There's been a growing brouhaha about recent progress in “Artificial Intelligence” (AI) research with the publication of Large Language Models (LLM) that thanks to “deep learning” are finally capable of producing outputs that are very credible to the superficial and/or untrained eye.

OpenAI, ChatGPT, DALL-E, Stable Diffusion, Midjourney have all made the news rounds a few times thanks to their presumed ability to “interpret” textual input and turn them into satisfactory visual or textual rendering of the prompted requirements. Three main lines of criticism have been moved to these kind of efforts, focusing respectively on (1) the nature of what is actually being produced (especially in terms of whether or not this could finally be considered true AI or not), (2) the ethical underpinnings of the training data selection, both in terms of breadth-and-scope (e.g. concerning bias in race or gender) and in terms of copyright (did the authors authorize the use of their work to train these models), (3) the future and intended use of the models (e.g. will they replace art or artists).

I'm only going to touch on the last point here, because there's an aspect that I think has gone missing in all the discussions I've seen so far, including the jokes (but are they really) about machines supposedly being intended to replace the boring, physically and mentally destructive work and instead being advanced to eliminate the creative endeavours while humans are relegated to the work the machines were supposed to eliminate. And I will argue that this is not really the case, although the net result will be quite similar to what would have happened if it had been.

There is a trend that has been going on for decades (half a century at least, in fact) and has seen a sudden jump in the last 15 or 20 years: the replacement of art with content, and artists with creators. Others have warned about this before, in more details than I will do here, but the gist of the point is that the growth of the Internet, and of social media in particular, has exacerbated the trend initiated by the mass commodification of creative output to a critical point.

This is the result of two apparently aligned interests: that of artists to be able to make a living out of their creativity, and the capitalist obsession with infinite growth incarnated by the publishers and distributors (be they formally recognized as such, or be they such by practical definitions of the term).

Art depending on the rich and powerful to thrive isn't news: it's why so much past visual arts have a religious theme, and why such an inordinate amount of words have been written to celebrate this or that local lord. If anything, actually, now more than ever the trend has been subverted thanks to new, distributed forms of patronage (see e.g. Kelsey's and Schneider's Street Performer Protocol as discussed by Cory Doctorow) that allow artists greater control on their creative endeavours rescinding the dependency from the interests of a single supporter: commissions still exist and for many are still the primary if not only way to make money out of their capabilities, but there are some artists that can afford to follow the “self-realization” principle: write/paint/compose what you want, and let those interested in your creative output come to you, as opposed to writing/painting/composing what may give better “engagement” on the platform(s) of choice.

In fact, it can be argued that the point of divergence between the mentioned apparently aligned interests is indeed the definition of engagement. For the artist, engagement materializes in an audience that is interested in and appreciates their creative output: For the publishers and distributors, engagement materializes in a returning consumer.

Some may argue that those are similar, e.g. in expectations about the production of new content over time. But despite the superficial parallels between the two, there are some crucial differences, starting from the more direct, “personal” relationship between an artist and their audience (which is not all fun and games, as it gives way to stalking, a sense of entitlement, and the myriad of downsides that come with less sterile connections). And this is something for which no metric exists, because it depends on the intangible quality of the interactions with the audience. Worse “engagement” metric do not correlate to a worse or smaller audience.

Scale is also very different: a few hundred supporters paying a few € each monthly can be sufficient to sustain an artist's ordinary life (conditional to location at the very least, of course). This is one of the pillars for the success of the digital Street Performer Protocol mentioned above: as long as enough members of the audience support the artist, the artist can thrive and their art remain accessible to all.

Finally, at least for the purpose of this discussion, there is a difference in expectations: although I'm sure any artist would be thrilled to reach the level of success that would allow them to live comfortably for the rest of their lives, from what I can see most of them would be content with just being able to make a decent living, without having to worry about whether they'll be able to cover rent next month and not depending on their partner's income or financial support from relatives.

The situation is very different for a commercial enterprise (into publishing or distribution, given the context), doubly so for one that has already achieved a certain success, and even more if it's publicly traded: there is no true “connection” between them and their consumer, operating costs are high, and most importantly there is an expectation of growth, e.g. to “increase shareholder value”. This leads to a particularly pressing need to publish and distribute new content, at an increasing rate.

Even before the Internet went mainstream, we've seen this trend manifesting e.g. for movies and animation with the spread of home theater solutions: while 40 years ago some cinemas still offered older classics, a decade later you would have been hard-pressed to find anything but new releases outside of film society screenings, despite there not being a significant change in production until the beginning of the XXI century, when massively production-cost-lowering technology improvements and expansion to “emerging markets” led to an explosive growth.

One of the key ways in which the Internet and digital media distribution has revolutionized the field has been through a shift towards subscription models. Once the domain mostly of periodic journals and service providers, the subscription and streaming model pioneered by Netflix has become the method for larger media conglomerates to cope with the new technology, after fighting it for years and trying to get governments to regulate it in the name of lost profits (and getting a long way in with it).

Subscription models are very convenient for the company, as they guarantee a more stable revenue stream compared to one-shot purchases, and may result in higher profits at equal consumption rates. But while ongoing payments can be justified by periodic journals on the basis of quantifiable periodic updates and by the service providers with an uninterrupted service delivery and continuous infrastructure maintenance, they are more difficult to justify in the case of spotty updates (release of a new movie or music album) or consumption (watching the movie, listening to the album). There's both a practical and psychological reason for this: while a subscription to e.g. a journal gives you something tangible that remains with you even if you cancel the subscription service, the digital subscription services are more like an access fee to a library owned by somebody else, and regardless of how vast that collection is, if one ends up frequently perusing the same material, the natural observation is that it is ultimately cheaper to pay for it once and own it forever than to access it in streaming.

To make the subscription model enticing, the company would thus need to either select a clientele that has the curiosity to go through their entire catalog (which no company is going to aim for, because numbers), or keep up the attention of the “generic” consumer with a continuous stream of fresh, palatable content that keeps them distracted from reconsidering the benefits of one-shot purchases.

Recommendations from the existing catalog (especially if it's a large, well categorized one) can take the role only up to a certain point, as consumption levels out. Hence the need for continuous content generation. This is not new: cable TV started to push out reality shows for the same reason; TV series were born out of the need to keep housewives engaged to sell more advertisement; and we can go as far back as the feuilleton at least as the first examples of content serialized over long runs to keep people engaged, i.e. transfixed to the specific media channel. What had been changing in recent time is the scale, and the scope, of the phenomenon, with vicious cycles that benefit neither the company nor the consumer: as the consumer aims to maximize the utility of the subscription fee, the demand for fresh content grows, and as the demand for fresh content grows, the attention to the quality of the product diminishes, lowering longer-term engagement. This reflects not only in a massive increase in production, often of debatable quality, but reaches grotesque peaks where entire series are sacrificed after the first couple of seasons on the altar of immediate engagement growth, even when the higher quality was appreciated by a meaningful number of consumers.

We are now in an era dominated by Continuous Content Generation, where engagement is not the result of sustained quality, but of a continuous renewal that doesn't even let products reach maturity before they are swapped out for a more recent one, in a desperate search for instant freshness gratification. And this is the era where ‘content’ replaced ‘art’.

Now, there are practical reasons to use the terms content and creator rather than art and artist, not least the fact that not all content in these distribution channels is art: for example, journalism and essays may use similar or related tools and medium to those used for novels and poetry, but regardless of the aspirations and capabilities of the writer, they'll rarely be classified as art, regardless of how important their role is in keeping readers engaged. So using the ‘c’ words is a reasonable way to talk about all the material when going into specifics about what is art and what isn't is unnecessary.

However, the choice of terms is indicative of a diminishing interest in the content itself, i.e. it's not just a way to indicate all the available content, but most importantly it's an indicator that it doesn't matter what kind of content it is: obviously not everybody can hire a Jules Verne or Alexandre Dumas to write serialized novels and keep selling copies of a newspaper, especially since at scale you'd need one for each kind of audience among your consumers, but we're way past the point in which this is just a matter of quantity over quality: the scale is such that what causes engagement is completely irrelevant.

Hence the spread of clickbait, mainstream trolling, enRagement algorithms, and any other strategy that helps keeping people coming back for more.

It should now be clear where I'm going with this, given the premises: this same need for Continuous Content Generation, regardless of type, form, or actual content, is the most immediate practical target for the AI/LLM that are at the center of attention today.

In this sense, these models do not represent a threat to the independent artist (or other “content creator”) that has built or can build a following of its own. The models can be considered aimed primarily at replacing the armies of scriptwriters, copywriters, “bloggers” and whatnot that provide ‘content’ for the media industries. As the models grow more sophisticated, we can expect more of the output of these industries to be produced by or with the assistance of such models, with humans relegated towards the roles of prompters and selectors/verifiers. And if superficially this may seem like a good idea to some («oh good, less effort for me to write the articles I need to publish to get paid»), it should be obvious that this will entail not only a massive cut in the workforce, but potentially its almost complete elimination, when publishers realize that they can leverage their own consumers to fuel the machine (think about user comments as prompts for the next set of articles, or how many of you are freely offering tables of possible prompts that you can rest assured are being logged for future use —how about a model that writes prompts next, for example?).

It's also debatable whether or how much or how many consumers would actually care about or even realize that the content is being produced by LLMs, although it may be important in the beginning that the automaton intervention be as subtle as possible. If any of you were hoping that the recently presented tools that detect LLM outputs would just be used to denounce the number of AI-written articles already in circulation, think again: a more likely use will be for the tools to be sold to the publishers to help identify the machine-produce content that cannot be detected, and can thus be published with less danger of repercussion from readers that do care about where the article comes from, at least until the new choice of writing system gets normalized.

While this does eliminate some of the threats that AI/LLM pose to independent artists and other “content creators” (we really need a better word for this), its influence will still need to be considered. The most obvious effect is that by reducing the opportunity for relevant employment within the industry, it potentially increases the pool of artists that will have to make a living independently (if the wish to live off their art), and the competition may make it harder to build one's livelihood from it. More importantly, however, these models are (still) incapable of original content creation, which puts a limit to the variety of what they create. This may not be apparent now that its use is very limited, but with an adoption at scale for Continuous Content Generation these limits are likely to be hit sooner than later, even with judicious use of human-direct prompts and selection.

It therefore becomes essential for the models to be periodically injected with “noise” (new training data) to increase the variability of its creation (this is something that anyone experimenting with particularly unusual prompts has noticed already). Now while it's possible that this can lead to the displaced scriptwriters and copywriters to find employment specifically to produce such “noise”, what is most likely is that the work of independent artists that continue in their art unassisted by LLM will be unceremoniously hijacked as training data —and this is not a potential threat: this is something that has already happened and is still happening, as illustrated by efforts done by nearly all art hosting and online editing services (from Adobe to DeviantArt) to change their terms of service to include wording that allows them to feed the hosted content to such models, and making this opt-out (thus enabled by default) rather than opt-in. This is an area (outside of the scope of this article, as it falls within the second theme mentioned in the first paragraph) where legislation potentially could curb the phenomenon, but it's likely that lobbying by the media companies will render it ineffective at best, and counterproductive at worst, despite the clear preference from most artist that AI be kept away from their work.

Curiously, the reason why I think that AI/LLM do not pose an immediate, direct threat to the independent artists is not only that the lack of originality and variety in the output of the models makes it more valuable for the “reprocessed art” that is behind much of the “content production” of the media industry (think for example of the speed with which subgenres saturate in mass-produced animation and comics, and that's with human work), but also because the target audience for the two forms is wildly different, and the audience that seeks out the more original and varied “content” produced by independent artists is also more likely to value it being “artisanal”: so not only it is more difficult that AI/LLM would be able to produce the content this audience seeks, but that same audience would value it more when produced by an actual human rather than mechanically (or procedurally) in response to a writing prompt.

And again, while the higher appreciation for artisanal production poses a problem for physical production, it is considerably more sustainable in the digital space, by the higher efficacy of crowdfunding.

There is something ironic, I feel, in the “techbro” enthusiasm for AI/LLM travelling often in tandem with an unhealthy obsession for cryptocurrencies and the Non-Fungible Token (NFT) craze that goes with it. (One of) the purported intent(s) of NFTs (as advertised by said techbros) is to help artists “monetize” their art by artificially restricting purchases of what amount to “certificates of authenticity”. The reality (to no one's surprise, except the fools that bought into the scam) has been very different: most of the NFT-“certified” content that has floated around so far has been either procedurally generated, unoriginal, uninteresting variants of thematic images designed for unsubstantial products with the only aim of generating noise before the rug pull (Exist scam), or outright “stolen” infringing on the copyright, licensing, and/or moral rights of the original authors.

The irony here is that what makes art truly valuable isn't restriction on consumption (making it inaccessible); in fact, it could be argued (but I don't want to get into a philosophical discussion about what makes art art) that art achieves its peak value when it reaches the widest audience. (In this sense, the infinite reproducibility of digital works of art have the potential to be valued as nothing before them.) And this value stems from the uniqueness of the work of art at creation time. Each work of art is unique because it could only have been created by that particular artist in those particular circumstances: a different artist, or even the same artist in different circumstances, would have produced something different. (Heck, even in the same circumstances, just because of small differences in context: think for example about the English and German versions of the same movie by Hitchcock.)

The digital work of art is unique because the work of art itself is not the bits encoding its representation, in the same way in which the novel isn't the ink and paper with which it is transcribed: the art is the story, the images, the sounds that result from the decoding of that representation, and how they reflect on the aesthete's mind —and these are unique by creation.

I could go off on a tangent here on how art is the antithesis of capitalism, and the principle of substitution not applying to creative work is just part of it, but (back on topic) coming to terms with this is what shows the irony of the AI/LLM+NFT fandom: art is intrinsically not fungible, by virtue of creative originality at inception. LLM-generated content is the epitome of unoriginality, and the artificial restriction imposed by “minting” NFT for it does nothing to compensate for this.

So maybe instead of worrying about how these glorified procedural generation models may threaten the livelihood of artists, we should focus on rethinking the socioeconomic system in such a way that creativity may be valued for what it is, instead of the perverse system of disincentives that has been built around it to force it into the capitalist concept of “value” by (artificially created) scarcity of access.

(And how's «digital (street) performer» as a better alternative to «content creator»?)

Appendix: I'm going to collect here links to Mastodon threads that are relevant to the points discussed in this post: