AI-Generated Art by AI Art House

1. The organization

The company under consideration in this digital ethics consultation is AI Art House (https://aiarthouse.com/pages/about). AI Art House is an online gallery for art that is generated by artificially intelligent algorithms, mostly generative adversarial networks (GAN). They sell the digital artwork alongside free access to a physical copy of the artwork. The digital artworks (or rather the access to the original artwork) are sold as non-fungible tokens (NFTs), which are pieces of data that are not fungible – meaning there is nothing of the same value for it to be interchanged with. Through NFTs the AI Art House sells the digital artwork authentically via the immutability of the NFT-blockchain technology. What makes AI Art House special is the way they open up the AI art market for secondary owners: they run a marketplace for the resale of AI artworks and second-hand owners of the NFT also gain access to the production of physical copies of the artwork. Driven by the belief that AI art breaks up the traditional art market for bigger audiences to embrace aesthetic progress, AI Art House advocates the artistic value of the GAN-produced artworks, which, depending on the training data, vibe like Renaissance paintings, 19th-century realism or color variations on Rothko. Furthermore, AI Art House has set a limit on the number of NFTs they will ‘mint’: 10.001 pieces. This restriction on the sale of AI Art is their idea of ensuring the market of AI art is not overflooded and the minted digital pieces will retain their value. The owner of the NFT is also licensed both commercially and non-commercially for the further usage of the artwork.

2. The AI technologies Employed

Two separate technologies play are directly involved in the service of AI Art House: (i) non-fungible tokens (NFTs) and (ii) generative adversarial networks (GAN)s. While NFTs are not strictly instances of artificial intelligence, the technology behind NFTs can be automated by AI models and automation. GANs are considered AI themselves.

(i) The NFT, as stated above, is a completely unique piece of data that is assigned a value based upon the digital object, an artwork in this case, that it gives access to and authenticates.

But what is this piece of data exactly and how is it created? An NFT becomes unique once the original file of a digital object is subsumed into a blockchain, typically Ethereum (for the same fact the main currency for trading NFTs and AI art). A ‘blockchain’ is a term used for digital registries or ‘ledgers’ that cryptographically stores information about transactions and their timestamps in order for it to have a unique representation. To have something enter the blockchain, it needs to be validated, which is done by decentralized ‘miners’: suppliers of computational resources. Blockchain validation is a computationally demanding process and what is being validated are block order (parts of the sequence of ‘entries’ in the ledger) and state transitions (transmission of data from one owner to another).
The merit of blockchain technology for NFT is that in order to breach the chain, an enormous amount of computational power (hence energy resources) is needed, which serves as very strong digital security. At the same time, the merit is also the most stringent problem with blockchain: the computing power required for blockchain technology requires a lot of electricity and is thus responsible for huge amounts of carbon dioxide production. Returning to how AI Art House utilizes the blockchain, AI Art House uses the security guaranteed by the blockchain to protect NFTs uniqueness and authenticity.

(ii) GANs are the machine learning tool that is used to generate AI art from a training set of images and artworks.

The software AI Art House uses is GAN-based and the company generates artworks based upon existing art traditions. But what is a GAN, how does it work and why is it the current standard for computer generative art? While computer generative art dates to the distribution of computers to the public in the past half century, GANs are the state of the art image generation tool. For starters, the GAN is so important for AI art because after its creation by Goodfellow et al. in 2014, it was used by an Parisian art collective called Obsidian to generate a portrait of a fictitious character called ‘Edmond de Belamy’ in the synthesized style of portraits from the 14th to the 19th century. Controversially so, because the collective did not credit the work of other programmers they had built upon; something returned to in the ethical concerns.
A GAN is type of neural network that combines two networks: one that generates candidate outputs (‘generator’) and one that evaluates the outputs (‘discriminator’). The generator generates ‘fake’ images based upon random noise. This is paired with images from the training dataset and the discriminator has to decide which image is the real one. The relation between the discriminator and the generator is thus a form of game theoretic fooling: the GAN goes through cycles of presenting images until the discriminator is fooled by the generator. Technically, the likelihood of the discriminator being incorrect is maximized and this is used for performing gradient descent (a form of backpropagation) on the generator. The generator henceforth performs better in fooling the discriminator.
Based upon training data of historical paintings, the software of AI Art House is able to generate (or, perhaps more aptly put, extrapolate) paintings in historical styles that have never existed before. Little details are known however about the specs of the GAN used by AI Art House or the nature of their training sets. For training sets, databases of publicly accessible images of historical paintings, such as WikiArt, are commonly used. The lack of transparency about the code used also raises question about its originality, given the number of open source GANs, such as code books.

3. Ethical concerns

The main question the consideration of the initiative of AI Art House raises, is what the consequences if AI generated art are for the traditional art market and the aesthetic value and singularity of original works of art. These questions, while clearly interesting, fall outside the scope of this consultation, since we are concerned with ethical issues, and these ultimately boil down to problems in the domain of aesthetics. AI Art House state that they do not think the digital art market will replace the traditional art market, but that the consequence of the digital art market will a democratization of access to fine art.

While there are major ethical concerns with the AI Art House, residing both in the AI technology used and the practice of selling digital art itself, it is doing well in protecting ownership and combatting market flooding of AI generated art. AI Art House recognizes the new NFT-market for digital art is without regulations and could quickly be flooded by generated products. Therefore they restrict their initiative to 10.001 artworks, thereby placing a maximum on the number of artworks they as gallery will market. While it is unclear if this restriction will be of any help to prevent market flooding, given the unpredictability of market of the blockchain-currencies and the youth of NFTs. However it is clear that they are aware of this danger of market decentralization, an ethical concern returned to in what follows. The way AI Art House is conscious of the difficulties of ownership is that in purchasing the NFT, they are also granted (non-)commercial rights over the usage of the artwork. Consequently, the purchased artwork is more than a file; it becomes the owner’s trading material, just like the purchase of a physical artwork would. Next, we turn to the three ethical concerns, having to do with (1) sustainability; (2) ethics of authorship; (3) economic markets.

The carbon footprint of the digital art market.

The first concern with the blockchain technology is that it essentially is a polluting technology, as caused by the huge amounts of computational resources required for the data mining to add information to the data ledgers. A single Ethereum transaction is estimated to have a footprint of 35 kWh and ‘minting’ and NFT costs 100 kg of CO2, the equivalent of a 2 hour flight. And here we have only started to talk about minting, not even about bidding, finalizing sales etc. This is an AI-related concern, because the way AI is used to automate blockchain processes is part and parcel of making it computationally demanding. As Michelle Kasprzak has noted, the concept of a carbon footprint is difficult to determine and was developed by fossil fuel giant BP. That being said, the blockchain craze has led to the rise of globally decentralised data centers expanding large amounts of electricity on computational power for datamining. While the concept of an NFT itself is neutral with respect to carbon emissions – i.e. non-fungibility could be realized in all sorts of ways, both digital and physical - the technology it is created by is. AI Art House is conscious of this concern, but they shrug it off, stating it is a problem of blockchain technology and not necessarily of their endeavor. I disagree, because the blockchain is a necessary condition for the current digital art market infrastructure and the pollution caused by it cannot be avoided if the digital art market continues like this. The moral problem present here then is: is it ethical to start or participate in a new market knowing the negative consequences for the climate, while there already exists an art market and the goal of democratization of the art market could also be approached in non-polluting ways. At least AI Art House should take responsibility for the negative effects on climate changes of their operation. Feigning ignorance or blaming ‘the technology’ as if their operations are not part of the reason for its existence are hence options not to be preferred.

transparency concerning training data and secondary authorship.

The second concern has to do with the lack of transparency concerning the training data and the origin of the GAN-algorithm used by AI Art House. As becomes clear from the controversy surrounding the high-priced artwork of Edmond de Belamy, GANs are tools used by people to generate the types of artwork they desire. The training data used, essentially artworks authored by others, and the GAN algorithm, possibly based on third party work, are essential components of the creation of the artworks. As McCormack et al. note in relation to the question of who or what the authorship of computer generated art could be ascribed to, a multiplicity of actors could be considered as authors: the creator of the software, the person who trained the network and artists who prominently feature in the training data, but not the algorithm itself, as it cannot be considered a creative individual, but remains a tool in the hands of others. Consider for example the collection of AI art sold by AI Art House from the collection of Pierre-Auguste Renoir. Who is the author here? AI Art House does not tell us, but on the account of McCormack et al. AI Art House itself (trainers of the algorithm), the long-deceased Renoir (due to prominence in the training data) or the mysterious creators of the software, either AI Art House or an uncredited third party, could all be considered authors. If we apply another argument made by McCormack et al., in selling these products, possibly partially authored by others, AI Art House has the responsibility to be transparent about the process of creation and the attribution of rights regarding authorship, but currently they state no such specific information on the web-pages of any of the artworks they sell. This aspect of the digital art market is currently unregulated, but closely connects to intellectual property rights and regulations regarding creative commons license (open source software). Legally permissible or not, a moral case can be made for an honest presentation of the authorship of artworks.

The consequences of market decentralization.

The third concern is an economic one: the lack of regulation of the digital markets made possible by blockchain technology. The Ethereum blockchain the digital art market depends on is hailed as ‘decentralized finance’ because the Ethereum transactions are not mediated by institutions such as banks or national financial regulations. This lack of regulation has ethical effects, as Tim Dean argues: from a lack of protection of buyers against the stealing of NFTs and the taking advantage of buyers in the form of malevolent ‘smart’ contracts. As Jay Perlman notes, the discourse surrounding the decentralized NFT market reeks of libertarian fantasies, free of government interference. Furthermore, the unregulated digital market is in its infancy, enabling malevolent individuals to do so called ‘rug pulls’: massive scams that make a lot of people lose a lot of money, simply because of the speed of the market, the lack of protection and the shadiness of verification. That being said, decentralization is not immoral per se, but currently decentralization comes in the form of a Hobbesian ‘scam of all against all.’ Probably even hardcore libertarians would be displeased by this situation.

4. Recommendations

Transparency about the ecological footprint of AI artworks

Dealing with Ethereum’s emission problem is hard, since currently there are no viable alternatives for making assets non-fungible. There has been lively debate however on the climate impact of NFTs and as a result several currencies that have a smaller carbon footprint, such as Tezos, which is used by the popular NFT-platform ‘Hic et Nunc,’ were put into use. NFTs produced using less carbon dioxide are called ‘eco-NFTs’ or ‘Clean-NFTs.’ CleanNFTs use a technology call ‘Proof of Stake’ – making users of the blockchain prove they are holders of certain authentificatory tokens - a more eco-friendly alternative to the required extensive data mining of Ethereum and Bitcoin. Tezos is even more in its infancy than Ethereum, but there are already hundreds of platforms using Tezos or other CleanNFTs. Because all its transaction processes are embedded in Ethereum, AI Art House will not be able to move there operations from Ethereum to Tezos overnight and Tezos is even more unstable than the most popular cryptocurrencies, Bitcoin and Ethereum. But, in considering sustainability as a technomoral virtue, Ai Art House is obliged to modify its operations. Also, Tezos is not to be seen as the end goal of the development towards CleanNfts. It falls within AI Art House’s responsibility scope to look for even more ‘eco-friendly’ options to produce NFTs. In the meantime, AI Art House should inform its customers of the impact on the climate each artwork has, possibly in the form of the helpful comparisons produced by CryptoArt.wtf translating the computational power into examples like the amounts of time of individuals’ energy consumption, flying or usage of electronics. Lastly, if AI Art House is thinking on expanding their platform to AI-art generated by third parties, then they should curate the artists’ works that are filed for admission into the gallery for their carbon footprint.
To sum up, AI Art House has a few options to combat the negative climate effects of the NFT minting: (1) swap Ethereum for ‘cleaner’ NFT currencies such as Tezos and (2) informing their customers about the carbon footprint of a digital art work; (3) if admitted, curate the work of other artists for its carbon footprint, setting their own maximum on what is admissible into the gallery.

Involve algorithmic background history into the presentation of AI artwork and reward the (partial) authors accordingly.

To tackle the issue of transparency concerning training data and secondary authorship, only some simple adjustments have to be made to AI Art House current way of working. For each artwork and collection specific information has to be included about: (a) who is the creator of the software used; (b) who trained the network; (c) what training data were used; and, if applicable, (d) on what authors is the artwork based primarily (i.e. provide general information about Pierre-Auguste Renoir). What is important is that this information package needs to be publicly verifiable. This means that for the software, include links to either open sources, GitHub repositories or in case of patented algorithms general descriptions of the working of the GAN. Similarly, links to open source training data or descriptions of the training data should suffice. For the potential additional authors, the case is more difficult, because of the difference between living and dead authors. Living authors should be rewarded according to their share in the (conditions of possibility for) creation of the artwork. With dead authors, it depends on who holds the rights over the intellectual property over works.
To conclude, some physical artworks are sold with a frame that was made to fit by specialized frame makers. Consider this informational context the frame that comes with the digital work of art.

Include small-scale insurances for the secondhand sales on AI Art House’s platform.

The solution for the greed-driven problems that arise from the unregulated, decentralised market in NFTs is, unsurprisingly, to introduce regulations. The hard part of this easy-made suggestion is what aspects of the AI-art market need to be regulated in order for it to remain an economic success while becoming unharmful to participants. Thanks to its newfound prolificity on the Internet, the NFT market is dynamic and so decentralized oftentimes there are no platforms or institutions to mediate transactions. Therefore it is difficult for a non-expert to determine which laws have to be extended so as to accommodate for the digital art network. Thus I will not explicate here how European or U.S. laws should regulate this market. Instead, this recommendation will not be a solution for the problems the NFT-market as a whole faces, it is aimed at improving the services of AI Art House for the better.
As a platform, AI Art House could go further to protect their customers against scams etc. by providing a mini insurance company for its customers, protecting the purchased amount against malevolent individuals. Introducing for example measures like that of Marktplaats that ensure the asset is transferred to the buyer first before the selling party receives the money that is temporarily stored digitally in the platform itself. This is mainly a solution to protect the secondhand marketplace that AI Art House makes available. As for the direct sales of the artworks they themselves produce, AI Art House should be clear about how transactions are protected and what rights customers have to potentially exercise against them.

Sources:

-AI Art House. “What Are NFTs? (How AI Art Is Involved).” Published on 20/03/2021. Accessed on 27/06/2022. https://aiarthouse.com/blogs/articles/what-are-nfts.
-Akten, Memo. “CryptoArt.wtf.” Accessed on 26/06/2022. https://cryptoart.wtf/.
-CleanNFTS. “Home page.” Accessed on 26/06/2022. https://cleannfts.org/.
-Dean, Tim. “Hype and hypocrisy: the high ethical cost of NFTs.” Published on 24/05/2022. Accessed on 25/06/2022. https://ethics.org.au/hype-and-hypocrisy-the-high-ethical-cost-of-nfts/.
-Epstein, Ziv, Sydney Levine, David G. Rand and Iyad Rahwan. “Who Gets Credit for AI-generated Art?” iScience 23, no.9 (September 2020): 101515. https://doi.org/10.1016/j.isci.2020.101515.
-Ethereumprice. “What is Ethereum?” Last updated on 19/06/2020. Accessed on 27/06/2022. https://ethereumprice.org/guides/article/what-is-ethereum/.
-Gandhi, Rohith. “Generative Adversarial Networks — Explained.” Published on 24/05/2022. Accessed on 26/06/2022. https://towardsdatascience.com/generative-adversarial-networks-explained-34472718707a.
-Kasprzak, Michelle. “Ethical Engagement with NFTs – Impossibility or Viable Aspiration?” Makery. Published on 31/03/2022. Accessed on 25/06/2022. https://www.makery.info/en/2022/03/31/english-ethical-engagement-with-nfts-impossibility-or-viable-aspiration/.
-McCormack, Jon, Toby Gifford, and Patrick Hutchings. “Autonomy, Authenticity, Authorship and Intention in Computer Generated Art.” In Proceedings of Computational Intelligence in Music, Sound, Art and Design 8th International Conference, EvoMUSART 2019, edited by Anikó Ekárt, Antonios Liapis and María Luz Castro Pena, 35-50. Cham: Springer, 2019. https://doi.org/10.1007/978-3-030-16667-0.
-Perlman, Jay G.. “The Moral Complexity of Ethics in NFTs.” Bueno. Published on 10/06/2022. Accessed on 25/06/2022. https://www.bueno.art/blog/ethics-nfts.
-WikiArt. “Home page.” Accessed on 26/06/2022. https://www.wikiart.org/
-Wikipedia. “Edmond de Belamy.” Last updated on 25/02/2022. Accessed on 27/06/2022. https://en.wikipedia.org/wiki/Edmond_de_Belamy.