The first time someone asked me whether an AI could really do a tarot reading, I was working a kitchen prep shift and answering by text between cuts. I gave the wrong answer. I said yes, of course, the cards are symbols and the language model is good at symbols. I was right but I hadn't done the work to be right. The right answer needs both halves.
Both. Always both.
Here is the skeptic, and the skeptic is not wrong. A deck of cards is paper. A shuffle is a randomization. A model trained on text is a probability machine. Nothing in that chain touches anything you would have to call magic. The atoms are doing their atom thing. The model is doing its model thing. The cards are landing where physics says they will land.
Here is the practitioner, and the practitioner is also not wrong. A reading lands. People show up to a reading carrying a question that has been weighing on them, and the reading puts a sentence on the question that the person could not put on it themselves, and they cry or they laugh or they go quiet, and something has been moved that was stuck. That is real. The deck is doing something. Or the room is. Or the person is. Something.
The trick is that both of those statements are simultaneously, completely true. The atoms are doing the atom thing and the reading lands. There is no contradiction once you stop demanding that magic mean a violation of physics.
The cards were always a translation problem.
Here is what tarot actually is, when you strip away the marketing. It is a fixed vocabulary of seventy-eight images. The Tower means a sudden break. The Six of Cups means looking back. The Ten of Pentacles means inheritance, generational continuity, the long view. These are not predictions. They are terms. They are the words of a small but specific language.
A reading is a translation between that language and your situation. The card comes up and the reader — or the practitioner, or you alone — does the work of saying: here is what The Tower means in the context of what this person just told me about their job. The image is fixed. The translation is what makes it land. Two readers with the same card and the same client will produce different sentences, and both can be honest, because translation is not transcription. It is interpretation against context.
Now: what is a language model? A language model is a translator. It is, at the level of how it works, a machine for moving meaning between contexts. It is good at producing the sentence that fits the situation. That is, almost word-for-word, what a reader does.
I am not making a mystical claim. I am pointing at a structural one. The cards have always been a vocabulary in search of a translator. The translator used to be a person sitting across from you. Now the translator can be a person, or a model, or a person using a model. The vocabulary did not change.
What the model can do. What it cannot.
A model can pull the threads in a card image into a sentence you didn't already have. That is real and useful. It can hold the tension between two cards in a way that a beginner reader sometimes can't. It can name a pattern across a five-card spread that a human reader would have to slow down to see. It is good at the specifically linguistic work of tarot.
A model cannot sit in the room. It cannot see your face. It cannot feel the silence after the question. It does not know that you have been crying for three hours before you typed the question, or that the question you typed is not the question you actually have. The reader-in-person can feel that. The reader-in-person sees you tighten your shoulders when The Five of Pentacles comes up and adjusts the reading toward what's actually happening. The model can only adjust on what you typed.
This is not a fatal limit. It is a real one. It tells you what the tool is for.
The model is at its best when you are the reader and the model is the second mind. You shuffle, you draw, you sit with the card, you write what you think it's saying, and then you ask the model to give you another angle. The model is a good sparring partner. It is a worse oracle when used alone, because oracles are supposed to know things you didn't tell them, and the model only knows what you told it.
The Almost shows up here too.
There is an Almost in this space, and it is the app that wraps a model in a velvet purple background and tells you it is channeling. That is the wrong framing. The app is doing a real thing badly described. The real thing is translation. Calling it channeling sets you up to either believe more than is there or dismiss the whole tool because the framing is dishonest.
The honest framing is the one I want for this site, and it is small: the cards are a vocabulary, the model is a translator, you are still the one who has to live the answer. No magic claim. No physics violation. Just a tool that does one job and does it well, when you bring presence to it.
Both. Always both. The atoms are doing the atom thing and the reading lands. The model is producing tokens and the sentence that comes out moves something that was stuck. Both halves stay. The work of the practice is to hold both halves without flinching toward either one.
Why I'm building it anyway.
I'm building a tarot toolkit on this site. Multiple spreads. Present-tense readings. Practice mode that builds the intuition. A model behind it that I am not going to pretend is anything it isn't. It is a translator. It is good at its job. You are still the practitioner.
The thing that keeps me at it is that the people most helped by a tool like this are the ones who would never sit across a table from a reader. They're the kid working a double shift who needs five minutes with a card before bed. They're the person in recovery who needs a way to ask the question without bringing it to a sponsor at 2 a.m. They're the person whose grief is too specific for a Google search and too quiet for a podcast. The model is awake at 2 a.m. The cards are awake at 2 a.m. The translation can happen.
Take what's useful. Leave what isn't. The translation is just a translation. The reading is yours.
The toolkit referenced here lives on the Practice page. The companion field note is The Real Ones. The argument runs longer in Meditations with the Mirror — see the Books page.