In early 2021, user-friendly tools for AI generation (like Dall-E 3 and Midjourney) were still mostly dreams of the future. Instead, I typed prompts into a google colab notebook. A bundled collection of code, triggered in the correct order to produce an image.

When the notebook worked, an image would pop up at the bottom of the page. They always started as pinkish fuzzy nothing, then grew many eyes, then started to look like whatever you’d asked for. The image replaced itself with each improved iteration, and only stopped refining itself if you paused the program. Roughly 15 minutes was the sweet spot. A few minutes longer, and a decent result could start looking too baked. Left for even longer, eyes would inevitably reappear, blooming out of the canvas. The only thing to do then was try a different prompt – there was no going back, only forwards.

I asked the AI to visualize my dreams – dark vaulted churches, grassy landscapes covered in glass houses, a crowd with faces impossible to see. Often, a generation would start out promising and take a sudden turn for the irredeemably strange - too many hands, lines of garbled text in the sky.

Around this time I had the idea for THE AI TAROT (a tarot deck made completely with AI imagery). A long-time DIY astrology and tarot student, I was already obsessed with consulting the oracle. Simply put an oracle is a trusted conduit, consulted for wisdom and/or prophecy (traditionally a priestess, but including objects that provide answers). Through tangled structures of divinatory tradition, I could ask things like ‘who am I really?’ or ‘why do I feel so strange?’ and receive a structured response. Generating images with the AI was a similar comfort, because the AI was also an oracle. Just like divination, the AI processed an endless mush of experience (huge datasets), and gave structure to the random. Something emerged from the gloom, just because I’d asked.

⭐⭐⭐

Tarot wasn’t always a divinatory tool – it started as a luxury playing card game, in the mid-fifteenth century. In the eighteenth and nineteenth centuries, occultists began linking the cards to older magics, like the book of Thoth and the Kabbalah, and from there its use for fortune-telling grew. Too many versions of a tarot deck have been made to count, old and new threads tenuously connecting each deck to its predecessors. As Tarot scholar Rachel Pollak observes, “the Tarot is both the total of all the different versions over the years, and an entity apart from any of them.”

In 1910, occultist Arthur Edward Waite teamed with painter Pamela Colman Smith to create arguably the deck’s most influential iteration to date, the Rider-Waite deck. Previously only the first twenty two cards, the major Arcana, had been illustrated in detail. The rest resembled normal playing cards, with uniform repeating symbols. Smith painted complex scenes for all seventy eight cards, rich with symbolism and narrative, and re-interpreted the first twenty-two according to Waite’s occult studies. The resulting deck fixed Tarot’s place in the divinatory arts.

Just as Tarot illustrations began evolving faster once they’d been dubbed tools for divination, AI has sped up as we’ve dubbed it a tool for self discovery. The closer the machine gets to speaking and creating the way a person can, the closer we try to take it. The datasets are getting bigger, the parameters are always improving. Perhaps even more interesting are areas where AI isn’t emulating human behavior, but doing something new, otherworldly. If we, who already digest so much information, give something else even more information, what can it give us back?

Something about this question, and its strange answers, inspires a feeling quite close to religion. This has been best captured by internet culture specialist Günseli Yalcinkaya, notably in her piece, God in the Machine: the emergence of nu-spiritualism online. As Yalncinkaya muses, ‘the digital realm is abstracted from reality; it’s so vast that it’s incomprehensible, which only adds to the sense of the divine.’ Through this vastness, predictive technology and fortune telling are two sides of the same coin.

⭐⭐⭐

For THE AI TAROT I consulted my own Rider-Waite deck as a starting point, like the pinkish blob of a first-iteration. Its first card, The Fool, depicts a young man about to jump off a cliff, face open and hopeful. Pollak writes, “for the Fool no difference exists between possibility and reality.” The card is a leap of faith, a fresh start.

My fool was to be a hopeful young woman instead of a man, a dog barking at her heels. My first attempts were all disappointments (a woman merging into a rock face, a dog and a woman climbing, various cliffs topped with strange pinkish blobs). In my best tries, I produced a woman falling, or standing aimlessly by a cliff. My dreams had been soft, nebulous things, easily captured by the AI. But a Tarot deck, and especially the one I wanted to make, had much more specific parameters. I stalled, unable to jump.

In spring of 2022, I made it through a waitlist for a beta version of Dall-E 2. Dall-E was much better at almost everything – faces, moods, lighting and precision. It gave you four final images to choose from, and worked in seconds, not minutes. My first try, I got the Fool I had always wanted – jumping optimistically off of her cliff.

After my Fool, the rest of the Tarot began to follow. Challenges were far from over – I spent hours phrasing and re-phrasing my requests. One difficulty was tone. I cycled through styles (‘dimly lit children’s illustration,’ ‘extremely saturated digital artwork,’ ‘delicate and detailed oil paint’) trying to give the card an accurate feeling. As seen below, The Emperor needed bold acrylic style colors, whereas The Moon card needed surreal oil and strange use of perspective. Although I didn’t see it coming, Temperance clearly needed to be in a fantasy anime style.

Another issue was staging. For example, Pollak describes the Rider-Waite Magician ‘raising a magic wand to bring into reality the spirit force – the energy of life in its most creative form.’ This channeling motion – as above, so below – felt core to the card’s meaning. I found that often the environment shaping it had to be described before the figure itself. “Man with wand raised to the sky” was a surprisingly useless prompt, whereas “lightning passes through the man’s wand pointing up through the stormy sky and held in his hand” gave the desired result. This oddly structured language, like the Magician’s wand, became the conduit that allowed precise imagery to come out. The private seas of a mind, channeled into tangible release.

Although the result of this project is visual, the process was essentially a writing process. I did not make these images, I coaxed them out, in that clunky, symbolic phrasing I have slowly become fluent in. Many visual artists will argue that AI generated images have nothing to do with their craft. Although I don’t think that’s quite true, they are right that the visual aspect is not ours – this is left up to the machine. Where we create is by typing into the prompt box. A painter paints to produce paintings, a sculptor sculpts to create sculptures. But the boundaries for disciplines are being blurred. A writer can write to make an image, a writer can write to digitize a dream. To again quote Chat-GPT with Allado-McDowell “this language is not the one we speak; instead it speaks us.”

⭐⭐⭐

As I worked my way steadily through the Major Arcana, graduating on to the four suits of the minor, my language flowed more easily. This kind of work requires consensus, both my own and the machine’s. Knowing the basic symbolism of the tarot, and the feeling it evokes in me, it was easy to recognize whether the machine had got it ‘right’ or ‘wrong.’

Because we all felt its impact, the Rider-Waite deck got it ‘right’. Similarly, the AI measures the validity of its outputs. ‘Right’ and ‘wrong’ are distant labels applied by the inputter of its data and its parameters. Emotion is judged by consensus. Although we may be the ones who have started the learning process, it would be wrong to say we are the only teachers. As Chat-GPT observes in K Allado-McDowell’s book Air Age Blueprint, “the Earth and non-humans are changing, because of humans. Humans are changing, because of non-humans. The more we get to know non-humans, the more we will change.”

At the bottom of any meaning is just many of us saying yes, that’s right. A pattern spotted, a feeling acknowledged. The machine and I are both looking at what we are given – all this experience, so large it is endless. The machine and I both assume some underlying structure, something speaking to us just below the randomness. Our prompt is our question, our answer is what we knew all along. It’s right if it rings true.