The Consequences of Generative AI Art
Not limited to intellectual property rights, there are also environmental effects.
1) Artificial, yes, but is it really intelligent?
2) But is it art?
Both questions miss the point.
Sketch for a video; generative art
Some artists and musicians, mostly amateurs, express their despair facing the realisation that the significant work they have invested in producing a song or an image has become pointless now that generative AI produces virtually the same results without effort. My impression is that most professional artists still do not feel that AI is about to replace them. Some are figuring out ways to incorporate AI as productive tools. Generative AI, as it exists today, produces digital content whole cloth, with their proprietary software offering no opportunity for users to tinker with the model itself. They offer consumption of products rather than tools for genuine creation. There appears to be little scope for the user to advance and become a virtuoso, the skills are absorbed into the software. That is also an argument in favour of maintaining classical artisanal skills, such as drawing or playing an instrument. Marshal McLuhan used to make a point about tools as extensions of our bodies, with computers and the internet as an obvious extension of the nervous system, but the extension often turns out to be an amputation. With AI this tendency is exhibited yet again. Certainly there are areas where AI outperforms humans, such as image analysis in medical diagnostics, but we still need to check the results.
In this article I will first discuss some philosophical problems of general artificial intelligence, then I will address more down to earth consequences in the arts.
General AI: doomsday scenario
According to the most fearsome prognostics, AI spells imminent doom to humanity. Details vary, but a typical imagined scenario takes the route of general artificial intelligence leading to a self-improving intelligence explosion. Nick Bostrom lays out the dynamics of this scenario, which, given some plausible sounding prerequisites, will happen when an AI is given the task of improving its own code. When it successfully improves itself slightly, the next iteration will more rapidly improve itself even more. The exponential curve of self-improvement means that the AI will overtake human intellectual capabilities in all conceivable areas in the blink of an eye and no-one will be prepared to deal with the outcome. What happens next is conjectural, depending on whether the AI can be ascribed an "agenda" or "intentions" of its own, and whether or not its agenda is in conflict with the well-being of humans. If there is a goal to be optimised, such as producing the maximum number of paper clips, to take Bostrom's example, the AI will go on pursuing that goal whatever the costs.
The scenario of a self-improving intelligence explosion cannot be entirely dismissed, even though it is unlikely to be an immediate risk. Human intelligence, however, is fundamentally different from artificial intelligence as it has been conceived and implemented thus far. Bostrom does have rather extraordinary technologies in mind, such as whole-brain simulation, which indeed sounds like an avenue to the take-off point of an intelligence explosion. Meanwhile one might argue, as Roger Penrose has tried to do, that human consciousness is such a peculiar phenomenon that it cannot be simulated by any computational process, no matter how advanced. Since we humans regard ourselves as being at once intelligent and conscious, we have no internal experience of intelligence without consciousness. Therefore AI, which clearly lacks consciousness, may be argued to also lack intelligence in any real sense. If Penrose's argument should turn out to be wishful thinking, then consciousness might just be an emergent phenomenon that could arise in silico once the computer and software is powerful enough. According to the fascinating Integrated Information Theory, consciousness arises in highly interconnected systems, such as parts of the human brain. It also predicts that lower degrees of consciousness would exist in other interconnected complex structures, leading to the view of panpsychism, that all matter carries consciousness, if only to the slightest degree. Current computer architectures don't even nearly approach the interconnectivity of a brain, which means that present forms of AI are at best rudimentarily conscious, perhaps at the level of a rock or some single-celled creature.
When it comes to assessing the risks of general AI, I think that speculations of its potential possession of consciousness are quite irrelevant. Useful applications as well as potential dangers do not depend on whether or not the AI reaches consciousness or if its intelligence is comparable in nature to human intelligence. Furthermore, the plausibility of the intelligence explosion scenario hinges on various factors, such as the material availability of sufficient computational resources, and the political unwillingness or failure to regulate AI research. Conceivably, there could be bottlenecks that would prevent the intelligence explosion from occurring, such as the scarcity of materials needed for computer chips, or excessive energy consumption, that cannot be easily overcome by more clever solutions, even if AI might be used to suggest more efficient alternatives. But this is where one of the real dangers lie, and it already affects us. It is easy to forget that computation takes material resources, draws energy, and has a CO2 footprint. Large server halls may be built on arable land, destroying fertile soil, using large quantities of water for cooling, depending on rare minerals from mines that cause environmental problems of their own.
Although not identical, the adverse sides of AI are intimately linked to capitalism's growth imperative. Ubiquitous computing and the internet of things is a case in point; even if it doesn't rely on AI there is an incentive for using more computation whether or not it is useful, necessary, or wanted. Cryptocurrencies based on proof of work and the minting of NFTs likewise draw enormous resources, while contributing only to a Ponzi scheme of sorts, if their critics are right.
Apart from its material side effects, what makes AI potentially harmful is how it is interwoven into society, for instance when important decisions are delegated to AI in policing, autonomous weapons, or self-driving cars. We have been warned about unemployment from automation since long, and renewed warnings are issued when not only physical work is taken over by robotics but more intellectual work such as writing summary news articles or the routine work of lawyers also turns out to be prone to automation. The widening wealth gap between low and high-skill workers is one of the serious consequences of automation, though not as spectacular as the most dramatic doomsday scenario where an evil AI mastermind takes over. Some work is so boring or hazardous that it should be automated, but replacing humans by automata in shops or telephone services usually does not make the experience any more pleasant or any simpler for the customer. The always excellent Peter Joseph is right about most things, steeped in systems thinking as he is, although I do not quite share his enthusiasm about automation.
Some believe that the current AI hype is just that, a market bubble ready to pop any time. The achievements are spectacular rather than revolutionary. CEOs feel inclined to catch up with the latest technology, Ploum writes, in the hope that AI will make things more efficient and cut costs, while it often turns out to bring more expenses and its output needs a manual check anyway.
Response to a survey on AI in art
Recent academic research is saturated with AI and automation. They follow the buzzword instead of engaging in a critical assessment of its consequences. At least that has been my general impression from the descriptions of available PhD positions in music technology I have seen lately. However, there are exceptions.
A year ago I was contacted by a researcher who had got the impression that some of my projects were related to AI, which, I would say, is a strenuous comparison. As this researcher clarified, there is an EU definition that casts the net wide enough to include even simple Markov models, which in fact I have used for analysing the transition probabilities of letters in a text and generating strange words in some imaginary Indoeuropean polyglottal language. But I think it would be misleading to slap the AI label on my simple program. Nevertheless, I got the opportunity to answer a questionnaire, parts of which I think are still worth sharing in a slightly edited form.
12. Do you see some downsides (or risks) in using AI in the artistic fields?
When using software written by others, I tend to prefer that which is more transparent about what goes on inside. For example, I use a small subset of Csound for sound synthesis and processing and, when feasible, I write my own programs in C/C++ where I (hopefully) have a detailed understanding of what the program does. With machine learning, it appears that no-one understands exactly how it comes that a model achieves what it does. In my algorithmic compositions, I program everything myself from low level routines for sound synthesis to the generation of the entire piece. A curiosity about algorithms or formalised representations of music has been an incentive for this work. Thus, it would be pointless to trust a third party ready-made composition program with a few parameters to tweak, instead of having to engage in the modelling from the ground up.
Evolutionary computing was in the vogue when I did my PhD, and since I worked on sound synthesis, feature extraction, and algorithmic composition, it was almost expected that I too should find use for evolutionary algorithms. But I didn't. Limited programming skills may have stopped me, but most of all, I needed a subjective evaluation in the loop which I never found a way to formalise away.
In general, I'm weary of advanced technology proposed as solutions to problems that may be solved by simpler means, and nudged upon us whether we like it or not.
Specifically, I see a few downsides with AI for artistic use. First and foremost, there is a risk that relying on tools makes us unlearn skills that take some practice to maintain. GPS causes people having trouble navigating by a sense of direction, pocket calculators become an excuse not to practice mental arithmetic, MIDI sequencers take over the need to practice instruments, and now AI will take over some aspects of artistic creativity. I understand and expect that creativity will be applied in other ways, at other levels of abstraction, but some skills will be at risk of atrophy.
Another downside is the reliance on big data and corpora of existing music or art, copyrighted or not. There is the so-called ethical problem of machine learning being applied to virtually any material available on the internet, without express permission, or even neglecting prohibitions against its use. But the reuse or rehashing of existing corpora is likely to produce interpolations of what already exists, instead of something radically different. If "creative" applications of AI take over in artistic circles, then there is less incentive to create something from scratch, by more laborious methods, and instead rely upon the AI tool to create its remixes.
I am aware of the high energy demands of machine learning on big data sets and I do consider it an additional, important reason to refrain from it or to be judicious in its use.
5. What is your knowledge of the environmental sustainability aspects of your artworks and artistic practice?
I work in a wide range of media, each with very different impacts on the environment. With prints on paper it's quite easy to reason about resource consumption. The printing technique itself is fully mechanical, a printing press is a durable machine that lasts for many decades. Digital or electronic art and music, on the other hand, is harder to assess in terms of material, waste, and energy consumption. Clearly, most computers do not last that long and then need to be replaced, implying toxic e-waste and mining for rare earths and other minerals. I am also aware that the electric power consumption of a laptop is not negligible.
Artist careers demand international presence, attending festivals, conferences, exhibitions. I have to a very large extent abstained from such activities and avoided air travel. This has been in part motivated by a concern about environmental impacts, which I have been aware of at least since the 1980's.
6. Do you keep track of energy usage (or other measures) during the creative process?
No. That would be nearly impossible, as well as a distraction. In particular, I do not keep track of my working hours. Such bookkeeping would only imply additional unproductive work. Artistic practice must have fluid boundaries between work and leisure.
7. What kinds of compromises or contributions could you imagine doing, or have done in your work for sustainability purposes?
A few things were mentioned in response to question 5, such as cutting down on long-distance travel. If CO2 emissions were the only thing to worry about, it might be enough to follow the advice of Wynes & Nicholas: to fly less, avoid having a car, avoiding eating meat, and having fewer children. Unfortunately, the narrow focus on CO2 emissions tends to distract from other looming catastrophes, such as biodiversity loss, fresh water shortage, top soil depletion and quite a few others. It may not be immediately obvious how artistic practices relate to this big picture, except that advanced technology (as well as some simpler technologies such as coal fuelled plants) and the demands of economic growth are generally detrimental.
I think the response to this situation is very astutely formulated as a compromise because, as mentioned, building an artist career requires an international presence and networking, not only by travelling but also by using the big corporate internet (social media, as opposed to the smolnet). Moreover, in music and art, the use of shiny new technologies tends to introduce a novelty value that translates into artistic value. It can be seen in the insistence on acousmoniums with 50-100 loudspeakers, the increasing image resolution in video for immersive effects, or in art's interest in science and technology for its felt relevance to the present state of society. I tend to see in it an unwitting propaganda for technology optimism. In stark contrast to this, there is Arte Povera, and performance traditions that require nothing but the physical presence of the performer.
As for myself, I have already committed to a few of the possible compromises, and I could take it further by turning even more to slow/simple technology. Raising awareness is also important, although pointing fingers and moralising will only put people on the defensive. I think it is better to quietly withdraw from whatever harmful circumstances one can.
11. What do you see as (your / the artist’s) responsibilities towards society when using AI in artistic work?
I prefer not to think in terms of responsibilities, but rather about actions and consequences. As I have argued above, there are significant downsides with the artistic use of AI, in particular that involving models trained on large data bases of other artists' creations. At the very least, it would be good practice to avoid data bases sourced without the express permission of each creator involved.
13. Who carries the responsibility if an AI artwork causes any form of harm or gets misused?
It would be useful to distinguish moral and legal responsibility, given that these might not agree. However, as AI regulations largely remain to be put in place, we can only reason about responsibility in subjective or moral terms. And who gets to decide whether or not "an AI artwork" has caused any harm? Cases of near intellectual property rights infringements by stylistic cloning of existing works are easily imagined, but the more systemic and subtle forms of harm I have mentioned (and also alluded to in this questionnaire's interest in environmental impacts) might be obscured by the focus on IP rights and other forms of harm to individuals.
Thus far the questionnaire.
The curses of generative AI in art
The debate about generative AI has largely revolved around issues of creativity, copyright, and artist economy. The fear of course is that generative models will make artists, designers, musicians, and writers obsolete. Again, the fearmongering may be overwrought, albeit more realistic than the most apocalyptic visions of a post-intelligence-explosion world. Reactions to the products of generative AI range from the terrified, "now there is no reason for artists to create their own images, or for composers to compose music when the machine does it so much better" – to the mildly disappointed, with complaints about the failures of AI to achieve anything near its promise. As far as digital content goes, it is true that some generative AI models have already achieved interesting results (anatomical imprecisions such as hands with six fingers notwithstanding – or included!). These images may be deemed good enough for musicians to use as an album cover, instead of paying a professional designer for the job. Likewise, muzak and the bland background music washing over documentaries from beginning to end, demanding no particular creativity, but rather an active restraint from creative shaping, is the perfect niche for generative AI. All kinds of mind numbing functional design is up for grabs. As the models improve, more "artistic quality" in the form of verisimilitude seeps in.
The output of generative AI lives in the medium of digital content, which is what allows its rapid and wide distribution. But, as most visual artists and composers of old-fashioned sheet music know, the work is not the digital file. A painting as a physical object is different from its representation as a digital image, and so far AI has not been put to the production of physical works of art in such a quantity that it worries anyone. Similar to how photography changed painting (briefly discussed in a previous post), and later the smart-phone changed photography, AI imagery may drive aesthetic changes in both painting and photography, even among artists who decline to use it directly. It also provides another reference for how to view images, as happens occasionally when an artist is wrongly suspected of having used generative AI in their work. The new insult is to say that something looks like AI art!
A deficiency in currently available generative AI models is their interface, or rather lack of interface allowing expert use, tweaking model parameters, rewriting parts of the code; in other words, for cunning users to use it creatively and not just as a consumer of semi-readymade products. The text prompt is like a convoluted jukebox that doesn't always play what you specify, especially if it happens to be far outside of its repertoire which is determined by the analysed corpus it is built upon. AI, as it is offered in proprietary software, is the diametrical opposite of IA, or intelligence augmentation, a design philosophy that arose at the same time as the first AI research, in the 1950's. Many IA inventions are today taken for granted, such as graphical user interfaces.
Viznut writes:
AI researchers wanted to create machines that would simulate human intelligence well enough to replace it. In their vision, humans and machines would compete against each other by the same rules and metrics – pretty much like in the good old industrial capitalism. On the other hand, the IA side wanted to create machines that would assist and "amplify" people's natural intellects – so, instead of competition there would be co-operation combining the strengths of humans and computers.
When computer art and computer music were new, these terms often evoked the impression that the machine itself had created the art as opposed to the artist-programmer who ultimately came up with ideas of how to use the machine, and who decided what output to accept. The practitioners themselves, such as Vera Molnar or Herbert Brün, had a more reflected understanding of the computer's role in the creative process. But the misattribution of creative agency to the computer has survived. However, in the case of generative AI, its plundering of existing works of art to serve as a data base from which seemingly new products can be interpolated implies that there is always a derivative aspect to the output. More people are able to see through the illusion that the AI model by itself has produced these dazzling results as hints of their human provenance shine through. If you ask for surreal imagery including ants, crutches, and melting watches in a desert, the result might not be as satisfying if Dali had been excluded from the training set. Insofar as there are commercial interests tied to copyright, legal actions or calls for regulation may be expected, as when record companies object to artists on their labels being used to train neural nets.
Generative art, which is what most early practitioners of computer art were working with, should be distinguished from AI art. While both rely on computers and programs, the former is made by writing code to generate images or music, and does not rely on big data sets of pre-existing artworks. Although most pioneers of computer art worked with generative art, one can find early predecessors in the approach of imitation by means of analysis of existing works, such as Lejaren Hiller's computer music.
When it comes to non-artistic utilitarian uses, let's say gathering accurate information or writing up an essay you might try to hand in for an exam, AI has some well-known vulnerabilities and problematic consequences. There are supposed to be safeguards against asking for practical advice on certain manufacturing processes such as you can find documented in the anarchist cookbook (which you should not try at home or anywhere else), but tricks have been discovered that allow the user to tease out this information anyway, so-called prompt injection attacks. Piecemeal countermeasures are being put into place, but as long as commands are sent over the same channel as the data the flaw persists.
Another problem concerns the reliability of answers, called into question by the occasional hallucination. Parody and irony are hard to distinguish from truthful content in the analysed texts. Furthermore, as these large language models sweep up written content from all over the internet, they eventually begin reading their own output. Then the feedback poisons the content, which rapidly deteriorates into nonsense. Some artists already use the strategy of deliberate data poisoning to prevent their works from contributing in any meaningful way if it should be collected by some AI enterprise. Maybe workarounds will be found against the inbreeding effect and data poisoning attacks, although then there might be new counter-attacks against these measures as well. It is a serious weakness of generative AI for artistic use that the content is already a melange of existing works, which would become even more homogenised when AI-generated output is re-analysed and incorporated into new iterations of the model.
Professional contemporary artists and composers have much less to fear from generative AI than designers and creators who produce for the commercial mass market. Partly that is because they do not produce "digital content" (with a few exceptions, obviously) as their primary artistic medium. The physical work of art still has something of an aura which no digital image on screen can compete with. There is also a performative aspect. Showing that you are able to produce a work of art with your own hands, even if it isn't a masterpiece, beats the fully machine generated output.
Generative models definitely stand on the shoulders of giants, namely all the appropriated artworks of other artists. The counter-argument – that all artists build on their predecessors in more or less overt borrowings so this is no different from generative AI – might appeal to some, but there are differences in the quantity of accessible material and the shallowness of engagement when writing a text prompt, in contrast to years of familiarity by exposure to previous art and painstaking attempts to mimic some of its qualities.
If artists are not directly threatened by generative AI content because they have advantages such as not being confined to producing only digital content, and being able to perform and put their works in meaningful contexts, the sheer volume of AI content may still be harmful. If spotify playlists become filled up with more automatically generated content there will be less room for artists who have actually created something themselves. Scarcity of art and music has not been a problem for a long time. The democratisation of the creative process in which everyone can become a bedroom producer and upload their content to the web already means that there is more content available than there are listeners out there to discover it. Likewise, the ubiquitous smart-phone with camera and an internet connection is a recipe for the inundation of images that assault our attention. In this context, what we least of all need is an even greater volume of digital content, much of it not of an impressive quality.
From an aesthetic point of view, perfection is not always the goal. Digital artefacts and errors are sought after in glitch art. Figures with way too many fingers on their hands is one of the easily spotted traits of AI art of a certain period, and is arguably more interesting than when it succeeds in creating realistic life-like scenarios. Translation services also have improved significantly over the last ten years. It used to be fun to engage them in a whispering game, translating a text into a foreign language, then to another language, and back again to the original language. Hilariously surreal misunderstandings were commonplace. Now, such a procedure too often reproduces the meaning of the original text, only paraphrased in more common and less expressive wordings. There is a regression to the mean, which happens for purely statistical reasons: Large language models are simply based on statistical analysis of the likelihood of certain words occurring close to other words, which means that rare words with a more precise meaning will likely be replaced by more common words with less precise meaning.
To conclude, I would like to stress again that the environmental effect of AI, in my opinion, is the strongest argument against it. Training a neural net to produce high quality audio or images is very costly (video must be orders of magnitude worse). Using it to produce content draws comparatively small amounts of energy per generated item, it's just that a lot of content gets generated.
A strange argument is sometimes advanced when raising the issue of energy demand: Yes, but if humans would do the work that AI has automated, then their energy consumption would be much greater! Not only does the human artist need to expend energy on moving their hand with a brush in front of the canvas, they also have to plan the painting in advance, in fact, they must spend years practising in art school before they can come anywhere near the neat results of the AI program.
I think this argument is flawed for various reasons, apart from the atrophy of skills I have already touched on. First, the efficiency of AI means that much more output will be produced. This is related to Jevons paradox, which states that increased efficiency is swallowed up by increasing use of the more efficient technology, resulting in a net increase in energy consumption or material use in the end. Another point is that when you consider the human labour which the automated AI replaces and makes more efficient, should you not take into account the labour of those programmers who spend years perfecting their machine learning algorithms?
Reasoning about energy consumption is far from my expertise, but it appears to be possible to calculate roughly the energy requirements of various human activities such as moving a brush back and forth over a canvas. Vaclav Smil might not have calculated that specific case, but he has educated readers about energy consumption in many other areas. With AI, Jevons paradox will apply, and perhaps we will eventually have names for other paradoxes such as its presumed utility which turns out to have more adverse effects than practical benefits.
Further reading
Nick Bostrom (2014). Superintelligence. Oxford University Press.
Vaclav Smil (2017). Energy and Civilization. A history. The MIT Press.