I used AI to analyse fashion reviews and got poetry instead!

In recent years I’ve become increasingly drawn to making text-based artworks. My first experimental series was called, Prints Based on Graffiti I Have Actually Seen. These prints feature odd, often misspelt phrases, tags, or ‘calls for help’ that I have seen sprayed and marked across blank walls, garages and public toilets – the weirder the better. Subsequently, I have ‘collected’ and used text from a wide range of sources to make my print-based works: those ubiquitous, trite self-help phrases you find pasted in so many office cubicles; subject headings from junk email, promises outlined on leaflets of obscure churches, as well as threatening (and sometimes lewd) text generated by Russian bots or hackers.

I am drawn to these texts partly because we don’t and can’t know the authors. The contexts in which the graffiti artist cries for help, Professor Makuba promises to ‘remove voodoo’ and ‘Твой кошмар I HACKED YOUR DEVICE SOPHIE’ mean these authors are secret, disguised or in some way performed. In the case of the self-help memes, the text is so banal and at the same time so extensively circulated that its relationship with its author is entirely untethered. These texts also excite me because of their fragmentary, part-communication and miscommunication. There is a chasm between intended audience and actual audience which makes meaning hard to access, a bit like Surrealist or Beat cut-ups. Perhaps these found texts tell us something about collective despair or desire; they are simmering and unauthorized by-products of an anxious twenty-first century psyche.

I was recently invited to ‘review the reviews’ for Louis Vuitton’s A/W 2020 Paris fashion week show for Warehouse Review, “a publication that aims to explore alternative modes of fashion criticism”. It seemed a good idea to use these techniques. But the dozen or so reviews I looked at from traditional print and online sources were too well-written, they communicated with the audience too efficiently and economically, they were too rational. There was no threat, weirdness or implausibility, no fissure between author, content and context and reader, no uncomfortable guessing at meaning. Could these reviews be ‘detourned’ somehow to reveal some collective angst, or piercing critique of the spectacle of Nicolas Ghesquière’s collection in the shadow of the oncoming Coronavirus crisis? In the spirit of Surrealist automatism I turned to artificial intelligence to automate a review for me, based on the existing reviews. This would function somewhat like the Surrealist game of ‘exquisite corpse’ where individuals write or make images based on a fragment of someone else’s work. Instead of attempting to access the human unconscious mind I would access the (by its nature unconscious) ‘mind’ of an artificial neural network, trained by colossal internet datasets and Silicon Valley developers. Surely this would tell me something either about the fashion show review, or at the very least, the dark recesses of the internet.

I used an online tool called Inferkit that, ‘takes text you provide and generates what it thinks comes next, using a state-of-the-art neural network’ called Megatron-11b. I was attracted by the developer’s suggestion that the text generator be used for, ‘writing stories, fake news articles, poetry, silly songs, recipes and just about every other type of content’. The developer of the tool warns that ‘the generator may produce offensive or sexual content’ because, of course, its learning is based on so much extant web content. This then seemed to be the perfect tool to create a new, irrational and even perverse review of Nicolas Ghesquière’s historic moment in the Louvre.

However, I soon became aware of the limitations of the tool. It can only comprehend 3000 characters at a time, so I wasn’t able to feed in those dozen or so reviews as ingredients for a new one. Using an ‘auto-completion’ function on a couple of paragraphs showed some promise – but too often the generated text seemed like a fairly plausible, if boring interview, movie or product review, with none of the unexpected nature of the found texts that so intrigue me. In some ways they were too good. According to James Vincent of tech news website The Verge, “at a fundamental level, the system doesn’t understand language or the world at large. The text it generates has surface-level coherence but no long-term structure.” What I needed was to find a way to exploit the tool’s limitations and make it disfunction.

Naturally, describing garments is a central part of a fashion show review. The described garment is often coupled with an adjective, sometimes two or more. The Louis Vuitton reviews contain technical descriptions of garments such as, “transparent, chiffon trousers” or “jewel-encrusted boleros”. Sometimes adjectives are used more poetically or as qualifiers for the garment, exemplified by phrases such as, “playful, feminine flounce skirts “or “beautiful bullfighter boleros”. There are other expanded noun phrases in the reviews that describe the atmosphere or mood of the show. Phrases such as, “time-collapsing sensation” and “hauntingly cyclical illusion” describe a kind of affective sensual or emotional ‘scape’. I decided to see what Inferkit would do with a selection of these phrases. Would these more fragmentary textual glimpses into the Louis Vuitton show encourage the generator into a different writing style?

With a few trials and tweaks I found that feeding just these descriptive phrases into the text generator tool jolted it into a more poetic mode. It generated lists, prose, direct addresses to the reader, text speak and advertisements as well as what could be described as the repetitive stutters of a collapsing machine mind or maybe some meth-addled poet. The texts are like modern American poems that riff on road trips, photography and pet grooming products. Others list garments or looks that would not be out of place in a real fashion show. There are beautiful and unexpected turns of phrase and wonderfully inappropriate juxtapositions. Yet, they are opaque, inaccessible – a pastiche of poetry.

You’ll have to wait for publication of Warehouse Review to read the complete set of poems but I have included one below as a taster. I have called them ‘Completion Poems’, poems started by found phrases from the Louis Vuitton A/W20 reviews and completed by the Megatron-11b artificial neural network. I am confused about my role in all of this. Perhaps I could be described as editor, selecting the phrases to enter into the text generator, selecting the best ‘poems’ and working only lightly on the punctuation. But that doesn’t seem quite right. It really is very weird to see a whole load of text generated in front you, that’s mostly not of you, but it’s not really of anyone else either. It’s a prediction generated by maths.

Using AI to generate text or ‘content’ is no longer the domain of programmers and engineers, it is becoming everyday reality for many media organisations and individuals. It is no surprise that the demand for ‘content producers’ and the ‘content industry’ is growing whilst the need for critics, journalists and journalism diminishes. One advertisement for an AI text generator claims that its “articles are usually very close to what a cheap article writer would produce, and you can publish them after only a few minutes of editing.” Another ad claims “up to 80% time savings” for AI-powered content generation. In a world where revenue is based on clicks, internet traffic and volume – quality and originality play second fiddle to speed. This echoes the relentless cycling of the fashion seasons, fashion as content and fashion houses as production companies – maximising brand visibility through the mass circulation of spectacle.

This article is not a lament for the dying art of the erudite critic, or an alarm call for the redundancy of human work. I neither fetishise or mourn how well AI generated ‘content’ imitates ‘real’ human output – that is beside the point. I don’t care about the real/simulated boundary – it’s really not the question here. The ‘poems’ demonstrate AI’s limits as well as its potential as a tool for curious writers and artists. This is not the death of the author, it’s the birth of hybrid, posthuman authorship and will require electronic and meat neural networks to create the fantastic, absurd and wonderful. I’d really like to see Nicolas Ghesquière’s next collection feature crop tops in flurries of lace, cross-dressing robot nurses, Belgian flower-children and tongue-in-cheek goggles. Let’s hope he remembers to “apply product to the outside of the animal’s leg” too!

COMPLETION POEM III

Generated using the phrases “Breathtaking tableau vivant”, “jarringly disconnected” and “ecclesiastical dresses”.

Breathtaking tableau vivant,

jarringly disconnected ecclesiastical dresses,

crop tops in flurries

of lace, frills of gold.

Pillowed sinew

pea coats and tartans too,

canary feathers, coronets

and petticoats, bustles.

The white sandals. I like this scene. This picture. This picture.

Bereft of so many dresses to keep them in frame —

scattering into the distant hills — at the last two moments the girls in crinolines make a sudden return to pin curls, to thread their foot in heels.

I know, we’ve seen this piece before.