I logged into Canva yesterday to make a cover for the winter playlist I am sharing today. (The playlist is at the end of the newsletter!) A pop up told me that Magic Write is now available on Canva. It promises to deliver,
Your first draft, fast
Looking for some words of inspiration to kickstart your creative process? Magic Write in Canva Docs is your very own AI text generator to help you get out a first draft, fast. Simply start with a prompt and watch as copy, blog outlines, lists, bio captions, content ideas, brainstorms, and more appear in seconds. - Canva
Magic Write is an AI text-generator created by OpenAI, the same artificial intelligence project that created the AI art generator DALL-E.
AI text-generators and art-generators are diffusion models. They’re artificial neural networks that function like a very simplified version of the human brain, processing information through nodes and synapses. The models are trained on billions of words and billions of images culled from the internet. Many of those words and images belong to people who did not give anyone permission to include them in any kind of model. Once grouped in the billions each word, art piece, photo of a 1st birthday party and stock image becomes part of a training dataset.
The models take apart each piece of data, learning about each word grouping and image through statistics. At each node, another bit of the text or image is destroyed again, until nothing but noise exists. When prompted by a user, the noise from the associated collection of images is removed and something is generated in the form learned during the destruction. Combinations and compositions that never existed before emerge, pushing us to ideas and visions we’d never reach without the collective.
And, despite what we’ve sometimes come to believe, the internet does not hold all of creation.
Well, that’s the idea. But the models only use data found online. And, despite what we’ve sometimes come to believe, the internet does not hold all of creation. And because of internet rot, it doesn’t even hold everything it once held. 72% of links from earlier than 1998 are dead.
1998 was just not that long ago! That was the year Bill Clinton said, “I did not have sexual relations with that woman.” I was thirteen and learned A LOT about sex because of that news cycle. Ahem. 1998 was only ten years before the Marvel Infinity Saga began! It was just yesterday! (And also forever ago, please stop making those movies.) And you don’t have to go back nearly that far to hit a high rate of link rot.
Even with internet decay, I know “billions” of images sounds like a lot. But it’s not really, when you consider the variety of human experience. And it’s really not a lot when you think about how variations of the same compositions clutter the internet because of viral trends.
I decided to give Magic Write and Midjourney prompts about influencers and mommy bloggers. I asked Magic Write to write instagram captions about influencer culture, mommyblogging and being a mom at Christmas time. The hash tags made me laugh.
But what the captions really helped me see - again - is how much of our content is already being produced by algorithms. There is a formula on Instagram and the algorithm awards those who represent it within those parameters. The captions were bland variations on the same conversation and captions we’ve been posting about influencing and moms since influencing became a thing. Which - by the way - was years AFTER The Infinity Saga started.
The images produced by my Christmas influencer prompts were similarly predictable. Every image was a woman - even when I didn’t specify gender. All but one image was a white woman. Of course, men are Instagram influencers. But men are rarely branded as “influencers” - so they’d not be processed that way by the neural networks. And of course, not all influencers are white women. But not every piece of data fed to the neural network has the same weight. And because platforms like Instagram have algorithms and users that privileged white women in that space, they’re going to take up more space in the neural network.
There was also a hint of Mary and Baby Jesus in all the compositions - a visual commentary on the way mothers are depicted online whether in influencing spaces or not.
Content over creation has hollowed out so many of our online spaces.
The text and images have me thinking again about what kind of content we need in our lives. I have only been able to begin a writing career because of the platforms built to support the content economy. Instagram was the first place I published my ideas. The results from the diffusion models aren’t just a testament to their inadequacies. They’re first and foremost proof that content over creation has hollowed out so many of our online spaces. These images are distillations of what there was to mine (illicitly!) online. And a lot of it is fool’s gold.
I am still researching diffusion models and AI generation. And I am still not against either, in theory. I like the technology that powers both Magic Write and Midjourney. I know that some of the words and images I’ve shared online have probably been gobbled up by each. That is wrong. And I think I should have been asked. But if, or when, someone creates a fully voluntary dataset for an open AI generator someday, I won’t wait to be asked. I’ll volunteer.
(Please, someone do that! Use public domain images and, for everything else, let people opt in!)
I’d submit my words to that future generator even though some of my work will be put at risk by things like Magic Write, a few versions from now. Especially once the paraphrase function comes online. (Although, I’ve definitely heard my work paraphrased without credit on several pretty big podcasts just this past year. So I don’t have to wait for Magic Write to get ripped off, hahaha AHHHH!)
Let me tell you why.
I was in the middle of figuring out how Magic Write outlines ideas when I started to cry. I’ve seen my daughter with multiple processing differences stare at a blank page in true horror. I’ve seen her approach a page full of words that way too. She just does not process writing or the written word the same way many other people do. Nothing is written for her, so she feels disordered. But she’s not. She just processes differently.
What if someone built a diffusion model for people with processing differences, like dyslexia? And my daughter could prompt the AI with a paragraph she doesn’t understand along with the setting /dyslexia? What if she could ask a diffusion engine to write the mechanics of a process - an equation or experiment - in a way that made sense to her. What noise could the neural network strip before returning something written for the way she reads? I’d give all my words to train a model that could act as an adaptive aid for her and others.
I realized I have aphantasia. I can’t visualize images in my head. Everything is words. There are no pictures.
I have processing disorders too. I’ve known about two of them for a while. But then, in the midst of researching AI art generators for a future newsletter, I realized I have aphantasia. I can’t visualize images in my head. Everything is words. There are no pictures. I thought when people said they could “see” memories and pictures in their heads they were speaking metaphorically. But where I only have a narrator in my head, apparently 98% of the population has some sort of projector.
I just realized this yesterday, and honestly, my stomach kind of sank. One more thing my head does differently. It’s a little unsettling to have not known about it for so long, you know?
Maybe that’s one reason I’ve found myself up so late at night, learning how to prompt Midjourney. I can’t seem to stop revising and rewording until the generated image matches the outline of the words in my head. It is something, isn’t it, to have a thought connected to an image in seconds? I didn’t even know I was missing it. An ethical dataset and ethical application of artificial neural network technology could help me see in ways I didn’t know I couldn’t see.
Of course, that’s what could be. Not what will be. It’d require all of our better angels to create a regenerative commons. All of our better angels and a different economic system. I’m going to hope and work for that until I’m disappointed, and probably hope and work for after too.
In the meantime, I think Magic Write and its ilk will become dominant enough in our online spaces to show all of us how hollowed out those spaces have become. And then what?
I think part of creation is supporting creators.
Maybe, we’ll stop accepting free content and start paying for real creation. I know that’s part of the bet I am making with this newsletter. Maybe another silly hope.* But it’s one I am practicing. I think part of creation is supporting creators. So I pay for over a dozen newsletters myself. I’ve also started paying for music instead of just streaming it. Like I am in high school again. And I’ve been putting away money to buy little bits of original art. I most recently purchased this piece from Elizabeth Sanchez.
Churning Twitter and Instagram content turned support into a streaming economy of likes and follows. And, frankly, I am finding my focus on supporting creation before content feels slow, and even limited, compared to what I’ve grown used to. But it also feels intentional and impactful. And I don’t know…lasting? Like link rot won’t touch it. Even twenty to thirty years after the last Marvel saga lights up the screen.
Like even when the stuff we created is fed into a machine, it won’t be reduced. It can’t be reduced, because the connection was part of the composition.
Here’s a little winter playlist. It’s mostly songs about being lonely or cozy or lost or all three.
Here’s the Apple playlist. I think it’s superior because it’s got Joni Mitchell on it. Her music isn’t on Spotify.
But I made the list sans Joni on Spotify too! Just imagine she’s in there, if that’s the platform you prefer.
And yes! I did buy each available song on my private account when I made this playlist for us. Because…I want to support the people lulling us into our frostbitten sleep!
