Ville Hytönen: I’ve been playing around quite a bit with Midjourney, an image AI application. Not so much professionally, just to generate a huge number of funny pictures and memes. Probably the most useful application for me so far has been creating the illustrations for a parody piece I wrote for Skeptikko magazine discussing the covert KGB connections of Constable Reinikainen [a well-known Finnish TV sitcom character in the 1980s]. I could well imagine that, at least for now, the majority of the text that AI is made to generate is used for ironic meme-like purposes.
However, in the future text-based AI may facilitate wholly new genres of literature that we can scarcely even imagine, such as genres based on extremely precise mathematical modelling or computation, or language hybrids in experimental literature.
Ano Sirppiniemi: AI advancements could very well foster new formats in music too, such as pieces that constantly adapt themselves within the constraints of parameters set by the creator and are different for every listener every time. Mind you, generative music has existed as a concept in both art music and pop music long before AI was even conceived of, but AI applications can offer new opportunities.
It’s interesting to ponder at what point AI models will be able to produce extensive structures independently. Most of the music models that currently exist can generate about 90 seconds of music at a time, and it still takes quite a lot of human control and choice to create any pieces of music longer than that.
Using text prompts to control the models is a skill of its own. I’ve tried out the Stable Audio model produced by Stability AI, which is trained on a catalogue of American production music. If you feed in text prompts that you make up yourself, the results are generally pretty bad, but if you copy or emulate the phrasing with which songs are described in the original catalogue, then the model can produce very much more convincing output.
VH: It’s the same with text; for now, AI can’t write a coherent novel. The best-known example in written art is the Aum Golly poetry collections created by Jukka Aalho using ChatGPT, but anyone who reads them will notice the same problem: a certain lack of authorship is clearly apparent in the text. There have been some decent efforts in children’s literature where both text and illustrations were generated by AI.
I’m coming around to the opinion that a person experiencing art needs there to be a human creator behind that art. It’s difficult for a viewer, reader or listener to contemplate AI as a creator, because it isn’t. AI is neither intelligent nor creative. It’s a tool, and without a user it’s empty.
This sets the stage for a further re-evaluation of the relationship between work and creator. We may see creators reassert themselves.
AS: I agree!
VH: I’m sure that AI will be used, and probably is already being used, quite a lot in genres of literature that are not so language-driven. Raw material for easy-reading fiction can be generated using language models, and I’m sure that experimental literature of various kinds is being created using AI even as we speak.
I’m coming around to the opinion that a person experiencing art needs there to be a human creator behind that art.Ville Hytönen
Can AI destroy genres of art – of music or literature?
VH: AI has already been used to create thousands if not tens of thousands of works of questionable quality, and Amazon constantly has to remove titles like this from its catalogue because of unfavourable feedback. The problem here is probably that writers who have no experience of how to become a writer are churning out substandard literature by the mile. Because AI generates literature by data mining existing literature, the poor quality of AI-generated works may in the long run impede the development of AI itself. A snake eating its own tail, if you like.
AS: The risk of creating a bubble is certainly a very real one. If all language and music models are trained with “all art created up until 2023” and they begin to generate new material on the basis of that, then it’s a no-brainer that the quality of the works will begin to erode, or they will converge into homogeneity.
AI could also be used to whip up extensive catalogues of pieces for new virtual musicians or indeed for musicians who are already dead. The sheer number of titles will quickly become a problem, though. It’s estimated that even now the number of songs uploaded to music streaming services is more than 100,000 per day. AI could easily multiply that number many times over. Finding music or literature of a high quality – particularly new works of a high quality – will quickly become an enormous challenge.
And as always, we have to ask: who cares? Do people really want to spend their time consuming content generated by AI? I feel that all this is also a social experiment on a massive scale as to whether the value of a book for a reader or of a piece of music for a listener is the same regardless of whether created by a human or by AI. I believe that the value and appreciation of human creators will persist and even increase if the markets are flooded with AI effluent.
Data mining, but whose rights?
VH: Authors see text data mining by AI as the greatest threat to copyright. It’s already been found that dozens of Finnish authors and their writings have been used for improving language models without their consent and without compensation. The actual figure is probably many times higher.
I’ve made motions at meetings of the European Writers’ Council and the Nordic Council of Authors and Translators that there should be an EU-level body to levy copyright fees from AI companies and to distribute the proceeds through copyright agencies to writers as grants.
AS: There’s similarly been a lot of talk in the music industry about the use of existing music as training material for generative AI models. Many models seem to produce identifiable copies of or extracts from existing works, which has already led to some infringement lawsuits against AI companies concerning images and song lyrics.
How the revenue from content generated by AI models could be shared with the original creators in practice is a huge issue. This may not even be possible if you wanted to drill down to individual songs or pieces. Obviously, you need to have permission from the original creators if you want to use their material for training an AI model. It’s also important that AI model developers be required to document what material they have used to train their model.
VH: My feeling is that it’s not even necessary to try to chase down the individual pieces or creators involved in data mining of text or music but that copyright agencies should instead be paid some kind of percentage-based royalties derived from turnover or profits and that the agencies would then distribute that money to creators in the form of grants. That way, the work of deceased creators would also benefit current creators.
AI and democracy
VH: Professor of Computer Science Hannu Toivonen has said that language models democratise writing, because they make it easier for everyone to achieve a certain minimum standard. I’m not sure whether it actually makes writing democratic that an incompetent and textually immature writer can pour out half-baked text with the assistance of AI.
AS: Language competence and being able to produce good language is surely fundamentally important for democracy, more so than being able to make music.
In music, we may similarly argue that AI will enable a huge number of people to express themselves through music who cannot currently make music at all. This can be a positive thing, but if the flip side of that is that traditional music-making skills and composing begin to decline or erode, then the outlook is bleak. Like language, music is an important part of all human communities and a means of self-expression that is firmly rooted in emotions and feelings of identity and community.
Maybe we might think that if, in literature and in music, we outsource the creating of new art to machines, we lose a lot of the plurality and freedom of art and human thinking. This would undoubtedly be a threat to democracy.
VH: Literature and more generally the ability to produce texts of different kinds requires a huge amount of technical competence, which I fear will somehow wither into a sort of copy-paste competence – editing, basically. Editing text generated by AI is not a problem in itself, except that if not even the professionals who are the most profoundly familiar with language produce language themselves, then language will begin to narrow in scope. Newspaper articles are already showing a reduction in vocabulary, although there may be multiple reasons for that.
Maybe we might think that if, in literature and in music, we outsource the creating of new art to machines, we lose a lot of the plurality and freedom of art and human thinking.Ano Sirppiniemi
Creative AI in 2028
AS: Personally, I don’t believe that any sort of general AI will exist as soon as in 2028, although looking at the pace of recent developments, I may be wrong. In specific areas of problem-solving, such as generating music, advancements may be very rapid indeed. Looking at how AI models have progressed in generating images and discussions, it seems certain that the quality of musical material produced by AI models will improve rapidly in the near future.
The music industry has always been an early adopter of new technologies in composing, recording production and music distribution. And indeed often the debate about new technologies has been just as polarised as it is now with AI: just think about the rapid proliferation of synthesisers and drum machines or sampling towards the end of the past century. I believe that AI tools will become quickly established among the practices and structures of the music industry.
VH: I remember when I was a teenager, there were stories about recreating the voices of dead singers on a computer and returning them to record new music in that way. There was talk at least of Freddie Mercury in the late 1990s. It’s surely much more of a possibility now than it was then?
AS: AI is already astonishingly good at copying singing voices – we’ve heard many examples of that this year, such as Johnny Cash singing Barbie Girl. It sounds utterly convincing in its way and serves as a brilliant showcase for what the technology can do.
VH: I’ve thought about the same thing in literature. When AI is able to analyse and reproduce the style of a specific author, perhaps it would be possible to reincarnate some of the world’s best-selling authors. Why couldn’t Sidney Sheldon or Barbara Cartland return to the best-seller lists? Why couldn’t Enid Blyton write 800 more Famous Five books? It is possible that all formulaic literature could be replaced with AI by the year 2028. But anything more challenging, more unexpected and more profound will still require a human brain for a long time to come.
AI may not steal anyone’s job in the creative industry as such, but someone who is able to use AI well may get a job ahead of someone who is not willing to learn how to use AI.
AS: I agree – the AI transition cannot be stopped, so while it’s important to delimit the threat scenarios, it’s equally important to actively experiment with and use these new tools and to try to understand their potential.
Featured photo: Ville Hytönen and Ano Sirppiniemi. Photo editing: Lasse Lehtonen
Translation: Jaakko Mäntyjärvi