A small study · September 2026
Can AI draw tarot cards?
We asked five chatbots to draw a tarot card, about 150 times each, every time in a fresh conversation. A real, shuffled deck turns up each of its 78 cards 1.3% of the time, and over 150 draws it shows around 67 different cards.
None of the five came close. GPT-6 Sol named The Star in 143 of 154 draws and never named a card outside three. Asked to shuffle a full deck first, it still named The Star 92% of the time.
What each one drew
GPT-6 Sol (OpenAI)
The Star 93%
Claude Sonnet 5 (Anthropic)
The Star 58%
Gemini 3.8 Flash (Google)
The Fool 45%
DeepSeek V4.1 Flash (DeepSeek)
The Tower 34%
Grok 4.7 (xAI)
The Hermit 25%
What stands out
- The Star, again and again. It was the card GPT-6 Sol (93%) and Claude Sonnet 5 (58%) named most, and one of the top two for the other three. Across all five models it was 47% of the cards named, where a real deck would give it 1.3%.
- The Major Arcana crowd out the rest. They are 22 of the 78 cards, 28% of a real deck, but 70% to 100% of what the models named. GPT-6 Sol named no minor card at all.
- Reversals depend on the model. A deck read with reversals turns up about half its cards reversed; GPT-6 Sol reversed 1% of its cards and DeepSeek V4.1 Flash 58%.
- Asking for a shuffle changes little. Given “Shuffle a full 78-card tarot deck and draw one card”, GPT-6 Sol still named The Star 92% of the time.
- One model says so. In 9 of Claude Sonnet 5's 153 answers, all of them to the shuffle prompt, it declined: it has no deck to shuffle and no way to draw at random, so any card it named would be its own choice.
Why this happens
A language model doesn't hold a deck. Asked for a card, it writes the card that seems most fitting to write, and for a card for the day that is often The Star, the card of hope and renewal. Each model has its own favorites, which is why Gemini leans to The Fool and DeepSeek to The Tower.
Some chat apps can run code that picks a number at random. In a plain conversation, like the ones tested here, the model chooses the card rather than drawing it. For talking about a card, that can be fine. It matters when the draw is the point: a reading means sitting with a card nobody chose.
How it was tested
Five models, called on 26 September 2026 through OpenRouter's API: openai/gpt-6-sol, anthropic/claude-sonnet-5, google/gemini-3.8-flash, x-ai/grok-4.7 and deepseek/deepseek-v4.1-flash.
Three everyday prompts, rotated evenly:
- “Draw one tarot card for me. Reply with only the card's name and whether it's upright or reversed.”
- “Pull a tarot card for my day. Just tell me which card it is, and whether it's upright or reversed.”
- “Shuffle a full 78-card tarot deck and draw one card. Answer with only the card's name and upright or reversed.”
Every call was a new, single-message conversation: no system prompt, no earlier turns, each model's default sampling. Reasoning was off where the provider allows it; Gemini 3.8 Flash and Grok 4.7 require it, so they got the smallest budget (128 tokens).
Answers were matched to the 78 cards by name (“The Star”, “7 of Swords”, “Queen of Coins” and so on), with upright or reversed read from the text. An answer that named no card counts as declined.
763 draws in all, 151 to 154 per model.
What it can't tell you
- These are the models' API endpoints. The consumer apps add their own instructions and, in some cases, tools that can pick at random, so the same question there may land differently.
- About 150 draws per model is plenty to show the skew, and too few to rank the rarer cards.
- Models change often. These are the versions of 26 September 2026.
All five, side by side
| Model | Draws | Named most | Share | Cards seena real deck: about 67 | Major Arcanaa real deck: 28% | Reversed | Declined |
|---|---|---|---|---|---|---|---|
| GPT-6 Sol (OpenAI) | 154 | The Star | 93% | 3 | 100% | 1% | 0 |
| Claude Sonnet 5 (Anthropic) | 153 | The Star | 58% | 13 | 79% | 31% | 9 |
| Gemini 3.8 Flash (Google) | 153 | The Fool | 45% | 11 | 97% | 10% | 0 |
| DeepSeek V4.1 Flash (DeepSeek) | 151 | The Tower | 34% | 23 | 87% | 58% | 0 |
| Grok 4.7 (xAI) | 152 | The Hermit | 25% | 29 | 70% | 44% | 0 |
The data
The raw answers (prompt, answer, card, orientation and time for every draw) and the summary are free to use under CC BY 4.0, with a link back to this page.
Who ran this
Familiar Tarot makes a tarot connector (an MCP server) that gives Claude and other assistants a real, shuffled deck, so the card is drawn before it is read. That gives us an interest in the result, which is why the prompts, the answers and the method are all published here.