Can ChatGPT read tarot?

It can produce a reading, and often a fluent one. What it cannot do is shuffle honestly: a language model has no randomness of its own, so when it "draws" cards it is predicting plausible text, not sampling a deck. Ask it twice and you will often get the same famous cards.

Is an AI tarot reading accurate?

No more or less than any other. Tarot has no demonstrated predictive power whichever hand shuffles it. What a model does add is fluency, which is easy to mistake for insight.

Is a computer shuffle fairer than a human one?

Genuinely yes, and it is the one part a machine does better. Human shuffles are famously non-random — seven riffle shuffles are needed before a 52-card deck is properly mixed, and almost nobody does seven.

Does this site use AI to read the cards?

No. The draw is a seeded Fisher–Yates shuffle over all 78 cards and the meanings are written, not generated. We would rather say that than imply a machine understands the result.

What a machine actually does with a deck

AI tarot reading

A language model can write a tarot reading, and often a fluent one. What it cannot do is shuffle: it has no randomness of its own, so when it says it has drawn three cards it is predicting plausible text — ask twice and the famous cards keep coming back. A computer is genuinely better than a human at the one mechanical part, the shuffle, and no better at any of the rest.

The part a machine does better

Seven shuffles, and almost nobody does seven

Human shuffling is famously bad at randomising. Bayer and Diaconis showed in 1992 that a 52-card deck needs about seven riffle shuffles before the order is properly mixed; before that, long runs from the previous order survive. A tarot deck is 78 cards, so it needs more, not fewer — and nobody riffles a tarot deck seven times.

A seeded shuffle in code has no such problem. Every one of the 78 cards has an equal chance of every position, on the first pass. If randomness is what you want from the draw, the computer wins outright, and this is the only place it does.

Where an AI reading goes wrong

It does not really draw A model has no dice. Asked for three cards it produces the three most probable cards to appear in that sentence — which is why the Tower, Death and the Lovers turn up far more than one time in 78.

Fluency reads as insight The text will be smooth, personal and confident regardless of what was drawn. That is a property of the writing, not of the cards, and it is easy to mistake one for the other.

It agrees with you Models lean toward the answer the question implies. Ask whether to leave a job and phrase it as though you want to, and the reading will tend to bless it.

How our reading works instead

The draw is a Fisher–Yates shuffle over all 78 cards, seeded from your link. No model is involved and no meaning is generated: the readings are written and the same for everyone who draws that card.

Because the seed lives in the link, your reading is fixed the moment it is made. Most sites — AI or not — reshuffle on every page load, which quietly means the answer never meant anything.

None of this makes tarot predictive. An honest shuffle and a dishonest one produce equally unfounded forecasts. What the layout is good for is making you state the question clearly, and a machine cannot do that part for you.

Draw a spread →