Source-linked AI summary
Autonomy, Authenticity, Authorship and Intention in computer generated art
Jon McCormack, Toby Gifford, Patrick Hutchings
TL;DR
The Belamy sale renewed questions about autonomy, authorship, authenticity, and intention, including whether deep learning changes decades of prior discussion. The authors address five questions about AI-generated art: authorship and rights, mimicry versus independent creativity, whether deep-learning techniques differ from other generative algorithms, and whether AI art is a new kind. Current popular machine-learning art systems trained on curated human datasets rely on mimicry, while artists, developers, trainers, and data contributors may share authorship.
Problem
The Belamy sale renewed questions about autonomy, authorship, authenticity, and intention, including whether deep learning changes decades of prior discussion.
Source: §1.1 · p. 4Method
The authors address five questions about AI-generated art: authorship and rights, mimicry versus independent creativity, whether deep-learning techniques differ from other generative algorithms, and whether AI art is a new kind.
Source: §1.1 · p. 4Results
Current popular machine-learning art systems trained on curated human datasets rely on mimicry, while artists, developers, trainers, and data contributors may share authorship.
Source: §3.1 · p. 12, §3.3 · p. 13Takeaways & Limitations
In contemporary Western contexts, attribution should attend to human contributors, since existing AI intelligence falls short of popular imagination.
Source: Summary · p. 10The paper leaves auction economics, creator motivations, and artistic quality outside its scope because they require broader market, behavioral, and art-critical research.
Source: §1.1 · p. 4
Abstract
from arXiv · showhide
This paper examines five key questions surrounding computer generated art. Driven by the recent public auction of a work of `AI Art' we selectively summarise many decades of research and commentary around topics of autonomy, authenticity, authorship and intention in computer generated art, and use this research to answer contemporary questions often asked about art made by computers that concern these topics. We additionally reflect on whether current techniques in deep learning and Generative Adversarial Networks significantly change the answers provided by many decades of prior research.
1 Introduction: Belamy’s Revenge
Christie’s 2018 sale of Portrait of Edmond Belamy revived longstanding questions about computer-generated art, prompting a review of prior debate and whether deep learning changes its conclusions.
1 Introduction: Belamy’s Revenge
- Christie’s sold Portrait of Edmond Belamy for US$432,500, more than 40 times its pre-auction estimate.Source: §1 · p. 1
- Obvious relied on code written by Robbie Barrat, who received neither credit nor sale remuneration; the code drew on Soumith Chintala’s DCGAN implementation [5].Source: §1 · pp. 1, 3
- Computer art’s acknowledged pioneers date to the early 1960s, while Cohen’s works typically sold for less than 1% of the Belamy price.Source: §1 · p. 3
1.1 Questions Raised
- The review poses five questions about authorship, developer rights, mimicry, deep-learning distinctiveness, and whether AI art is a new kind of art.Source: §1.1 · p. 4
- Questions about the auction price, creators’ motivations, and the work’s artistic quality fall outside the paper’s scope and require broader research.Source: §1.1 · p. 4
1.2 AI Art
- Here, “AI Art” means art made using any AI technique, though the paper notes there is no consensus on what the term includes.Source: §1.2 · p. 4
2 Autonomy, Authenticity, Authorship and Intention (AAAI)
The paper examines how autonomy, authenticity, authorship, and intention shape the aesthetic value of computer-generated art. It distinguishes machine self-organisation from human-like intentional autonomy and treats attribution as culturally and legally contingent.
2.1 Autonomy
- GAN autonomy is bounded by training-data statistics and does not establish intentional autonomy, though autonomous generative processes can yield more challenging and interesting art than using computers as mere tools.Such systems may enable expression unavailable through other human tools, but that does not give them human-like intentional autonomy.Source: §2.1 · p. 5, Autonomy of Deep Learning Systems · pp. 6–7
2.2 Authorship
- Authorship is culturally and legally contingent, so whether a computer can count as an author may be settled through cultural argument rather than objective or universal criteria.The paper uses Hobbes’s definition of personhood and the Namatjira case to illustrate how authorship and personhood reflect ambient power and culture.Source: Authorship and Individuality · pp. 7–8
- In contemporary Western contexts, authorship may be shared among artists, algorithm developers, training-data contributors, and sometimes algorithms, making accurate attribution important to authenticity.The paper also argues that software and datasets are attributable contributions, and that commercial gains may increase interest in legal standards.Source: Legal and Moral Ownership · pp. 9–10, Summary · p. 10
2.3 Authenticity
- Authenticity in computer art depends substantially on the generative process: derivative or memetic processes tend to produce generic work and weaken claims to artistic integrity.The authors urge practitioners to disclose process and correct misconceptions that current AI art is intelligent or creative in the human sense.Source: §2.3 · p. 11, Truth about Process · pp. 11–12
2.4 Intention
- Claims that AI systems intend to make art are unsupported here: programmed systems must generate, while human-like intention would include the choice not to make art.The authors suggest that the shift from “generative” to “artificial intelligence” terminology may affect perceptions of intention.Source: §2.4 · p. 12
3 Questions Answered
The answers distribute authorship across contributors, require accurate process attribution, and characterize current trained systems as mimicking human creative artefacts. They also find no significant GAN-specific novelty in many works and argue technical differences alone rarely make a new kind of art.
3.1 Who is the author of a work of art made by an AI algorithm?
- The software creator and person who trained and modified its parameters can both be authors; artists whose works feature in training data may also count.AI systems are not broadly accepted as authors by artistic or general-public communities.Source: §3.1 · p. 12
3.2 What attribution or rights should be given to developers of an “art” generating algorithm?
- Authors should accurately represent the generation process, including machine and human labour, while maintaining all authors’ moral right to integrity.Open-source software can reduce or negate legal disclosure responsibilities.Source: §3.2 · p. 13
3.3 To what extent is any AI artwork relying on mimicry of human creative artefacts?
- Systems trained on curated datasets of human-produced content rely on mimicry rather than autonomously searching for novel expression.Source: §3.3 · p. 13
- Many GAN artworks introduce no significant new process or artefact aspects compared with established machine-learning art systems.Media, auction houses and designers frame GANs as AI, a distinction the authors attribute more to terminology and marketing; Sheridan earlier described “Copy Art” exploiting a marketable, recognizable product.Source: §3.4 · p. 13
3.5 Do current artworks made by AI represent a new “kind” of art?
- GAN art is probably not a new kind in any major way; technical differences rarely suffice, while cultural and social factors typically matter more.Computer-generated art is more than 50 years old; the authors leave judgement on Belamy’s merits and Obvious and Christie’s claims to the reader.Source: §3.5 · pp. 13–14