Introduction

Artists are among the many professions that are embracing artificial intelligence. AI is beginning to undertake tedious repetitive work, without replacing human’s creativity (Abramovich 2018). Recently, the art market has displayed an interest in purchasing AI art. Christies sold Edmond de Belamy, an algorithm-generated painting, and Sotheby’s sold Memories of Passersby I, an AI video installation, at $432,500 and $51,012, respectively (Moxley 2019). With a rapidly increasing variety of AI art being created and transacted, the need to preserve such art in differentiated ways is emerging. 

Preservation vs. Conservation: Challenges in the Language

While people sometimes use the words ‘preserve’ and ‘conserve’ when it comes to protecting, maintaining, or storing artworks, they are not interchangeable. That is because born-digital components of AI art cannot undergo conservation in a narrow sense – interventive conservation involving hands-on treatment directly to an object to stabilize its condition (Kroslowitz n.d.). Unlike physical art pieces, digital objects are intangible, and are likely to get altered for future use (Baucom 2019). Subject to changes of available technology, they require digital preservation  by means of new formats or hardware that seek to maintain an output of authentic information (Thibodeau 2002). Therefore, preservation can better capture the sense of change tolerance when it comes to AI art as digital art. 

While ‘conservation’ is still often used for digital artwork, especially works that have been damaged, obsolete electronics that need physical treatment hardware for AI art is new and more suitable for preventive efforts (Guerrieri 2019). Given that Kroslowitz (n.d.) defines the prevention work as preservation in contrast to direct invention, discussing how to ‘preserve’ AI art is better in understanding the practices used today.

Nonetheless, there are AI artworks that have been finalized and collected as static pieces, for which traditional methods of conservation are more applicable, while their digital versions can still be preserved.

How to categorize AI art 

While using genres as criteria for categorizing traditional art, a different logic needs to be used in categorizing AI art. As Carla Dobronauteanu pointed out, AI art can be classified as AI-assisted or AI-powered based on whether AI is a tool for artists to produce a final piece or  rather part of dynamic work. In this sense, Edmond de Belamy--sold in the form of a print on canvas--is a typical AI-assisted piece, only involving AI prior to the painting’s debut. Memories of Passersby I then belongs to the latter category, as AI continually powers a never-ending, never repeating stream of human portraits(Onkaos 2018). 

What underlies the ongoing task by AI is the work’s characteristic of having time duration in addition to its being physically visible. AI-powered art thus falls into the genre of time-based media art, which depend on technology and have a durational dimension (TATE n.d.). Also included in this genre are videos, sound artworks, film, slide-based installations, software-based art and other forms of technology-based artworks (Dover 2014). It can be helpful to consider AI-powered art under the framework of time-based art because it focuses the conservation of digital art within constant replacement of ever-changing technology.  

It is important to note that not all AI-assisted artwork must have temporal attributes. Memo Akten’s Deep Meditation, a one-hour long film generated by AI but played without further involvement of AI, is an example of time-based AI-assisted work. Therefore, one way to avoid unclarity and overlaps when discussing these concepts is to use both the timeline of AI’s existence in an art piece and the work’s temporal length as criteria. It can thus be summarized into three categories of AI art: non-time-based (or static) AI-assisted art, time-based AI-assisted art, and AI-powered art. These categories lay a foundation for analyzing preservation challenges.

Preserving static AI-assisted art

Since Edmond de Belamy was printed out and sold as a single physical object by the French collective Obvious, to preserve it becomes a task in the traditional realm of conservation and the responsibility of the print’s new collector. But it is more complicated to determine who should instruct the preservation for an AI artwork’s digital counterpart, because various parties may have contributed to producing it apart from those who finalized the work. For example, while Obvious trained an algorithm with large datasets to generate Edmond de Belamy, the algorithm was created by Robbie Barret before Obvious retrieved it from an open-source platform (Chang 2019). Anyone intending to preserve the AI along with the painting’s digital version, needs instruction from both the trainer and the creator of the AI to fully understand the work’s mechanism and nature. However, it is possible that a collector knows little about the original creators--like Robbie Barrat, who did not receive any formal acknowledgement--or intention contained in an algorithm for artistic use (Chang 2019). The collector may thus lose part of the authentic information. 

While some expect Blockchain to ensure authentic digital artworks as it will add a unique, unerasable history, this will only ensure digital objects with respect to anti-counterfeiting (Moxley 2019). Collectors may still fail to seek instructions from AI artists or creators of the algorithm shared in open-source domains if Blockchain only records legal sellers of the work. Therefore, additional measures need to be taken when dealing with AI generated digital pieces. 

 Preserving time-based AI-assisted art

Unlike static works, time-based AI-assisted art requires that collectors keep digital files, software, and hardware for playing the files. The project files are not needed, because the artist exports and provides only the final work. These digital objects, similar to common time-based artworks, can be obsolete, unreadable, or unrestorable when the media technology carrying them exit the market and new platforms do not support previous file formats (Guerrieri 2019). Renewal of the formats or the platforms is thus necessary. However, the cost can sometimes be higher than the value of the art piece, thereby intimidating owners or potential buyers and potentially leading to loss of these artworks altogether (Sabah 2019).

Preserving time-based, AI-powered art

Time-based AI-powered art takes place in real-time, as AI programs are continually running to present dynamic effects based on ongoing calculation. The data input can be preset, such as what was used by the machine of Memories of Passersby I. Similarly, Refik Anadol’s pre-collected data from the public domain for his interactive architectureMachine Hallucination. The scale of the dataset was significant with 300 million photos and 113 million other raw data points (Haigney 2019). Metadata and the threshold of computer capacity results in high fixed costs and barriers to preserving AI-powered artwork like Machine Hallucination. This likely contributes to the fact that the installation is in the for-profit space Artechouse as a permanent exhibition. Apart from the challenge of data volume, software and hardware’s capability of sensing and analyzing instant data is also essential, especially for artworks using a live feed. An example of this is the interactive installation Learning to See by Memo Akten. The machine-learning software acquired data input from a camera shooting a table with objects continually and randomly rearranged by viewers, before reinterpreting and reflecting the changes by projecting them on a wall to achieve a real-time effect (Goodman2020). The responsiveness of technology infrastructure, including computers and display equipment, is the basis for presenting the dynamics of such live artworks smoothly, thus proving that responsiveness is essential for the preservation. 

Conclusion

AI-based art involves different levels of artists’ creativity and curation based on what data sources are input, what role AI will take in an art piece, and what form artists will showcase a work. Utilizing AI for art making is a part of the human phenomena and efforts of innovating artistic practices by manipulating cutting-edge tools, and it is those efforts that are worth discussing, and comparing and differentiating. During the process of handling an AI artwork’s complex apparatus or temporal elements, museums, galleries and collectors should also advance their existing system of digital preservation, collection management or conservation. Rather than a marketing gimmick, AI art is an opportunity for organizations or collectors to welcome changes and break through their standards of practice, which is one of the reasons why AI-based artworks can be valuable. For the general process of creating and exchanging AI artworks, see the infographics below.

Resources

Abramovich, Giselle. “Technology and Creativity Go Hand In Hand: Study.” Accessed March 6, 2020. https://cmo.adobe.com/articles/2018/10/adobe-pfeiffer-ai-creativity-study.html#gs.yn3f73.Baucom, Erin. “Preservation as Risk Management.” American Libraries, September 3, 2019.

https://americanlibrariesmagazine.org/2019/09/03/preservation-risk-management/.Chang, Vanessa. “Ghost Hands, Player Pianos, and the Hidden History of AI.” Los Angeles Review of Books, October 5, 2019. https://lareviewofbooks.org/article/ghost-hands-player pianos-and-the-hidden-history-of-ai/.

Dobronauteanu, Carla. “Research Update: Artworks Powered by Artificial Intelligence in the Marketplace.” Arts Management and Technology Laboratory, November 28, 2018. https://amt-lab.org/blog/2018/11/research-update-artworks-powered-by-artificial-intelligence-in-the-marketplace?rq=AI%20generated%20art.

Dover, Caitlin. “What Is “Time-Based Media”?: A Q&A with Guggenheim Conservator Joanna Phillips.” Accessed March 6, 2020. https://www.guggenheim.org/blogs/checklist/what-is-time-based-media-a-q-and-a-with-guggenheim-conservator-joanna-phillips.

Goodman, Matthew. “How Does AI change the Way we Perceive Art.” Arts in America, January 21, 2020. https://www.artnews.com/art-in-america/features/ian-cheng-simulation-ai-art-1202675838/. 

Google. “Revisiting Unreasonable Effectiveness of Data in Deep Learning Era.” Accessed March 7, 2020. https://arxiv.org/pdf/1707.02968.pdf.

Guerrieri, Vince. “The Secret Tech Problem at Modern Art Museums.” Popular Mechanics, June 24, 2019. https://www.popularmechanics.com/technology/design/a28168772/techproblem-art-museums/.

Haigney, Sophie. “Refik Anadol Trains AI to Dream of New York City.” Art in America, September 18, 2019. https://www.artnews.com/art-in-america/features/refik-anadol-machine-hallucination-artechouse-shows-how-ai-dreams-60204/.

Kroslowitz, Karen. “Preservation, Conservation, Restoration: What’s the Difference?” The Computer History Museum Blog, October 26, 2012. https://computerhistory.org/blog/preservation-conservation-restoration-whats-the-difference/. 

Moxley, Mitch. “The AI Art Boom.” Barrons, September 24, 2019. https://www.barrons.com/articles/the-ai-art-boom-51569348001.

Onkaos. “Memories of Passersby I by Mario Klingemann.” Accessed March 7, 2019. https://vimeo.com/298000366. 

Daily Sabah. “Sakıp Sabancı Museum offers panels on protection of digital arts.” Daily Sabah, December 1, 2019. https://www.dailysabah.com/arts-culture/2019/12/01/sakipsabanci-museum-offers-panels-on-protection-of-digital-arts.

TATE. “Time-Based Media.” Accessed March 7, 2020. https://www.tate.org.uk/art/art-terms/t/time-based-media.

Thibodeau, Kenneth. “Overview of Technological Approaches to Digital Preservation and Challenges in Coming Years.” The Council on Library and Information Resources. Accessed April 5, 2020. https://www.clir.org/pubs/reports/pub107/thibodeau/.