When Asking a Chatbot to Draw, Cards Are Mostly Spoken into Existence
Search for "ChatGPT tarot" and you will find many people typing "draw a card for me" into a chat window, wondering if that draw actually counts. This guide looks at what is directly observable: how the cards are determined, and who writes the sentences. There is no percentage score for accuracy, and we will not manufacture one.
When you ask ChatGPT or another conversational AI to draw a card, there is usually no 78-card deck being shuffled behind the scenes. The model reads your prompt, generates a card name in its next response, and explains it. That is more like a card being spoken into the conversation than an observable shuffle.
Some tools connect a separate drawing program that fixes the cards first before asking the model to write. The visible difference is whether you can verify that the cards were established before the prose was drafted. Looking only at a fluent paragraph, you cannot tell the two steps apart.
It helps to break down what "AI tarot" means. One part is how cards are chosen; the other part is how the explanation is composed. When both are lumped under "AI," it is hard to tell which part can be checked.
Cards May Shift Across Turns, and Names May Not Match a Real Deck
If you ask the same question a few minutes later, the chatbot may name a different card. It has no obligation to treat its previous response as a fixed draw. If you treat it like continuing with the same physical deck, the explanations will conflict.
The card name might also not belong to the standard 78-card Rider–Waite–Smith system (22 Major Arcana and 56 Minor Arcana). Models sometimes mix up suits, ranks, or court cards into names that do not exist in the deck. This can happen; our purpose is not to single out products, but to show how you can verify.
The verification is straightforward: look up the card name in our card library. If it exists, read the explanation. If it does not exist, set that paragraph aside.
Orientation can also shift within the same conversation. A model might write upright in one sentence and reversed in the next, or change direction if you ask for something more encouraging. If an orientation changes with conversation, it was not fixed at the moment of the draw.
How Chatbots Can Actually Be Helpful
Chatbots can still be useful, just in a different role. Treating them as text organizers is usually far more helpful than treating them as card shufflers. If the card is already in your hands, the AI does not need to invent one.
You can draw physical cards yourself, paste the card names, orientations, and your question into the chat, and ask the model to structure reflections around those specific cards. You can then check its output against the cards in front of you. If its points do not connect to your cards, keep the cards and set the paragraphs aside.
Frame your question around what you can actually do. As shown in our guide on how to clarify a question, turn "Will they come back?" into your own purpose and boundaries, including how you will respect their space if they decline. The vaguer the question, the more likely the model will fill the void with generic romance clichés.
You can ask the AI to produce two columns: facts you explicitly provided, and its own interpretations. When an assumption is presented as established fact, highlight it. This check will not make the model infallible, but it prevents you from swallowing the whole text uncritically. If no new events have occurred, finish your current reading before deciding on repeating the same question.
How stilltarot Draws Cards via Code
At stilltarot, three cards are drawn programmatically from a 78-card deck. The moment a card is drawn, its orientation is fixed; it is not flipped over because an interpretation sounds uncomfortable. A single reading will never draw the same card twice. The technical steps are detailed in how this reading is created.
Descriptions and sources for every card are available on public pages without needing to complete a reading first. You can read the card meanings directly, then see whether the AI text stays grounded in them. Discard sentences that stray; do not adjust your memories just to make the article feel coherent.
Here, the AI does not choose cards. It receives fixed cards, your question, and your optional context, then composes a reflection. It might overlook a limitation you stated, or present a conjecture very smoothly. Those boundaries are discussed in what AI does in a reading.
If the AI reading service is unavailable, the site displays its status. You can still consult public card meanings; those texts are not answers generated for your question. There are no human tarot readers endorsing the text, and no user testimonials offered as proof of credibility.
If you only want to explore a slight inclination, you can consider a simple yes-or-no question. That offers a tendency and rationale rather than a final verdict. When your question is clear and you want to look at your current state alongside where things are heading, draw three cards from the home page.
Judging Value by Whether It Crosses Boundaries
If you want a percentage score for whether ChatGPT tarot is accurate, this guide does not offer one. There is no verifiable metric, and we will not invent one. Feeling understood by a sentence is very different from an event being predicted accurately. For more on this distinction, see how to look at tarot accuracy.
Whether a reading is useful comes down to three small checks: Did it address the specific matter you asked about? Did it offer a next step you can realistically take this week? Did it treat guesses as guesses, rather than claiming to reveal someone else's inner thoughts?
Language models are skilled at writing complete, polished sentences. Completeness can simply mean fluent phrasing. If your stated limits were ignored, or if meetings that never happened were inserted, treat those sentences as errors. You do not need to change your life to match the reading.
Card meanings are naturally broad. A phrase like "speak your truth" might make someone in love think of confessing feelings, while someone at work thinks of talking to their manager. Both feel the card was spot-on because the sentence left an open space. That space can help you start a conversation, but it is not a prediction.
What Neither Approach Can Do
A chatbot cannot tell you whether someone will message you tomorrow. A tool that fixes cards beforehand cannot tell you either. Whether an event will happen and on what date it will fall are not things models or cards can measure. A confident tone merely reflects confident phrasing.
The same applies to other people's feelings. Whether someone likes you, whether they told the truth, or what their unspoken thoughts are must be voiced by them directly, or confirmed by what actually occurs between you. Putting your question back onto what you can do frees the reading from having to pretend it knows what someone else is thinking.
Physical discomfort, signing contracts, and financial decisions should never be outsourced to any tarot reading. Those require evidence and qualified professionals. Cards can help clarify what worries you—what you fear or what fact is missing. They cannot provide medical diagnoses or legal verdicts.
If someone has clearly asked for no further contact, do not search for loopholes in another chat window. Clear boundaries do not require new paragraphs to reopen closed doors. If you are grieving, bring the question back to yourself: what support do you need today, and how can you take care of yourself right now?