AI limitations in tarot readings

TarotAxis uses Anthropic's Claude family of large language models to turn the cards you draw into a personalised reflection. AI is a powerful interpretive engine, but it is not a mystical instrument and not omniscient. This page exists so that you know, in detail, what the model is actually doing — and what it cannot do no matter how the question is phrased.

What the model can do well

Claude is trained on a very large body of text that includes thousands of works on tarot, symbolism, mythology and depth psychology. Given a specific spread, it can recognise the cards as named entities, recall their canonical upright and reversed meanings, weave together the keyword sets of multiple cards into a coherent narrative, and ground that narrative in the question the seeker provided. It can also adjust tone — gentler, sharper, more practical — and translate the same reading idiomatically into English or Spanish.

In other words: it is an unusually good explainer of established symbolism. The card meanings it draws on were not invented by the model; they were inherited from the same Rider-Waite-Smith and Jungian sources a human reader would consult.

What the model literally cannot do

It cannot perceive you. There is no facial expression, no tone of voice, no body language, no knowledge of your history, your relationships, your medical chart, your finances, or anything outside the few sentences you typed. Every claim it makes about "your situation" is inferred entirely from the cards and your stated question. When it sounds uncannily specific, that specificity comes from the symbolism, not from any direct knowledge of your life.

It cannot predict the future. The model has no causal model of your world. When it uses future-tense language it is gesturing at possibilities suggested by the cards, not making forecasts. TarotAxis prompts are explicitly written to avoid hard predictions ("you will", "in three months"), and to speak instead about invitations, patterns, energies and choices.

It cannot give you facts it has not been given. The model has no access to your calendar, the news after its training cutoff, the weather, your messages, or anything happening on the live internet. If a reading sounds like it knows something it shouldn't, that impression is pattern matching — not information leakage.

Where the model tends to fail

Three failure modes show up consistently in AI-assisted tarot readings:

Over-confidence. Language models default to a fluent, confident voice even when the underlying inference is weak. A spread answered with "the cards strongly suggest" sounds more certain than it should. Our prompts ask the model to speak in invitations and possibilities, and to name uncertainty when it exists — but the default tilt is towards certainty and we cannot fully eliminate it.

Sycophancy. If a question is framed in a leading way ("the cards say I should leave my job, right?"), an unguided model often agrees. TarotAxis prompts redirect leading questions back to the cards themselves rather than confirming the framing — but a determined user can still push the model towards the answer they wanted to hear. The cure is to ask open questions, not yes/no prompts.

Confabulation. When the model is unsure of a symbol's historical detail, it sometimes invents one that sounds plausible. We mitigate this by feeding the model a compact, edited meanings table built from the human-reviewed card library, so the model has a stable reference and does not have to recall from memory. But fine-grained claims about historical sources or specific deck editions should be treated as starting points for research, not established fact.

What we do about it

Several engineering choices are in place to keep readings honest. The system prompt explicitly forbids medical, legal, financial and relationship-leaving advice; if a question lands on one of those topics, the model redirects to "what the cards invite you to reflect on" rather than what to do. The card meanings the model sees are the same human-reviewed entries published on the cards library, which keeps the AI in step with the rest of the site rather than drifting. Responses are length-capped so the model cannot bury weak inferences under volume. The temperature and decoding parameters are tuned to encourage hedged, reflective phrasing.

We also publish our reasoning so readers can pressure-test it. The methodology page describes the interpretive frame in detail, and the reflection-vs-prediction page sets out the philosophical stance behind those choices.

The bottom line

An AI tarot reading at TarotAxis is best understood as a structured prompt for self-reflection rendered in the voice of a thoughtful reader. It is not a forecast, a diagnosis, or a substitute for professional advice in any domain. If a reading helps you think more clearly about a situation, name a pattern, or ask a better question — it has done its job. If it tries to tell you what to do, treat it with the same scepticism you would extend to any stranger with strong opinions about your life.