Ruminating on creativity and AI

I was pleasantly surprised to come across about an hour of music – ostensibly original – by a YouTube channel I had never heard of. It was their first upload, with no other information available. I enjoyed the mystique of an extended play thrown together by an anonymous group. I’m hesitant to link the video, however; and I’m hesitant to label the uploader as an “artist” or “creator” because as far as I can tell, this music was generated using AI.¹

I have no concrete evidence for this claim, but, circumstantially, the setlist had many track titles that came across as AI-generated or at least lazy. None of the album art they used is original, either. It was ripped – unmodified – from accomplished artists, their credit left as a byline in the description of the upload. Missing from the description was any mention of people or assistance in producing the album: no assistants, no musicians, no writers, no audio engineers, no studios, no labels.

Most damning, however, is the sheer volume of content. Since the channel’s inception, they have uploaded five hours of purportedly original music within a month or so. Only abnormally prolific artists and collectives can hold a flame to this; Black Sabbath² comes to mind.

It would seem I am not entirely alone in this feeling. Given the age of the account, there was not much discussion online about it, but most of the deeper discussion I could find seemed to be around whether it was AI-generated or not. The account is seemingly one of many conspiring to shovel lofi-inspired sets into the ether, or so the conspiracy goes.

If this is, in fact, AI-generated, then I am deeply troubled I found it so enjoyable.

In the mainstream popularization of LLM technology, the layman has gained access to models that produce generally “good” artwork and music: inoffensive, generic, popular, and appealing.³

In this proliferation of access, however, critical elements of the creative process have been harmed; namely synthesis and authorship. “Prompt engineers” come up with some word soup that excites the right paths in a neural net to spit out the closest thing to what they had in mind. In this case, the prompt author has performed synthesis to a certain degree: they have crystalized what they saw in their mind’s eye into written form. Some prompt writers, though, have no end-product in mind. They saturate their query with trendy keywords that a forum post said would improve model performance, and then trial-and-error it out until they get something they believe others will enjoy.⁴

There is a strong argument that in neither case can authorship truly be claimed by the prompt writer; it was an LLM that did the legwork. However, a trend in subsets of “AI-creatives” is to disingenuously present an LLM’s creation as though it were entirely their own, either through omission or blatant falsehood. In some cases, it is intentionally mislabeled to garner attention more aggressively.

Ignoring the host of meta-arguments to be had around what art even is, there are obvious downsides in this approach. Writing that prompt to generate some music does not take nearly as much time and effort as it does to learn an instrument, create a tune, record it, mix and master it; but, if acclaimed, both will provide significant gratification in the creative process.⁵ Presenting it as though it were one’s own creative endeavor makes this gratification nearly indistinguishable.

Animal life is biologically incentivized to stimulate reward pathways in the brain with as little effort as possible. However, the body will build tolerance to and dependency on these stimuli, accelerating the means by which intelligent life exploits its reward systems. Social media started us down a path with quick gratification for low-effort or self-aggrandizing content and incredibly fast feedback loops. Decreasing the loop’s latency was the primary mechanism by which social media evolved for about two decades. If not Facebook, then Twitter. If not Twitter, then Instagram. If not Instagram, then Snapchat. As of this writing, vertical short-form video dominates the Internet, largely motivated by TikTok.

What LLMs’ exposure to the mainstream has achieved is a transformation of the creative process itself into an artificially exploitable mechanism by which we engage our reward paths. By significantly reducing the absolute quantity of effort demanded to produce creative works and increasing the latent demand, creative works themselves become a byproduct of human consumption. This is problematic for both humans and LLMs: as the amount of publicly-accessible content is polluted with AI-generated works, it is accompanied by a sharp decline in information with which state-of-the-art LLMs can effectively train, poisoning the well.⁶

This potential for net-negative codependency is what worries me. LLMs rely on a steady inflow of new ideas in order to innovate. If humanity grows increasingly addicted to this form of instant gratification, though, the frequency and quality of original works will decrease. As a year-over-year deficit in human creation builds, so too will the volume of AI-generated content uploaded to preserve the perceived standard level of output.

But LLM-based derivative works cannot be truly novel – until this changes, the absolute peak of our ingenuity will remain tied directly to the output of humankind. Allowing this peak to fall over time and doping it with AI slop risks exposing humanity – at least in part – to the same existential threat as LLMs themselves.

If this album was AI-generated, then I am deeply troubled; complacency aids dependency.