D.N.F. (2025). Beyond Human Authorship: Exploring Computational Creativity and Machine-Led Aesthetics. Harmonia, 3(3).
    [4] Hintze, A., Proschinger ร…strรถm, F., & Schossau, J. (2025). Autonomous language-image generation loops converge to generic visual motifs. Patterns, 6, 101451.

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 7 keywords โ†’