Ekaterina Kolesnikova

Institute of Culture and Arts,

Moscow City University

Moscow, Russia

Abstract: This article examines artificial intelligence as the latest stage in the long history of interaction between artistic creation and technology. It traces major technological thresholds - from prehistoric pigment and classical metallurgy to linear perspective, print, photography, cinema, computer art, and generative machine learning and argues that each threshold redistributed creative labor rather than simply replacing the artist. The article distinguishes three contemporary models of human-AI practice: AI as an extended instrument, AI as a collaborative system, and AI as a putatively autonomous agent. It then considers the implications of these models for authorship, copyright, training data, aesthetic homogenization, cultural diversity, deepfakes, and artistic education. Drawing on recent English-language research and policy documents, including work by Aaron Hertzmann, Marian Mazzone and Ahmed Elgammal, Anil Doshi and Oliver Hauser, UNESCO, the U.S. Copyright Office, and the European Union, the article proposes a human-centered cultural framework. Its central claim is that the future of art will depend not on a simple opposition between human beings and machines, but on transparent, critically directed, and institutionally accountable forms of hybrid creativity.

Keywords: artificial intelligence; generative art; artistic authorship; digital culture; art education; copyright; cultural diversity

Art at the Threshold of a New Paradigm

Art is one of the fundamental forms through which human beings interpret experience, construct collective memory, and communicate values that cannot be reduced to purely instrumental information. It has therefore always responded with sensitivity to technological and social change. The discovery of mineral pigments, the development of bronze casting, the codification of linear perspective, the printing press, photography, recorded sound, cinema, and digital media did more than supply artists with new tools. Each innovation altered the conditions under which images and narratives could be made, circulated, authenticated, owned, and interpreted.

Generative artificial intelligence represents the newest and perhaps the most controversial stage of this history. Systems capable of producing images, music, video, and text from natural-language instructions have moved rapidly from specialist laboratories into ordinary creative practice. They are now used by illustrators, photographers, designers, game developers, filmmakers, architects, musicians, advertisers, educators, and nonprofessional users. Their spread has produced polarized interpretations: predictions of a new renaissance coexist with warnings about the “death of the author,” mass unemployment, plagiarism, cultural standardization, and the erosion of artistic skill.

The debate is frequently distorted by two symmetrical errors. Technological enthusiasm treats the model as an autonomous intelligence whose outputs are equivalent to human creation. Technological pessimism treats the same model as a machine that can only steal, imitate, and destroy. Both positions simplify the actual creative process. Generative AI is neither an ordinary neutral instrument nor a conscious artistic subject. It is a complex technical medium whose cultural meaning depends on training data, interface design, institutional ownership, human direction, and the social context in which outputs are selected and presented.

The purpose of this article is to place AI within the longer history of art and technology, examine how it is transforming artistic labor, and identify the educational, ethical, and policy measures required to ensure that technological innovation strengthens rather than impoverishes human culture.

Historical Context: Technological Thresholds in the Evolution of Art

The earliest surviving works of visual culture were inseparable from material knowledge. Cave paintings at Chauvet, Altamira, and other Paleolithic sites were produced through the controlled use of charcoal, ochre, mineral color, engraving, stencils, and the irregular topography of rock surfaces. These works demonstrate that artistic representation was already technological: it required the selection and transformation of matter, the management of light and access, and an understanding of how a surface could activate the movement of an animal body. Chauvet, whose decorated areas date back more than thirty thousand years, is especially important because its images combine observation, abstraction, repetition, and spatial staging at an unexpectedly early date.

In ancient societies, developments in metallurgy, quarrying, engineering, ceramics, and architectural measurement expanded the scale and permanence of artistic production. Greek bronze casting made possible dynamic sculptural poses that were difficult to achieve in stone, while Roman concrete transformed architectural space. The distinction between “artist” and “craftsperson” was far less rigid than it would become in modern Europe: technical mastery, workshop organization, patronage, ritual function, and aesthetic value formed a single cultural system.

The Renaissance introduced not only new materials but also new technologies of seeing and knowing. Linear perspective, theorized by Leon Battista Alberti and associated with the experiments of Filippo Brunelleschi, converted pictorial space into a measurable geometric construction. The technique was not a transparent reproduction of vision; it was a cultural model that organized the relation between spectator, picture plane, and represented world. Anatomical study, optics, chiaroscuro, oil painting, and new workshop practices similarly contributed to the period’s expanded realism and expressive range.

The printing press transformed artistic communication by multiplying texts and images, standardizing formats, and enabling knowledge to move across regions with unprecedented speed. Woodcuts and engravings circulated compositions beyond the physical location of an original object. The authority of the unique artifact began to coexist with the cultural power of the reproducible image. This tension would become central to modern debates about photography, mechanical reproduction, and digital copying.

Before the nineteenth century, the visible world was represented primarily through the trained hand, even when artists used optical aids. Photography introduced a different chain of production: light itself could inscribe an image through a chemical and mechanical process. The new medium immediately generated anxiety. If a machine could reproduce appearances with greater speed and precision than a painter, would painting lose its social purpose? In practice, photography did displace some forms of commissioned portraiture and documentary representation. Yet it also democratized access to images, created new professions, transformed journalism and science, and encouraged painting to pursue color, movement, abstraction, and the materiality of the surface.

Walter Benjamin later argued that mechanical reproduction altered the “aura” of the artwork by changing its relation to uniqueness, ritual, exhibition, and mass spectatorship. Whether or not one accepts every part of his argument, the historical pattern is clear: reproduction technologies do not merely copy art; they reorganize the institutions and habits through which art is experienced (Walter Benjamin, 1968).

Cinema, recorded sound, radio, and television extended this reorganization. Artistic production became increasingly collective and industrial. A film is not the expression of a single hand but the coordinated result of scripts, performance, cinematography, editing, sound, set construction, finance, and distribution. The romantic image of the solitary artist therefore coexisted with large-scale systems of technological collaboration long before the arrival of AI.

From the mid-twentieth century onward, artists used computers to generate patterns, drawings, animations, and interactive systems. Early computer art often made its procedures visible: the artist designed rules, parameters, or chance operations, and the machine executed them. Harold Cohen’s AARON, Vera Molnár’s algorithmic drawings, Frieder Nake’s plotter works, and Lillian Schwartz’s computer graphics demonstrated that automation could itself become an artistic subject.

Aaron Hertzmann notes that earlier technologies associated with automation—including photography and computer animation—initially appeared to threaten artists but ultimately created new expressive forms and professions. This historical analogy is useful, although generative AI introduces an important difference. Traditional tools execute relatively stable operations; machine-learning systems infer patterns from large datasets and can return outputs that were not explicitly specified by the user. The artist therefore works not only with a device but with a probabilistic system shaped by millions or billions of prior cultural examples (Aaron Hertzmann, 2018).

The expression “generative AI” refers to machine-learning systems designed to produce new content by modeling statistical regularities in training data. Generative adversarial networks (GANs), introduced in 2014, trained a generator and discriminator in competition and became influential in synthetic image production. More recent diffusion models learn to reverse a process of noise addition, gradually constructing images from probabilistic representations. Large language and multimodal models extend this logic across text, image, audio, and video (Ian J. Goodfellow, 2014)

These systems do not possess intention, memory, embodiment, or cultural understanding in the human sense. They can model relationships among words, visual features, and styles with extraordinary sophistication, but their “knowledge” is operational and statistical. The tendency to describe a model as imagining, remembering, or wanting can be rhetorically convenient, yet it risks concealing the human and institutional structures behind the system: dataset selection, annotation, model architecture, safety filters, interface design, computational infrastructure, and commercial policy.

At the same time, it would be misleading to reduce generative AI to an ordinary brush. A brush does not reorganize itself through training data or produce complex alternatives in response to a linguistic prompt. Marian Mazzone and Ahmed Elgammal therefore propose understanding AI not only as a tool but also as a medium whose possibilities and constraints include code, data, model architecture, curation, and critical interpretation. This formulation helps explain why AI art requires both technical literacy and aesthetic judgment (Marian Mazzone and Ahmed Elgammal, 2019).

Three Models of Human - AI Artistic Practice

The most common model treats AI as an extended instrument within a human-directed workflow. An illustrator may generate visual alternatives, a filmmaker may create storyboards or backgrounds, a photographer may remove distractions or reconstruct missing detail, and a designer may test compositions before producing a final version. In such cases, generation is only one stage. The artist defines the problem, writes and revises prompts, provides reference images, selects among alternatives, edits the result, combines it with other materials, and determines its context of presentation.

The widespread claim that “the prompt is the artwork” is therefore too narrow. Prompting can be an important creative skill, but authorship may also reside in the construction of a dataset, the sequencing of iterations, the rejection of predictable outputs, the integration of handmade material, the manipulation of individual elements, and the conceptual framing of the final work. The decisive issue is not whether AI was used, but what kinds of human decisions shaped the expressive form.

A second model understands the process as hybrid collaboration. Artists may train or fine-tune models on their own archives, family photographs, drawings, texts, field recordings, or scientific data. The system produces variations that the artist did not fully anticipate; the artist interprets those variations, adjusts the parameters, and returns them to the model. The process becomes iterative and dialogical, although the “dialogue” remains asymmetric because only the human participant possesses intention and responsibility.

Refik Anadol’s large-scale data installations and Mario Klingemann’s experiments with neural portraiture illustrate different versions of this approach. In these practices, the artist’s contribution includes the choice of data, the design of the environment, the orchestration of movement and sound, and the decision to present model behavior as an aesthetic event. The generated image is not necessarily the whole artwork; the artwork may be the complete system through which data, architecture, spectatorship, and computation interact.

The third and most controversial model minimizes visible human intervention and presents the AI system as an autonomous creative agent. Projects such as Creative Adversarial Networks attempted to model novelty by generating images that remain recognizable as art while deviating from established stylistic categories. Such experiments are intellectually significant because they test computational definitions of creativity. Yet they do not prove that the system has acquired artistic consciousness (Ahmed Elgammal, 2017).

Margaret Boden distinguishes combinational, exploratory, and transformational creativity. Computational systems can combine known elements and explore a rule-governed space with remarkable effectiveness; whether they can independently transform the values that define that space is more difficult to establish (Ahmed Elgammal, 2017).

Hertzmann adds a social criterion: art is not merely an object with certain visual properties, but an action attributed to a social agent within a community. On this account, current AI systems can participate in art-making without themselves becoming artists in the full cultural sense.

The effect of AI differs across creative professions. Commercial illustration, stock imagery, advertising mock-ups, translation, basic copywriting, and some forms of concept design are especially exposed because clients often value speed, volume, and price over singular authorship. In these sectors, AI may reduce the number of entry-level assignments through which young professionals traditionally developed experience. The threat is therefore not simply the disappearance of an entire profession but the erosion of stages within a professional career.

Other fields may experience augmentation rather than direct replacement. A theatre designer can test scenic alternatives before construction; a composer can explore orchestration; a museum can reconstruct damaged artifacts; an independent filmmaker can visualize sequences that previously required a large studio. These uses lower technical barriers and may increase access to production. Yet lower barriers do not automatically produce cultural equality. Access to high-quality models, computing power, proprietary datasets, and distribution platforms remains uneven.

The most plausible near-term outcome is a redistribution of value. Routine execution becomes cheaper, while concept development, distinctive style, art direction, editing, verification, performance, and trusted authorship become more important. New roles emerge around model supervision, dataset curation, AI-assisted postproduction, provenance, rights clearance, and the evaluation of synthetic media. The artist of the future may need to be simultaneously a maker, director, researcher, editor, and systems critic.

Experimental evidence suggests both gains and risks. In a 2024 study, Anil Doshi and Oliver Hauser found that access to generative-AI ideas improved the judged creativity and quality of short stories, particularly for participants with lower baseline creativity. At the same time, AI-assisted stories became more like one another, creating a tension between individual improvement and collective diversity. This result is highly relevant to visual culture: a tool can help many individuals produce more polished work while making the total cultural field less varied (Anil R. Doshi and Oliver P. Hauser, 2024).

Authorship is both a philosophical and a legal category. Philosophically, it concerns intention, responsibility, and the relation between a work and a creative subject. Legally, it determines ownership, registration, licensing, and remedies. These questions should not be collapsed into a single argument about whether an output “looks creative.”

The U.S. Copyright Office’s 2025 report on copyrightability concludes that copyright does not extend to purely AI-generated material or to material over which a human lacks sufficient control. It also states that prompts alone, given the operation of currently available systems, generally do not provide enough control to make the user the author of the output. Human-authored selection, coordination, arrangement, and substantial modification may nevertheless be protected on a case-by-case basis. This approach reinforces a process-based understanding of authorship: legal protection follows identifiable human expression rather than the mere act of initiating generation.

For artists, the practical implication is clear. They should preserve evidence of their process: drafts, prompt histories, reference materials, compositing files, edits, annotations, and explanations of which expressive elements were created or controlled by a human. Such documentation can support not only legal claims but also scholarly transparency and audience trust.

A different question concerns the copyrighted works used to train generative models. The ethical problem arises because large datasets may include images, books, recordings, and other works gathered without meaningful consent, attribution, or compensation. Artists object not only to literal copying but also to the extraction of recognizable stylistic and professional value from entire bodies of work.

The legal status of training varies by jurisdiction and remains unsettled. The U.S. Copyright Office’s 2025 pre-publication report on generative-AI training emphasizes that fair-use analysis is fact-specific: some training uses may qualify as fair use and others may not, depending on the works, their sources, the purpose of the use, safeguards on outputs, and market effects. Consequently, confident universal claims that all training is either lawful or unlawful are premature.

Ethically responsible systems should move toward greater dataset transparency, meaningful opt-out mechanisms, licensing pathways, and compensation models appropriate to different sectors. Collective licensing may be practical for some repertoires, while direct agreements, public-domain collections, or creator-controlled datasets may be preferable in others. The central principle is that culture should not be treated as a free raw material whose contributors disappear from the technological value chain.

Training data do not represent the world neutrally. They reflect historical patterns of visibility, inequality, language dominance, colonial classification, commercial popularity, and platform moderation. Kate Crawford and Trevor Paglen’s influential critique of training sets showed how apparently technical categories can encode political judgments about bodies and identities. Although specific datasets and claims require careful examination, the broader lesson remains valid: curating data is a cultural act (Kate Crawford and Trevor Paglen, 2019).

Generative systems can reproduce stereotyped professions, beauty standards, gender roles, racial hierarchies, architectural clichés, and geographically narrow ideas of “traditional” culture. Correcting these effects requires more than adding diverse images. It requires documentation of provenance, participation by represented communities, evaluation in multiple languages and cultural contexts, and mechanisms through which harms can be challenged.

Photography once derived much of its authority from the causal relation between light and a photographed scene. Digital editing had already weakened any simple equation between photograph and truth, but generative AI intensifies the problem by producing photorealistic images and voices without a corresponding event. The result is not merely an artistic issue. Synthetic media can affect journalism, historical memory, elections, reputation, and evidence.

The European Union’s Artificial Intelligence Act responds through transparency obligations. Article 50 requires providers of certain generative systems to make synthetic outputs detectable in machine-readable form and requires deployers of deepfake systems to disclose that image, audio, or video content has been artificially generated or manipulated. Artistic and fictional works receive a more limited disclosure obligation intended not to obstruct normal presentation or enjoyment.

Disclosure alone is not sufficient, because labels can be removed and metadata can be lost. Technical provenance standards such as the Coalition for Content Provenance and Authenticity (C2PA) allow creators and publishers to attach cryptographically verifiable information about origin and editing history. Such systems do not prove that depicted events are true, and they cannot identify every unlabeled fake. Their value lies in providing a positive chain of provenance for participating media.

Cultural institutions should therefore combine visible labeling, machine-readable credentials, archival preservation of source files, editorial review, and public media literacy. Authenticity in the AI era will increasingly depend on documented processes rather than on visual appearance alone.

Generative models are optimized to identify and reproduce patterns that occur frequently in their training distributions. This makes them powerful but also predisposes them toward statistical centrality. Popular visual conventions, cinematic lighting, familiar character types, and globally dominant design languages may be easier to generate than local, hybrid, or historically marginal forms. When millions of users rely on similar models and prompt conventions, cultural production can drift toward an immediately recognizable “AI look.”

Homogenization is not inevitable. Artists can build specialized datasets, introduce non-digital materials, use local archives, deliberately preserve errors, and combine systems with embodied performance or site-specific practice. Nevertheless, the economic structure of AI matters. UNESCO’s 2025 report on artificial intelligence and culture warns that concentration of AI infrastructure among a small number of corporate actors limits the ability of diverse cultural stakeholders to participate in development and governance. It calls attention to equitable access, public infrastructure, training opportunities, cultural rights, and benefit sharing.

Cultural diversity therefore depends on more than diverse outputs. It requires diversity in who designs systems, who controls datasets, which languages receive technical support, which communities can afford advanced tools, and whose standards determine whether a result is judged useful or beautiful. Without such structural pluralism, AI may increase the quantity of cultural content while narrowing the range of cultural imagination.

Art education faces a fundamental question: if a machine can simulate drawing, color harmony, stylistic variation, or compositional polish, what should an artist learn? The wrong response is to abandon foundational skills. Drawing, observation, material practice, movement, sound, and craft train attention and judgment; they are not merely inefficient routes to a final image. The equally wrong response is to exclude AI and leave students unprepared for the conditions of contemporary production.

A renewed curriculum should integrate six areas of competence:

1. Conceptualization and research. Students must learn to formulate meaningful questions, situate projects historically, construct narratives, and connect form to social experience. In a field of abundant images, the quality of the problem becomes more important than the speed of execution.

2. Material and embodied practice. Hands-on work with drawing, painting, sculpture, photography, performance, sound, and installation should remain central. Embodied skill provides sensory knowledge and gives artists alternatives to platform-dependent production.

3. Algorithmic and data literacy. Students do not all need to become programmers, but they should understand training data, model limitations, probabilistic generation, bias, versioning, privacy, and the difference between model output and factual evidence.

4. Critical curation and editing. The ability to reject, compare, sequence, revise, and contextualize outputs is a core artistic skill. Students should be assessed on their decision-making, not merely on the visual polish of a generated result.

5. Ethics, copyright, and provenance. Courses should address consent, attribution, disclosure, dataset rights, contractual terms, synthetic likenesses, and documentation of human contribution. Students should learn to use provenance tools and to describe their AI involvement accurately.

6. Development of a distinctive human voice. The highest educational priority is the formation of an artistic position grounded in lived experience, cultural knowledge, ethical responsibility, and sustained practice. AI can imitate many surface features, but it cannot substitute for the student’s relationship to a community, body, place, memory, or historical situation.

Assessment methods must also change. Instead of grading only the final artifact, instructors can require process portfolios, prompt and revision logs, source statements, oral defenses, comparative analysis of rejected outputs, and reflective accounts of where human judgment entered the work. Such methods reward learning and authorship rather than effortless appearance.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence specifically calls for AI education and digital training for artists and creative professionals, while emphasizing cultural diversity, artistic freedom, intellectual-property research, critical thinking, and the relational value of education.

UNESCO’s guidance on generative AI likewise advocates a human-centered approach in which systems support rather than replace human agency and where educational use is co-designed, evaluated, and governed.

The integration of AI into art cannot be governed by individual choices alone. Museums, universities, publishers, festivals, platforms, funding bodies, and governments all shape incentives and norms. A responsible cultural ecosystem should include the following measures:

· Transparent disclosure. Institutions should require clear statements describing whether and how generative AI was used, while distinguishing assistive editing from substantial generation.

· Process documentation and provenance. Creators and institutions should preserve source materials, model information, version histories, and editing records, and adopt interoperable provenance standards where appropriate.

· Rights-respecting data practices. Public and private systems should develop licensing, opt-out, attribution, and compensation mechanisms appropriate to the affected creative sector.

· Support for independent and local creators. Public investment should expand access to open tools, multilingual resources, cultural archives, and training outside the largest commercial platforms.

· Human review and accountability. Decisions about publication, historical representation, cultural heritage, and sensitive identities should remain subject to identifiable human responsibility.

· Protected spaces for non-AI practice. Cultural institutions should continue to support handmade, embodied, local, and slow forms of creation rather than treating technological adoption as an automatic criterion of innovation.

These measures do not imply hostility toward technology. On the contrary, they create the conditions under which innovation can be trusted. The strongest cultural response to AI is neither prohibition nor uncritical adoption, but institutional design that makes responsibility visible and distributes benefits fairly.

Conclusion

Artificial intelligence is not an apocalypse for art and not a universal solution to its problems. It is a powerful medium that accelerates some stages of creation, destabilizes others, and forces renewed attention to authorship, labor, authenticity, cultural rights, and education. Like photography and earlier forms of automation, it will eliminate some tasks, create new professions, and transform the meaning of established skills. Unlike a camera or a brush, however, it operates through vast cultural datasets and computational infrastructures whose ownership and design have direct political consequences.

The history of art suggests that artistic value has never depended solely on technical difficulty. A work matters because it organizes attention, communicates experience, transforms a convention, creates a social relation, or reveals something that could not otherwise be perceived. AI can contribute to these processes, but it cannot absolve human beings of intention and responsibility. Even when generation is automated, the cultural questions remain human: What is selected? What is excluded? Whose work trained the system? Who benefits? Who is represented? Who is accountable for the result?

The future of art therefore lies neither in defending an idealized past nor in surrendering creativity to automated production. It lies in a critically informed symbiosis in which human beings use computational systems to extend perception and experimentation while preserving cultural diversity, transparent authorship, and the freedom to create outside algorithmic norms. The art of the AI era has only begun, and its trajectory will be determined less by what the technology can.

References

1. Alberti, Leon Battista. On Painting. Translated by Cecil Grayson. London: Penguin Books, 1991.

2. Benjamin, Walter. “The Work of Art in the Age of Mechanical Reproduction.” In Illuminations, edited by Hannah Arendt, translated by Harry Zohn, 217–251. New York: Schocken Books, 1968.

3. Boden, Margaret A. The Creative Mind: Myths and Mechanisms. 2nd ed. London: Routledge, 2004.

4. Crawford, Kate, and Trevor Paglen. “Excavating AI: The Politics of Training Sets for Machine Learning.” September 19, 2019.

5. Doshi, Anil R., and Oliver P. Hauser. “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content.” Science Advances 10, no. 28 (2024): eadn5290.

6. Elgammal, Ahmed, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone. “CAN: Creative Adversarial Networks, Generating ‘Art’ by Learning about Styles and Deviating from Style Norms.” In Proceedings of the Eighth International Conference on Computational Creativity, edited by Ashok Goel, Anna Jordanous, and Alison Pease, 96–103. Atlanta, GA: Association for Computational Creativity, 2017.

7. Goodfellow, Ian J., Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. “Generative Adversarial Nets.” In Advances in Neural Information Processing Systems 27, 2672–2680. 2014.

8. Hertzmann, Aaron. “Can Computers Create Art?” Arts 7, no. 2 (2018): 18.

9. Manovich, Lev. The Language of New Media. Cambridge, MA: MIT Press, 2001.