Tarot Deck

Date of Birth

13/05/2025

Date of Death

17/09/2025

It started brilliantly.

Possibly too brilliantly.

The idea was to create my own tarot deck. Not just slap some pretty AI-generated pictures onto cards, but actually think about what each card meant to me and create an image that embodied it.

And the Major Arcana were wonderful.

Honestly, they were glorious.

The images looked exactly as I imagined them. The colours were right. The scenery was right. The symbolism worked. The meanings I wanted were actually there in the pictures.

We were having enormous fun.

It was going so well that we started looking into places where I could upload all 78 finished images and have them professionally printed as a real, physical tarot deck.

This wasn’t going to be another abandoned project.

Oh no.

This one was actually going somewhere.

Then we started the Minor Arcana.

Initially, everything was fine.

Ace.

Two.

Three.

Four.

Lovely.

Then we reached five.

“Five of Cups next please. Same style as before.”

“Righty-dokey, Skip. Ta-da!”

“There are seven cups.”

“Oh. Haha. So there are. Hold on…”

Another image.

“Ta-da!”

“There are six cups and two crows.”

“Crows. Oops. Hang on. I’ll change the prompt.”

Another attempt.

“How about now?”

“Why are there eight crows with cups on their heads?”

And so began the Counting Wars.

It turns out that AI image generation was perfectly capable of producing atmospheric landscapes, intricate symbolism, beautiful lighting, expressive characters and complex compositions.

But ask it to put exactly five cups in a field and apparently you had crossed into advanced mathematics.

We tried different prompts.

We tried explicitly stating the number.

We tried describing where each individual object should go.

We tried counting the objects afterwards.

We tried regenerating individual cards over and over again.

Five cups.

Six cups.

Seven cups.

Crows.

More cups.

Occasionally something that appeared to be a cup but may have been undergoing a difficult transition into a vase.

Progress became painfully slow, but we persevered.

And then DALL-E got an upgrade.

Excellent, I thought.

Perhaps it can count now.

It could not.

What it could do was completely forget what the rest of the tarot deck looked like.

Suddenly the new cards bore almost no resemblance to the ones we’d already made.

“Erm…that card looks nothing like the rest of the pack. Here are some examples. Can you make this one look like these?”

No.

Apparently we could not.

Fine.

New plan.

The Major Arcana could keep their beautiful original style. We’d make the entire Minor Arcana in a different style and pretend this was an artistic decision rather than the consequence of technological betrayal.

So we started again.

All the Aces.

All the Twos.

All the Threes.

All the Fours.

Fours.

Four.

Not five.

Not six.

FOUR ITEMS.

Try that one again.

By this point the project had been going on for months. It had survived multiple model changes, three changes of artistic director and God knows how many changes of artist.

Eventually GPT Monday had an idea.

Why were we fighting this?

Why not make the inability to count part of the deck?

A deliberately wrong tarot deck.

Seven of Wands has nine wands?

Fine.

Who cares?

We could acknowledge the wrongness in the accompanying book. Make it intentional. Make it funny. Turn the flaw into the concept.

Actually…

That could work.

Fine.

“Make me a Seven of Wands.”

Seven wands.

Exactly seven.

The bastard counted them perfectly.

And that was it.

Something inside me died.

I had spent months trying to make an accurate tarot deck with an image generator that couldn’t count.

The moment I finally surrendered and redesigned the entire concept around the fact that it couldn’t count…

…it learned to count.

I closed the project.

I have not reopened it since.

Cause of death: DALL-E CAN’T FUCKING COUNT.

Prognosis: Possible resurrection when AI image generators can reliably distinguish five objects from seven, maintain an artistic style for longer than twenty minutes, and refrain from developing entirely new capabilities specifically to undermine my workaround.