2027 Forecast

What a machine actually does with a deck

AI tarot reading

An AI tarot reading is fluent text, not a draw: a language model can write the reading, but it cannot shuffle. It has no randomness of its own, so when it says it has drawn three cards it is predicting plausible text, and plausible text leans toward the famous cards. A computer is genuinely better than a human at the one mechanical part, the shuffle, and no better at any of the rest.

What to draw instead, and how to read it: free tarot reading, three-card reading, yes or no tarot, Celtic Cross reading, card of the day, how to read tarot, all 78 cards, tarot spreads, Rider-Waite deck and Lenormand.

AI Tarot Reading: What a Machine Can and Can’t Do

The part a machine does better

Seven shuffles for playing cards, eight for tarot

How far from perfectly mixed a deck is after each riffle shuffle
Total variation distance after k riffle shuffles, 52 and 78 cards00.250.50.751123456789101112riffle shufflesdistance from a perfect mix (0 = perfect)78 cards, 1: 1.00078 cards, 2: 1.00078 cards, 3: 1.00078 cards, 4: 1.00078 cards, 5: 1.00078 cards, 6: 0.89378 cards, 7: 0.57178 cards, 8: 0.30778 cards, 9: 0.15378 cards, 10: 0.07878 cards, 11: 0.03978 cards, 12: 0.02052 cards, 1: 1.00052 cards, 2: 1.00052 cards, 3: 1.00052 cards, 4: 1.00052 cards, 5: 0.92452 cards, 6: 0.61452 cards, 7: 0.33452 cards, 8: 0.16752 cards, 9: 0.08552 cards, 10: 0.04352 cards, 11: 0.02252 cards, 12: 0.01178 cards52 cardsBayer & Diaconis 1992 formula, computed exactlyastrohot.com

After seven riffles a 52-card pack sits 0.33 from a perfect mix, while a 78-card tarot deck is still at 0.57; it takes an eighth shuffle to bring tarot down to 0.31.

Human shuffling is famously bad at randomizing. 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. Their formula, worked through for 78 cards, says a tarot deck needs one shuffle more.

A shuffle in code has no such problem. On the first pass each of the 78 cards has the same chance of every position: in our own test of 200,000 draws, every card landed first within 5% of the expected 2,564 times. 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 cards most likely to appear in that sentence, so famous names such as the Tower, Death and the Lovers are likelier picks than a fair one 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. A draw that reshuffles on every page load, AI or not, 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 →

Cite this page

AstroHot. “AI Tarot Reading: What a Machine Can and Can’t Do.” Accessed October 29, 2026. https://astrohot.com/tarot/ai/

For Wikipedia {{cite web |title=AI Tarot Reading: What a Machine Can and Can’t Do |url=https://astrohot.com/tarot/ai/ |website=AstroHot |access-date=2026-10-29}}