Hands up who has been putting to the test the patience of their favourite AI text-to-image applications, feeding bots with the most impossible prompts and then keeping on tweaking the generated images for hours. There's probably quite a few of us out there, but it looks like there are also rather unsuspectable people among us, exasperating ChatGPT – crafters, and in particular crochet and knitting enthusiasts and fiber artists.
Rather than asking the popular AI-based text generator to do their homework and write essays for school, for a few months now, crafters have been asking ChatGPT to produce patterns for knitwear and crochet designs.
TikTok is currently full of short videos about what can ChatGPT do for arty crochet and knitting projects. Some TikTok users only generated patterns, others went through the effort of actually creating the designs suggested by the AI. The results are rather surprising.
Earlier on this year, TikTok user and textile artist Alexandra Woolner (Generated Crochet account on TikTok), asked ChatGPT to produce the pattern for a narwhal amigurumi.
Her prompt was: "Write me a crochet pattern for a narwhal stuffed animal using worsted weight yarn". The application provided instructions that seemed to respect the language of the crochet pattern with abbreviations ("Ch" for "chain"; "sc" for "single crochet") and symbols (such as (*) indicating a repetition).
Woolner followed the pattern, but the results were disappointing and she called the final craft, "AI generated narwhal crochet monstrosity". Yet other users seemed to like the experiment: fans dubbed the little narwhal Gerald and the original video hit over 900,000 views (at the time of writing this post), with mixed comments and with some users highlighting how this was the proof that the AI wasn't good at generating art.
Woolner didn't desist, though, and prompted ChatGPT to generate more patterns, including one for a cat and another for an octopus that, she says, has "booby eyes". Other crochet enthusiasts on TikTok started suggesting her ideas for her patterns and even started buying her handmade renditions of Gerald. Woolner is also selling the downloadable pattern for Gerald for users who want to make their own nonsensical narwahl.
Other crochet fans on TikTok had very different experiences: Lily Lanario, a London-based crocheter, has been developing a series of amigurumi with patterns made by ChatGPT with mixed results. Lanario has so far done a convincing frog (with too many limbs, though, if you look at it from one side…), a very accurate rendition of Jigglypuff from Pokemon, a frighteningly scary version of Pikachu and an eyeless cat with an uncanny body.
ChatGPT reproduces the abbreviations standing for different stitch types, but tends to turn them into weird strings of codes that, once translated into stitches do not make sense. Apparently, there are specific reasons why ChatGPT produces inconsistent patterns and this is not because it received very few training data.
The coded language of crochet and knitting instructions should be easier to understand for an application using code, but it's not as simple as that.
In reality, AI applications struggle with patterns because there are numbers included, something that Artificial Intelligence finds harder to elaborate than words, as AI applications are trained on texts and not on numbers and they tend to predict which words come after each other (ChatGPT is a Large Language Model and, as you may remember from a previous post, LLMs are artificial intelligence tools that have the ability to read, summarise and translate texts and predict future words in a sentence by generating sentences similar to how humans talk and write).
To put it simply, ChatGPT, like many of us, is simply bad at maths (knowing that the AI struggles with ratios and numbers makes it suddenly more human to our eyes…). So ChatGPT creates the crocheting or knitting pattern trying to predict what may come after and eventually failing at it (would it become more precise if the AI app were trained with crochet patterns with complete words rather than abbreviations and numbers written down as words?).
Yet some of the experiments carried out by Lanario led her to prove that ChatGPT can actually rework and improve instructions, adding elements that it had forgotten in a previous version of the pattern (the pattern generated for the cat was missing its eyes, even though the instructions inconsistently mentioned the eyes towards the end; quite often when asked to generate amigurumi ChatGPT seems to forget to include the eyes or doesn't explicitly say where to put them, or suggests to stitch two parts of the body of an animal together in an incongruous way – yet these are minor mistakes as you can figure that out and reattach the parts where you think they should go…).
But stories about ChatGPT and crocheting have sparked a debate among experts as well: Margaret Wertheim, a physicist and a science author, and part of the team behind the Crochet Coral Reef project, highlights that ChatGPT doesn't understand crochet instructions because it can’t imagine they represent a physical object. Daina Taimina, a Latvian mathematician, retired adjunct associate professor of mathematics at Cornell University and author of Crocheting Adventures with Hyperbolic Planes, proposed maybe a more interesting and simpler solution - instructions given for the amigurumi animals done so far weren't good enough.
Giving specific instructions to AI apps generates indeed the best results (if you go on Midjourney's general chats, you realise that the most accurate prompts with precise adjectives, instructions about the light and information about ratios, produce very detailed images). Woolner then tried to improve the instructions for Gerald, but ChatGPT came up with more complex instructions that included more stitches than needed, something that produced an even more nonsensical piece than the previous one.
What’s remarkable and surprising in all the stories involving AI applications is the fact that they progress pretty fast: nonsensical patterns may indeed become more reliable little by little. ChatGPT has anyway progressed compared to previous systems and seems to be able to remember different parts in a crocheted animal (albeit the main body always seem to be a sphere of some sort, often with another sphere attached for the head).
In 2019, a bot called HAT3000 was trained to create crocheted hats. Despite having been trained with several hundred patterns, one day it created a piece that didn't make any sense. Its developer, AI experimenter and writer Janelle Shane eventually understood that - without having been trained on Taimina's crocheted model of non-Euclidean geometry, made increasing the number of stitches at a constant rate to create a hyperbolic plane - the system had started imitating it.
Trained on hat patterns that got bigger with each row, the AI grasped the concept of adding stitches and started producing hats that didn’t make sense. ChatGPT doesn't seem to have adopted this habit of adding stitches, even though in the second narwhal pattern it did add some stitches.
Yet, so far, we have looked at amigurumi patterns, finding it hilarious the way AI apps positions body parts and paws, but what about other types of patterns? Would artificial intelligence produce something better if encouraged to create a series of granny squares?
Blog Crochet Concupiscence tried to make an owl amigurumi that came out as a head with wings that resembled massive floppy ears. The blog crocheters also made a granny square with a pattern that proved extremely difficult to read and produced a kite shape with a raised ball at the center.
But which would be the applications of the patterns created by ChatGPT? Well, they could be created to experiment more and come up with a human-machine collaboration. Will AI take over fiber art? Too early to say, but expect some experiments will be showcases sometimes soon in art exhibitions (that will probably start having sections dedicated to AI-generated art).
In a way, it would also be possible to create incredible images of crocheted/knitted pieces with an AI text-to-image application and then work out a way to do the stitches by yourself, even though that would be a long and complex process that only advanced crocheters and knitters may want to venture in.
Whatever happens, you have a feeling that soon we will have pocket books about ugly crochet creatures made with AI. After all, Miuccia Prada made a living out of ugly chic, crochet artists collaborating with AI applications may try earning a living with hilarious or simply ugly amigurumi. Or what about redeveloping a Nintendo knitwear system that works with AI applications? The possibilities may be endless.
The twist in the tail of this story? None of us probably ever expected that the greatest adversaries, but also the greatest fans of ChatGPT, an application that has been worrying many teachers out there, would have been crafters and crochet, knitting and textile artists.
A final note: if you buy AI-generated patterns online always get them from certified sources as scammers involved in "get rich quick" schemes using AI may be trying generating faulty crochet patterns that would be impossible to make and you would realise only after buying the pattern that it's useless.
Comments