Shuichiro Ogawa
日本語

Notes · updated 2026-06-28

Overview

This note organizes the generative art ecosystem from an industry perspective along four axes: evaluation, copyright/authorship, democratization, and exploration. Sources were collected across three tiers — T1 (public agencies and primary sources), T2 (industry research and platforms), and T3 (practitioner first-person testimony) — yielding 49 data points (3 excluded as heavy, 10 items requiring primary verification). The scholarly lineage (historical origins, technical genealogy, theoretical frameworks, and post-LLM research) is covered in generative-art-literature.

T1: Public Agencies and Primary Sources

The legal landscape surrounding copyright in AI-generated works diverges by jurisdiction.

United States: The requirement of human authorship as a foundational condition for copyright has been firmly established. Thaler v. Perlmutter (2023 district court, 2025 appellate court, 2026 Supreme Court cert denied) rejected registration of AI as sole author at every level. However, the appellate court explicitly stated that the ruling “does not prohibit copyright in works created with the assistance of AI,” leaving open the possibility of protection where human involvement is sufficient1. The USCO’s March 2023 guidance mandated explicit exclusion of AI-generated content, and the Zarya of the Dawn letter (February 2023) excluded individual Midjourney-generated images while protecting the text and arrangement.

China: The Beijing Internet Court (November 2023, Li Yunkai v. Liu Yuqian) held that Stable Diffusion-generated images are copyrightable, reasoning that “prompt selection and parameter settings reflect the individuality of human expression.” In contrast to the U.S. approach, this ruling recognizes copyright in AI-generated images where human input is deemed sufficient.

Japan: The Agency for Cultural Affairs’ “Approach to AI and Copyright” (March 2024) confirmed that non-expressive uses under Article 30-4 of the Copyright Act do not require permission, while clarifying that the exception does not apply when outputs reproduce the expression of training data. LoRA-based style imitation falls outside the exception.

EU: The AI Act (Regulation (EU) 2024/1689), Article 50, establishes marking and disclosure obligations for AI-generated content (applicable from August 2, 2026). An exception for artistic and creative works limits disclosure to “appropriate methods that do not impair enjoyment.” The impact of Article 50 on design practice is examined in eu-ai-act-design-impact.

United Kingdom: A March 2026 report examined four policy options based on 11,520 public consultation responses. The CDPA Section 9(3) already provides existing protection for computer-generated works.

Museums and Cultural Institutions

Major museums are simultaneously conferring institutional legitimacy on generative/AI art and situating it within art-historical context.

  • MoMA: Accessioned Refik Anadol’s “Unsupervised” into its permanent collection in October 2023 (the first generative AI artwork in MoMA’s collection).
  • Tate Modern: “Electric Dreams: Art and Technology Before the Internet” (November 2024 — June 2025) positioned the AI boom within a “pre-history.”
  • Whitney Biennial 2024: Commissioned Holly Herndon and Mat Dryhurst’s “xhairymutantx,” a work that embedded consent to AI training models as its subject matter.
  • Ars Electronica: Established the “AI in ART Award” in 2024. AI works constituted approximately 25% of entries in the New Animation Art category.
  • Sony World Photography Awards 2023: Boris Eldagsen won with an AI-generated image and then declined the award, stating: “AI images and photography should not compete. They are different entities.”

Provenance Standards

C2PA (Coalition for Content Provenance and Authenticity) leads the standardization effort for tracking AI-generated content. Specification v2.1 (September 2024) defines a framework for attaching cryptographic signature metadata to digital assets. OpenAI and Amazon joined the Steering Committee in 2024. NIST AI 100-4 (November 2024) organized risk mitigation techniques for synthetic content, though its underlying Executive Order 14110 was rescinded in January 2025.

T2: Industry Research and Platforms

The Generative Art Market

Art Blocks recorded cumulative transaction volume of approximately $1.47B (as of July 2024), with 495 flagship projects and approximately 177,000 works. However, trading volume has declined 95% from its 2021 peak, with sales count down 88%. fxhash reports over 25,000 projects, approximately 8,000 artists across 80+ countries, and cumulative circulation of approximately $40-45M, though the source is a VC (investor) blog and requires independent verification.

The NFT art market overall contracted 93%, from $2.9B in 2021 to $23.8M in Q1 2025. Active traders fell 96%, from 529,101 to 19,575. However, Art Blocks Curated flagship works continue to maintain high floor prices: Fidenza at a floor of $34,525 (market cap approximately $34.5M) and Chromie Squiggle at a floor of $4,446 (market cap approximately $44.5M).

Auction houses have conferred institutional legitimacy on generative art. Beeple’s “Everydays” sold for $69.3M at Christie’s (March 2021), Dmitri Cherniak’s “Ringers #879 ‘The Goose’” sold for $6.2M at Sotheby’s (June 2023), and Sotheby’s “Grails” Three Arrows Capital collection totaled approximately $17M.

The global art market was estimated at $57.5B in 2024. A survey of 3,100 HNW collectors found 51% had purchased digital art, which ranked third in spending share (approximately 14%).

Creative Coding Tools

Processing launched in 2001 and has reached its 25th anniversary. p5.js has approximately 23,700 GitHub stars. openFrameworks has approximately 10,300 stars. TouchDesigner and Cables.gl do not publish quantitative usage data. The Processing Foundation describes its user base as “tens of thousands of students, artists, and designers” but does not disclose specific figures.

AI Art Tools (Text-to-Image)

Over 15 billion images have been generated across all AI text-to-image tools (as of 2024, Everypixel estimate), at approximately 34 million per day. The breakdown is Stable Diffusion variants at 12.6 billion (approximately 80%), Adobe Firefly at 1 billion (self-reported 22 billion assets reached by April 2025), Midjourney at 964 million, and DALL-E 2 at 916 million. Midjourney has approximately 19.83 million registered users (estimated, as it is a private company [requires primary verification]), with 2025 revenue of $500M. Civitai has 4 million users and 87,000 models.

T3: Practitioner First-Person Testimony

Democratization: The Design Intent of Processing/p5.js

Casey Reas described Processing’s purpose as “bringing ideas and technology out to the larger world beyond MIT” and “putting code into the hands of artists, architects, and designers around the world as a conceptual medium rather than a technical one”2. He understands software as “a medium in its own right,” distinct from “a digital version of the darkroom.” His shared interest with Ben Fry was “how to teach programming to students with visual ideas,” and Processing was designed as a “digital sketchbook” for interactive graphics.

Zach Lieberman, through openFrameworks and the School for Poetic Computation (SFPC), stated: “Technology should always be in service of ideas. Ideas should be poetic, about what it means to be human”3. The founding of SFPC was motivated by disillusionment with the commercialization of universities.

Exploration and Evaluation: The Tension between Generation and Curation

Tyler Hobbs articulated the distinctive constraints of long-form generative art: “A design that produces 95% garbage and 5% treasure does not work for long-form. Every output must meet the standard”4. He spent approximately two months on quality assurance for Fidenza. On curation, he observed that “in generative systems, a substantial portion of the creativity shifts to a curatorial role,” characterizing curation as “everything, the truly creative act.” For QQL, working under Art Blocks’ no-curation constraint, he identified emergence — “those moments when random elements sync in just the right way and something truly new and unexpected is born” — as the most important quality.

This structure of “constraint-setting, generation, selection” is shared with historical predecessors. Vera Molnar stated: “In life you cannot do everything. You always have to choose. And then the algorithm comes”5. Frieder Nake formulated the principle: “Think the image, don’t make it. In algorithmic art, making is delegated to the machine”6. Harold Cohen initially described AARON as “only I can change the rules. AARON is the rules themselves. The most remarkable artist’s assistant in history, not an artist,” but in his later years shifted to: “I no longer think much about the program’s autonomy. I think of it as a collaborator”7.

Holly Herndon positioned training data as “children of the mind sent into the future” and built a consent protocol for AI training data through Spawning.ai (which also influenced the EU AI Act)8. Mat Dryhurst distinguished “sampling and spawning are fundamentally different,” arguing that “being pro-AI and being pro-consent are not mutually exclusive”9.

Refik Anadol stated that “data is not numbers but a form of memory,” positioning machines as “extensions of the mind” and collaborators. Memo Akten, by contrast, stated: “I’m interested in using machine learning as a mirror that reflects back how we understand the world,” demonstrating equal concern with AI’s cultural and ethical implications as with the technology itself10.

Resistance Movements: Anti-AI Art

In the December 2022 ArtStation protest, thousands of artists simultaneously uploaded “No To AI Images” banners in response to AI-generated images appearing on the homepage. As a technical countermeasure, Ben Zhao at the University of Chicago developed Glaze (style protection, 8.5 million downloads [requires primary verification]) and Nightshade (data poisoning, 2.5 million downloads [requires primary verification]). His motivation was “a deterrent to shift the balance of power from AI companies back to artists,” sparked by contact with the Concept Art Association in November 202211.

Answers to Six Key Questions

1. What is the market size of generative art? Art Blocks cumulative transaction volume is approximately $1.47B. Fidenza market cap is approximately $34.5M; Chromie Squiggle approximately $44.5M. However, the NFT art market overall contracted 93% from a peak of $2.9B (2021) to $23.8M (Q1 2025). The global art market totals $57.5B (2024), with 51% of HNW collectors purchasing digital art.

2. How many users do creative coding tools have? p5.js has approximately 23,700 GitHub stars; openFrameworks approximately 10,300 stars. The Processing Foundation describes “tens of thousands” but does not disclose specific figures. TouchDesigner and Cables.gl likewise lack quantitative data. By contrast, AI image generation tools operate at orders-of-magnitude greater scale: Midjourney at approximately 20 million users, Stable Diffusion variants at 12.6 billion images generated, Civitai at 4 million users.

3. What are the copyright/authorship rulings and their implications? The U.S. requires “human authorship” and has denied copyright to AI as sole author (Thaler, finalized). However, AI-assisted works are not excluded. China recognizes copyright where human input is sufficient. Japan restricts fine-tuning for style imitation. The EU establishes transparency obligations (Art. 50) with an exception for artistic works. Inconsistencies across jurisdictions create international legal uncertainty.

4. How do practitioners describe their evaluation processes? A three-stage structure of “constraint-setting, generation, selection” is common. Tyler Hobbs: curation is “the truly creative act,” with two months devoted to quality assurance. Vera Molnar: “You always have to choose.” Frieder Nake: “Think the image, don’t make it.” Harold Cohen: a shift in perception from assistant to collaborator.

5. What are the barriers to entry for generative art? Processing/p5.js explicitly aimed to lower barriers to entry (“teaching programming to students with visual ideas,” “code as a conceptual medium”). Its use as educational material is mapped in generative-art-education-literature. However, the need for programming skills persists. AI text-to-image tools eliminate the programming requirement and dramatically lower the entry barrier, but they involve a fundamentally different creative process from “generative art.”

6. How has AI art (text-to-image) disrupted the generative art community? Disruption has occurred along three dimensions. (a) Copyright uncertainty: Prompt input alone does not establish copyright (U.S.), whereas generative artists who write their own algorithms retain conventional copyright protection. (b) Market disruption: The flood of AI-generated images triggered the ArtStation protest and gave rise to technical countermeasures such as Glaze and Nightshade. (c) Conceptual conflation: There is a risk that “AI art” and “generative art” are being confused. Generative artists design the algorithms themselves, while text-to-image users input prompts to pre-trained models. Tyler Hobbs’ “two months of quality assurance” and Midjourney’s instant generation represent fundamentally different processes.

References

T1: Public Agencies and Primary Sources

T2: Industry Research and Platforms

T3: Practitioner Testimony

Footnotes

  1. U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 2: Copyrightability” (2025-01-29). https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf ↩

  2. Casey Reas, “A Modern Prometheus,” Processing Foundation (Medium), 2018-05-29. https://medium.com/processing-foundation/a-modern-prometheus-59aed94abe85 ↩

  3. Zach Lieberman, ARTECHOUSE interview. https://www.artechouse.com/all-about-creative-coding-artist-zach-lieberman/ ↩

  4. Tyler Hobbs, “The Rise of Long-Form Generative Art,” 2021-08. https://www.tylerxhobbs.com/words/the-rise-of-long-form-generative-art ↩

  5. Vera Molnar, Right Click Save interview, 2022-08. https://www.rightclicksave.com/article/an-interview-with-vera-molnar ↩

  6. Frieder Nake, DAM Museum interview, 2021-11. https://dam.org/museum/essays_ui/essays/frieder-nake-interview-lines-fiction/ ↩

  7. Harold Cohen, Computer History Museum. https://computerhistory.org/blog/harold-cohen-and-aaron-a-40-year-collaboration/ ↩

  8. Holly Herndon, AnOther Magazine (in conversation with Hans Ulrich Obrist), 2024. https://www.anothermag.com/art-photography/15858/art-in-the-age-of-ai-holly-herndon-mat-dryhurst-with-hans-ulrich-obrist ↩

  9. Mat Dryhurst, Inverse, 2022. https://www.inverse.com/input/culture/mat-dryhurst-holly-herndon-artists-ai-spawning-source-dall-e-midjourney ↩

  10. Memo Akten, Artnome interview, 2018-12. https://www.artnome.com/news/2018/12/13/machine-learning-art-an-interview-with-memo-akten ↩

  11. Ben Zhao, MIT Technology Review, 2023-10-23. https://www.technologyreview.com/2023/10/23/1082189/data-poisoning-artists-fight-generative-ai/ ↩

← All Notes · Home