remember when we could tell?

“Remember when we could tell?” This question became a theme during my practice-led PhD on AI. Encountering glitchy AI artefacts evoked a sense of anticipatory nostalgia for a time when these flaws revealed their creation process—a cosy, uncanny feeling.

I have since created several physical objects based on AI artefacts that have some of these qualities. If nostalgia is rooted in comforting memories that provide stability during change, then these artworks are kitschy reliquaries of those moments. They serve as nostalgic reminders of when AI’s flaws were apparent and comforting, like treasured relics of a bygone era. These include Jukebox Frank (below) and Five Reliquiae.

Jukebox Frank

Jukebox Frank is a single sided vinyl pressing of ‘classic pop in the style of Frank Sinatra‘ uploaded in 2020 to Soundcloud by Open AI ⁽¹⁾, together with a cover photo of Frank Sinatra ⁽²⁾ remixed using aging AI algorithms to their maximum slide value of ‘100 years old’.

I was inspired to release it as a physical record by some of the questions it provoked in terms of ‘style’ copyright, and data usage. Researchers used convolutional neural networks to encode and compress Frank Sinatra audio ⁽³⁾ . It seemed fitting that ingesting the ouvre of Frank Sinatra using this technique spits out the fragmented lyrics “human flesh for sacrifice”.

If Open AI were using copyrighted recordings of Frank Sinatra to produce endless Frank Sinatra style songs, then a pressing of their derivative work questions idea of subsequent copyright under fair use. The arguments are central to the future monetisation and current bankrolling of the AI everywhere juggernaut.

These have been contested ever since these tecnologies transitioned from nascent to commercially mainstream in just three years. Sean O’Connor (2022) points out that music generators using machine learning are capable of producing music that’s recognisably in the style of composers and performers, and therefore indicates that “style standing alone can be quantified and fixed” ⁽⁴⁾. He argues that musical style should be protected for several reasons, up to and including what he terms the “unjust enrichment of tech companies at the expense of human artists.”

Given how things have played out so far, the dulcet tones of Jukebox Frank crooning “human flesh for sacrifice” could be Open AI’s brand motto.

At each stage there is a physicality to the recordings, from the human energy of Frank Sinatra moving the diaphragm of a microphone to the computational energies of the model Frank. The way we encounter these artifacts and how they are framed influences our perception. If Open AI Frank escapes the experimental framing on Soundcloud, does this make for easier listening as a jukebox installation? How does human flesh for sacrifice make us feel emanating from a 7″ single playing out on a record deck?

Five Reliquiae

Hello Everybody

Back in the old days ⁽⁵⁾, the fidelity of text to image programs improved at an almost unnerving rate. They were fed vast amounts of other people’s data scraped from various sources by non-profits like Open AI, which were backed by investors and organisations with pockets that were billions deep. As models improved and new tools like MidJourney and Stable Diffusion also emerged, photorealism and style mimicry became common, mundane even. However, these tools often still struggled with human limbs, especially hands. Mistakes felt to some extent reassuring. Even as less obvious fake images circulated virally as real, alongside this there was a discourse that there were telltale signs that could be read by those sufficiently in the know, or tools that could tag images as ‘AI’. A missing or extra digit is nothing though for those trading in the speculation and dollars that use Arthur C Clarke’s third law, that “any sufficiently advanced technology is indistiguishable from magic”⁽⁶⁾ as a way to attract investment and steer where human energy goes.

The comfort of failure

If the definition of AI is the horizon of the things it can’t appear to do yet, the fiddleiness of hands seemed, for a long time to offer the holdout of a reassuringly human problem. When Patti Smith first posted an image online, she chose a photo of her hands with ‘Hello Everybody!’ as an Instagram post. A virtual signifier of the very human, for Smith, hands are “a direct correspondence between imagination and execution” ⁽⁷⁾. If a machine fumbles hands, then the human touch still feels, at least for a time, valid.

Like any artist starting out, there’s a tradeoff between looking at example of something and trying to copy it or looking to understand why something looks the way it does, by understanding things like anatomy and structure. The lack of actual understanding was particularly apparent with text to image models. This problem was attacked with more comprehensive datasets, and as specific techniques such as HanDiffuser which injects detailed hand parameters, such as 3D shapes and joint positions, into the model to ensure more realistic hand renderings to ‘align models more closely with human preferences’ ⁽⁸⁾. All this to overcome the fact that the models may have seen more images of hands than a human can in a lifetime. But they have never (yet) shaken many.

Five Reliquiae

In the beginning of 2025, I began to look back at this period of controversy and conversation of ‘firsts’ such as the inevitable winning of photography prizes ⁽⁹⁾, with a nostalgia for the discourse surrounding it. As many of the text to image models and techniques have improved and do ‘align models more closely with human preferences‘ ⁽¹⁰⁾ to the point where we cannot tell the difference, I wondered if I could create something that circled back to the genesis of the story.

I had previously made Jukebox Sinatra, turning an early machine learning derangement or rearrangement ⁽¹¹⁾ by Open AI back into the physical medium from which it was once contained and encountered as. In a similar way,

It feels a bit like the attributions made of some bits of things resembling body parts to saints that also handily help to attract money and prestige to a particular place of worship. Religious relics are purported to be the physical remains or personal effects of a saint or similar. These artfacts are preserved or created so they can be venerated as a tangible memorial. They are props that stand in as the physical proof of an old story.

Five is a series of five individual six fingered hands that in one way charts this. Silicone models made from text to image computational images, seeded in diffusion models that incorporate many many individual hands to give a sense, the seed of human handiness. Of course, some humans have six fingered hands. Or toes. Polydactyly runs in my family. And as the old saying goes, ‘five fingers good, six fingers better’.

Text to image models had begun to be developed in the 2010’s from advances in deep neural network techniques. Fast forward to the beginning of 2021, the trailing and release of Dalle-E, Open AI’s transformer based system captured widespread public attention as it began to approach a quality threshold nearer to to the human-made images Open AI had been scraping from the web and training its models with.

As the iterations of their model improved, and as other companies released accessible tools (such as MidJourney and Stable Diffusion), photorealism, or highly capable mimicry of different styles became standard. In these ‘early days’, whilst these tools were able to render realistic or stylistically accurate landscapes, famous faces, remixing what they had ingested, the running joke and the uncanny burps that were conjured up by the reality that these tools don’t have an actual understanding of anything per se really became apparent when they would, like anybody learning to look early on, in the difficulty of rendering human limbs, and most especially, human hands.

Inputting ‘Why do people find it difficult to draw hands and fingers?’ into Microsoft’s Co-pilot on 01/04 2025, it outputted: Drawing hands and fingers can be challenging due to their complex anatomy, varied angles, dynamic movements, and fine details. Practicing regularly and studying hand anatomy can help improve your skills. Some artists use simplified shapes to build the structure before adding details.

Inputting ‘Why do machine learning text to image programs find it difficult to draw hands and fingers?’ it outputted: AI and machine learning text-to-image programs often struggle with drawing hands and fingers due to the complex anatomy, variability in positions and angles, limitations in training data quality and diversity, and the challenge of translating three-dimensional structures into two-dimensional images. Despite these difficulties, ongoing improvements in AI algorithms and more comprehensive datasets are helping to enhance the accuracy of AI-generated images.

Related projects: NoiseGait, algorithmic surveillance gait detection evasion strategies using shoes with different heels.

Footnotes:

Jukebox Frank

1) Jukebox: A Generative Model for Music
https://arxiv.org/abs/2005.00341

2) Aged cover based on photo: Frank Sinatra: performing live onstage, waving, with audience behind. (Photo by David Redfern/Redferns) Sept 1st 1980 Royal Albert Hall, Getty ImagesImage #: 85002198

3) Pressing audio source: https://soundcloud.com/openai_audio/classic-pop-in-the-style-of-frank-sinatra

4) O’Connor, Seán M., AI Replication of Musical Styles Points the Way to An Exclusive Rights Regime (February 15, 2022). Research Handbook on Intellectual Property and Artificial Intelligence, Ryan Abbott ed. (Edward Elgar 2022), Available at SSRN: https://ssrn.com/abstract=4066915 or http://dx.doi.org/10.2139/ssrn.4066915

Five Reliquiae

5) DALL-E was revealed by OpenAI in a blog post on 5th January 2021

6) Clarke, A. C. (1968). Clarke’s Third Law on UFO’s. Science, 159(3812), 255. https://doi.org/10.1126/science.159.3812.255-b

7) Patti Smith, The New Yorker, Smith, P. (2022, November 10). Things I’ve seen. The New Yorker. Retrieved from https://www.newyorker.com/culture/culture-desk/things-ive-seen

8) and (6) Narasimhaswamy, S., Bhattacharya, U., Chen, X., Dasgupta, I., Mitra, S., & Hoai, M. (2024). HanDiffuser: Text-to-Image Generation With Realistic Hand Appearances. arXiv. https://arxiv.org/abs/2403.01693

9) Eldagsen, B. (2023, April 14). How my AI image won a major photography competition. Scientific American. Retrieved from https://www.scientificamerican.com/article/how-my-ai-image-won-a-major-photography-competition/

10) OpenAI released the SoundCloud recording titled “Classic Pop, in the style of Frank Sinatra” on May 5, 2020. This track is part of a series where OpenAI’s models generate music in the style of various iconic artists.