No AI
Every word here was written by a person
Not one sentence of this site was generated by a language model, and the deck is not weighted. Here is exactly what is written, what is computed, and how you can tell the difference for yourself.
Three claims, and how to test each one
Every word existed before you drew
The 78 card meanings, upright and reversed, the spread guides, the lessons and the combination pages were all written by a person, against the Rider-Waite-Smith deck and the Golden Dawn attributions it was built on. Nothing is composed while you wait, because there is nothing here capable of composing it.
Check it: Draw the same cards twice. The words match exactly.
The relationships are computed, and only those
What a reading adds is what it can see across the whole draw: elemental dignities between neighbouring cards, how many trumps came up, which suit dominates or is missing entirely, repeated numbers, Aces, court cards. These cannot be written in advance because there are sextillions of them. They are calculated for your spread from the written pieces.
Check it: Change one card and watch the observations about the whole spread change with it.
It refuses questions rather than answering them
No reading here will diagnose an illness, say when somebody will die, answer a question about a pregnancy, or advise you on a legal matter. It will point you toward somebody qualified instead. A model cannot hold that line reliably, because declining is not what it is for.
Check it: Ask it one of those questions and see what it does.
Why written beats generated, even though generated reads better
A language model will produce a more beautiful paragraph than we will. That is what it is optimised for, and it is exactly why the result is empty: it produces the same quality of beautiful paragraph about a card you did not draw. The fluency is unconditional. It never had to be about your spread and nothing in the process required it to be.
A written meaning had to be decided. Somebody had to take a position on what the Five of Pentacles is doing, know that Waite and the Golden Dawn do not entirely agree about it, and say which reading this site follows and why. That is a slower, plainer, more argumentative kind of writing, and it is the only kind that can be wrong, which is what makes it worth reading.
What we are not claiming
That tarot predicts the future, that a card knows anything about you, or that writing a meaning by hand makes it true. It does not. The claim on this page is narrower and it is one we can actually stand behind: the draw is honest, the words were written by somebody who had to think about them, and nothing is being invented while you wait. What the reading means is yours to decide.
See for yourself
The quickest test is to pull the same spread twice, then read a card page and see whether it commits to anything.
The sources the meanings were written against are in the editorial policy, the mechanics are in how it works, and the spreads are listed on spreads. Why the site exists at all is in about.
Questions about how this is built
The five that come up most.
None. No language model is called at any point, for anything. There is no AI package among this site’s dependencies and no request goes out to an AI service, because there is nothing to request. Every card meaning, spread guide, lesson and combination page was written by a person and stored. A reading selects from that writing according to the cards you drew.
No, and the arithmetic is why. Three cards drawn from seventy-eight, in order, with reversals, is about 3.6 million distinct spreads. A Celtic Cross is roughly 4.7 sextillion. The individual card meanings can be written because there are only 156 of them counting reversals. What happens between the cards cannot be, so it is computed for your draw: the elemental dignities between neighbours, the ratio of Major Arcana, suit dominance and absence, repeated numbers, Aces, court cards, the direction the numbers travel. That is where most of the meaning in a spread lives.
Draw the same cards twice and compare the readings word for word. They are identical, because nothing is written on the spot. Then try it on a site that generates readings: the same cards and the same question will produce different text every time, because the text was never anchored to the cards in the first place.
Considerably cheaper, which is the problem. A model will write you a fluent, moving paragraph about The Tower whether or not The Tower came up, and it will write a different one tomorrow. Tarot has real content: a documented deck, the Golden Dawn attributions, traditions that disagree with each other for reasons worth knowing. A system producing plausible text regardless of the draw is not reading tarot, it is finishing a sentence.
No. The Tower is exactly as likely as The Sun, and the shuffle is a real one. Tilting the deck to make the site feel better would be the same failure as generating the words: an outcome chosen for how it reads rather than by what actually happened.