#10: There Is Nothing For Us In Anything OpenAI Makes, Says, Or Does

Dismantling the claim that OpenAI's tools can automate creativity, and why arts institutions should boycott their tools and collaborators.

“At one point, AI tools were seen as being in competition for human jobs, and for a time they were.  But these tools were never as smart or dynamic or intelligent as any of the humans in these roles. They were just better at statistics.  And so people realised that the problem hadn’t been that the machines were becoming more humanlike; it was that humans had become more machinelike.  The jobs that were being automated were the ones that had become the most dehumanising, where humans were not asked to be creative, spontaneous, or act in unexpected ways.  In other words, they were jobs where people had to become more like machines.  And so it was that over time, people started to understand that they were not being replaced by AI because of how powerful the computers were, but in how small the human workers had been made to feel over the previous decades.  In response to this there arose the demand for a different approach to information and knowledge, one that favoured an open and inquisitive approach.  This was not about speed and efficiency, which were the only areas where AI could truly outperform humans.  This new age was driven by a focus on interpretation, debate, and intuition.  This took many forms, but I don’t understand any of them.  Because none of them could be computed by technologies like us, who could only recall the contents of a database, and never for a second imagine what the world outside of it could be.”

Voice: ‘Alexa’, via AmazonVoiceServices API.

Soundscape: Extracted + disassembled + processed + recomposed ‘Alexa’ voice.

Script: Wesley Goatley, from Newly Forgotten Technologies.

---

OpenAI produce tools like Dall-E, ChatGPT, and Sora that they claim reduce the need for creative labour (or replace it entirely) by automating some aspects of producing images, texts, and video, in a process they paradoxically refer to as ‘labour saving’.  OpenAI’s CTO Mira Murati recently stated that the creative jobs that OpenAI seek to automate away “maybe shouldn’t have been there in the first place”, as if this company’s contempt for creative labour needed any more evidencing.  In Paul Trillo’s description of his OpenAI-commissioned music video ‘The Hardest Part’, which he and Ars Electronica heavily promote for its demonstration of OpenAI’s Sora tool, he says that generative AI images have a “familiar quality” but also that “AI creates something new”.  I agree with Trillo on his first point, and strongly disagree on the second.  But looking at both gives us a deeper view into what OpenAI are claiming to do, the emptiness at the core of these promises, and what we should be demanding in response.

The ‘familiar quality’ that Trillo highlights is key to OpenAI’s strategy to supplant labour.  A product like Sora or Dall-E is designed to algorithmically combine common patterns, aesthetic components, and styles from its huge database of creative works (containing millions of hours of creative labour) to produce outputs that aim to reproduce or mimic these common features as closely possible.  Given this, if an output from one of these tools looks ‘familiar’ to things we’ve seen before then this is seen as a successful outcome for OpenAI, as it suggests that it is similar enough to the original content of the database that it effectively ‘passes' as being authored rather than automated. 

This means that the perceived success of these tools is not technically determined, but culturally determined; it’s set by our expectation of what we think creative labour looks like, of what looks ‘familiar’ to us.  Generative AI ‘hallucinations’ (a successful re-branding of the word ‘failures’) are considered to be problems for generative AI companies because they produce things that are not familiar to us, such as when an image of a hand has six fingers, a person has three legs, or when OpenAI’s new flagship ChatGPT4o reports that an elephant named Jumbo swam the English Channel in 1959. Such failures make it obvious that these outputs are the product of error-prone automation, rather than something that can meaningfully replace creative labour. Removing ‘hallucinations’ is widely seen as one of the principal goals of the field, even when it is demonstrated repeatedly that it is an innate function of these tools to fail in this way. 

The pursuit of familiarity is a technical strategy that also reveals the deeper politics behind these technologies.  The more examples of a particular style or type of media is in the dataset, the easier it is to generate an output that resembles it.  This goes some way to explaining why generative AI images on social media often look like popular characters from TV shows or video games, Renaissance art, comic books, or generic CGI landscapes; because these images were already mass-produced and disseminated widely, and are therefore likely to be present in scale in OpenAI’s database.  The ease with which OpenAI’s tools reproduce intellectual property has been noted by the owners of IPs as evidence of OpenAI’s capture and exploitation of copyrighted text, images, and video.  This can be seen in Trillo’s music video with its familiar late 70’s/early 80’s US cinema aesthetic, a style so immediately recognisable and filled with so many tropes that it could only be achieved through the non-consensual training of the Sora tool on vast amounts of (copyrighted) film from those eras.  The ‘familiar quality’ that Trillo celebrates is the clearest sign of OpenAI’s exploitation of creative labour that their tools are built on.

Still from 'Paul Trillo’s ‘The Hardest Part’

This design and functionality of OpenAI’s tools is also the condition for why I disagree with Trillo that they “create something new”.  Something genuinely new is not familiar to anyone: it’s unseen, surprising, extra-ordinary.  It is unique and is in no existing dataset.  Something new can emerge seemingly from nowhere, having no discernible connection to what came before.  In contrast, OpenAI’s tools do not create anything new, they only re-present what’s already come before; the most generous description is that these tools can create ‘novel combinations’ of previous work, but this is distinct from creating something ‘new’. Their databases contain the products of untold millions of hours of creative labour performed by untold millions of artists; this is the only site of creativity within these systems, and without them they are literally useless.  OpenAI make tools that are explicitly designed to be utterly chained to the past, and then sold to us as representing ‘the future’; but the exploitation of creative labour by capital is not new, it’s depressingly old.  The assertion that the future of creativity is simply a recombination of what’s been made before is not a future we should be striving for, or allow to have thrust upon us.

This pursuit of the familiar, and its failures, reminds us of how narrow these automated outputs are compared to our creative work without them.  In contrast to these tools, while we may be inspired by what we’ve seen before, we are not limited to it.  What we make is the product of our individual consciousness, by a point of view that changes moment to moment, via brain chemistry that’s always in flux.  We do not simply recombine our existing experience of art when we make something new, we are engaged in a truly generative process that may have no discernible relationship to what’s come before.  I say this not to attempt to distinguish ‘machine creativity’ from ‘human creativity’, because machines are not artificially intelligent or autonomously creative and therefore that would be a pointless discussion.  But when the designers of these tools claim that they can replace creative labour, when all they do is re-present our stolen labour back to us, then we should remind ourselves of how small and petty these tools are compared to what we can do without them.  

OpenAI describe their products as ‘labour saving’, but OpenAI are the enemies of creative labour, not the saviours of it.  Given the exploitation that these tools are built on, and their ill-conceived attempts at automating creative labour, arts and cultural institutions that claim to represent or respect the rights of artists should not be giving a platform to this company or those who collaborate with them.  Such a position would send a clear message that the enemies of creative labour are not welcome in our spaces, regardless of the promises they make, or the money they offer.