The legal fight over AI-generated work feels new because the tools are new. The copyright problem is much older.

Source note: Zvi S. Rosen. “The Future Was Then: The Recurrent History of AI Authorship.” Indiana Law Review, Vol. 59, p. 701, 2026. https://mckinneylaw.iu.edu/practice/law-reviews/ilr/pdf/vol59p701.pdf

Why This Paper Matters

Most debates about AI and copyright assume generative AI has created an entirely new authorship crisis. A person types a prompt; a model returns an image, story, or code. The question follows: who is the author?

Rosen’s paper uses a historical lens to slow that conversation down. Instead of treating AI authorship as an unprecedented break, it categorizes generative AI as mediated authorship. This occurs when an intermediary sits between the human creator and the final work, materially shaping the result.

This framing matters because copyright has handled mediated creation for over a century. Cameras, phonograph rolls, aleatory music, computer art, and database compilations forced courts and the Copyright Office to decide where human authorship begins.

The point isn’t that AI authorship is easy to solve, but that the legal system isn’t starting from scratch.

The Idea in Plain English

The paper’s core claim is that the AI authorship debate has a recurring structure.

When a new tool appears to produce expressive material, the legal system must separate three things:

  1. The human’s judgment, selection, or arrangement;
  2. The machine or process that produces the final form;
  3. The finished work.

When these three are tightly bound, authorship is obvious. When a machine contributes expression, the boundary blurs.

Rosen argues that the Copyright Office has used a stable answer for decades: human authorship is necessary. Machine output alone is insufficient. However, human selection, coordination, editing, or expressive prompting can be protectable when they contribute original expression.

What the Researchers Tested

This is a legal-history and doctrinal paper. The evidence is archival and institutional.

Rosen traces several layers of the authorship problem:

  • Current AI cases, including Stephen Thaler’s Creativity Machine and Kris Kashtanova’s Zarya of the Dawn;
  • The Copyright Office’s 1965 annual report, which addressed computer-generated works;
  • Older mediated-authorship examples like photography, player piano rolls, and animal authorship;
  • Early computer works from the 1960s, including software and computer-generated music;
  • Policy lessons for current AI legal battles.

The paper asks one consistent question: when a work comes through a tool, what part of the final expression belongs to the human?

What They Found

The core finding is continuity. Tools change faster than the legal tests for authorship.

The Copyright Office’s 1965 annual report sounds modern. The Office recognized computers could generate works but insisted copyright depended on whether the work reflected human authorship. The 2025 Copyright Office report follows the same path: did a human determine the expression, or did the machine supply it?

Current AI Cases Follow an Older Pattern

Recent disputes follow an established pattern. Thaler’s Creativity Machine claim focused on nonhuman authors. Kashtanova’s Zarya of the Dawn addressed “mixed-work” questions involving text and AI-generated images. Jason Allen’s Theatre D’opera Spatial tested how much iteration and post-processing counts as human authorship.

The paper treats these as the latest round in a recurring cycle: a tool expands what can be made, then the law asks if the human contribution is expressive enough.

Aleatory Music Provides a Useful Model

Composers like John Cage used chance and rules to create music. The final sound varied, but the human provided the structure. This is useful for AI because it separates control from authorship. Copyright doesn’t require a human to determine every detail, but it requires enough human-originated structure to call the work theirs.

Early Computer Art and Music Were Functional

Computer music and art from the 1950s and 60s show that machine-mediated expression isn’t new. People were registering and arguing about computer-assisted works long before chatbots. This history explains why modern legal language feels inherited rather than invented.

The Human-Authorship Anchor

Copyright consistently returns to human authorship. Protection attaches to the human contribution, not autonomous machine generation. In mixed works, the question is which parts reflect human selection, arrangement, or editing.

Why It Happens

Copyright usually ignores the authorship question because the author, tool, and expression are fused. A novelist writes words; a photographer frames a scene. AI makes that fusion visible. The user has an intention, but the model supplies the details. This creates pressure on a system built for human creators.

Historical examples show why the law resists protecting machine-made works. Copyright isn’t just a reward for output; it is tied to human judgment and labor. AI outputs look like “works” but make it difficult to identify an “author” in the legal sense.

What This Means for Builders

For AI product teams, the lesson is clear: authorship needs provenance.

If a tool helps people create protectable work, it must preserve the history of human contribution—prompts, iterations, edits, and selections. The legal value often sits in the process rather than the final model output.

Builders should avoid “one prompt, one work” workflows. Products that support structured selection, editing, and audit trails make it easier to distinguish human-authored elements from machine residue. There is also an opportunity to make authorship visible by helping users document their specific contributions.

What This Means for Buyers and Operators

For companies using AI content, the risk isn’t whether a vendor calls an output “copyrightable.” The real question is which parts can be traced to a human.

This matters for marketing, publishing, and software teams. A deliverable may be useful even if parts aren’t protectable, but buyers should know which rights they actually hold.

Procurement teams should ask for process records rather than just license language. Who selected the output? Who edited it? Was it a compilation, a derivative, or an unmodified model output? The answers change the risk profile.

What to Watch Next

Watch how courts handle “mixed” AI works rather than pure machine-output claims. The harder, commercially relevant cases involve human-machine workflows where authorship is shared.

Monitor whether prompt authorship develops as a practical category. Prompts themselves may be protectable if they contain enough original expression. Finally, watch Copyright Office registration practices, as these will likely shape behavior before appellate courts settle the major issues.

Limitations and Caveats

This is a doctrinal article, not a quantitative study. It doesn’t predict how courts will rule on specific workflows or solve edge cases in multimodal generation.

While historical analogues are useful, they aren’t perfect. A camera and a diffusion model mediate authorship differently. However, copyright already has the vocabulary needed to ask where human expression sits in a mediated process.

Source

Rosen, Zvi S. (2026). The Future Was Then: The Recurrent History of AI Authorship. Indiana Law Review, Vol. 59, p. 701. https://mckinneylaw.iu.edu/practice/law-reviews/ilr/pdf/vol59p701.pdf