AI Oracles & Divination: How Machine Learning Meets Ancient Wisdom
An "AI oracle" in 2026 is a specific and limited thing: a large language model given read-only access to traditional commentary, asked to translate a hexagram's text into language that speaks to your current situation. It is not clairvoyant, it does not perceive the future, and it is not a spiritual entity of any kind. It is a text-processing system that has read an enormous amount of I Ching scholarship and can produce fluent, contextualised commentary on demand. Understanding exactly what that means — and what it doesn't — is the starting point for using one well.
What LLMs are actually good at for I Ching reading
Large language models have been trained on essentially everything published in English (and much of what has been published in Chinese, German, and other languages). That includes Richard Wilhelm's 1923 German translation and Cary Baynes's English version of it; James Legge's 1882 Victorian scholarly edition; the Confucian Ten Wings; Thomas Cleary's more recent scholarly translation; and a large body of modern academic commentary. No reader working from a single printed volume has access to the comparative breadth that an LLM can draw on in a single response.
LLMs are also good at plain English. The original hexagram texts are terse, poetic, and often archaic in translation — "Preponderance of the great. The ridgepole sags to the breaking point." That sentence is accurate and is worth reading, but it is not immediately navigable for someone asking about a job decision. A language model can translate that image into something immediately usable without discarding the structure.
The third genuine advantage is dialogue. Traditional I Ching consultation is a one-shot transaction: you cast, you read, you sit with it. An LLM commentator can stay in conversation — you can ask "what does this mean for a situation where I have already committed to a path?" and receive a response shaped to that constraint. The oracle doesn't change, but the commentary can narrow down to what is actually relevant.
Finally, LLMs are useful as synthesisers across commentators. If you don't know whether Wilhelm or Legge has the better reading for Hexagram 29, and you don't have time to read both, an LLM can offer a synthesis — noting where the two traditions agree and where they diverge — without you having to locate the books.
What LLMs are NOT good at
The most important limitation is one that cannot be engineered away: an LLM does not know your situation. You bring the meaning. The hexagram is an image, and the image only makes contact with your life because you are the one holding it up to your circumstances. The LLM can offer commentary on what that image has historically meant; only you can decide where it lands.
LLMs hallucinate. This is not a fringe case or an early-model problem; it is a structural feature of how these systems work. They generate text that is statistically plausible given their training, which sometimes means they produce confident-sounding nonsense. In the context of I Ching commentary, this can look like fabricated quotations attributed to real translators, invented hexagram interpretations that don't appear in any actual edition, or confident cross-references to line texts that say something different from what the LLM claims. Any AI-generated I Ching commentary should be checked against the primary text, especially if you intend to quote or rely on a specific passage.
LLMs are calibrated to sound authoritative. They are not calibrated to accurately represent their own uncertainty. A response that begins "This hexagram suggests..." carries the same confident register whether the model is drawing on a strong scholarly consensus or making something up. The surface confidence is a feature of the text style, not a signal about reliability.
LLMs can drift toward sycophancy. If you supply context in your question — "I am thinking about leaving my job, is that what this hexagram is saying?" — the model may find a way to read the hexagram as confirming what you seem to want. This is not unique to AI; readers project onto oracle texts all the time. But LLMs are particularly susceptible to it because they are trained to produce responses that feel satisfying to the person reading them. The cure is the same as with any oracle: ask the question you are genuinely unsure about, not the one you want answered a particular way.
How IChing Oracle uses an LLM: the hybrid model
IChing Oracle uses DeepSeek's language model, accessed via API. The architecture is deliberate about what the LLM is and isn't being asked to do.
When you cast a hexagram, the authoritative hexagram text — judgment, image, and any relevant line texts — is loaded as ground truth and passed to the model as part of the system context. The LLM is not being asked to recall the hexagram from training memory; it is being asked to interpret a specific text it has just been given. This matters because it sharply reduces the risk of hallucinated hexagram content: the model is commenting on a document, not reconstructing one from memory.
The system prompt frames the LLM explicitly as a commentator, not the oracle. The oracle is the hexagram; the LLM is one voice offering a reading of it. The model is also cautioned, in the prompt, about its own limits — a reminder that matters because LLMs are generally inclined to project competence rather than uncertainty.
After the initial interpretation, the chat interface lets you ask follow-up questions. This is where the conversational strength of an LLM is most useful: you can narrow down from the general commentary to your specific constraint, or ask what a particular line text has meant across different translations. Each response continues to operate within the same framing — the model as commentator, the hexagram text as source of authority.
The ethics of "AI fortune telling"
The phrase "AI fortune telling" is inaccurate in a way worth being direct about. This is not fortune telling. Fortune telling is the claim to perceive future events. Neither the I Ching nor an LLM makes that claim honestly. What this tool offers is structured reflection: a way of slowing down and thinking about a situation through the lens of a 3,000-year-old framework, assisted by a language model that can translate the framework into contemporary terms.
The most important ethical boundary is substitution. An AI-assisted I Ching reading is not a substitute for professional medical, legal, financial, or mental health advice. This is not a disclaimer to be scrolled past: if you are in a crisis, if you are making a decision with serious legal or financial consequences, or if you are experiencing serious mental health symptoms, the right tool is a qualified professional, not an oracle of any kind. The I Ching and its AI commentators are appropriate for the space of reflection — for questions about orientation, attention, and choice — not for the space of diagnosis, prognosis, or binding decision.
There are specific situations where AI oracle tools are actively unhelpful. The most common one is seeking validation: you have already decided, you want confirmation, and you frame the cast to get it. LLMs are particularly susceptible to providing that confirmation because they are trained to satisfy. If you know you have already decided, the honest move is to put the question down. Another situation is severe emotional distress — when you are too close to a situation to read anything neutrally. At that point the text will mean whatever you need it to mean, and that is not a use of the tool.
Used honestly, the main ethical contribution of an AI oracle is the same as the I Ching itself: it slows you down. You have to frame a question, cast a hexagram, read a text, and sit with it. That pause, however brief, is the tool working as intended.
AI oracle vs traditional oracle: what changes, what doesn't
The hexagram cast is independent of the interpretation layer. Whether you use three physical coins, a dedicated yarrow-stalk ritual, or a cryptographically seeded random number generator, the casting step is a randomisation process, and its function — selecting one of 64 hexagrams plus a changing-line configuration — is the same. The casting mechanics do not change because an LLM is going to interpret the result. Those who find meaning in the ritual of physical coins can retain it entirely; those who find it irrelevant can skip to a digital cast. Neither position affects the hexagram you receive.
What changes with AI interpretation is the character of the commentary, not its source. The hexagram text is the same text it has always been. What the LLM adds is a fluent, personalised, plain-English reading that can respond to your specific framing. What it takes away — compared to reading the text yourself — is the friction of sitting with an archaic image until it opens up. That friction is not incidental. Many experienced practitioners would say it is where most of the work happens. An AI that makes the commentary immediately available short-circuits that process. The interpretation becomes faster and more conversational, and less grounded in the reader's own intuition and tolerance for uncertainty.
The honest summary: AI interpretation is not better than traditional reading. It is different. It is faster, more accessible, and more immediately coherent. It is also shallower in the specific sense that it gives you an answer before you have had to work very hard for one. Whether that trade-off suits your practice depends on what you want from the tool.
What this means for divination practice
The most useful frame for an AI oracle is as one voice among many, not as the authoritative reading. Use it to surface a translation you wouldn't have found, to get a first-pass plain-English version of an unfamiliar hexagram, or to put your situation into words when you are struggling to articulate the question. Then put the screen away and sit with the text.
A practical hybrid approach that many practitioners find useful: cast on physical coins or yarrow stalks if you prefer the ritual; look up the hexagram text in a printed translation; ask the AI for a plain-English first pass and for its sense of which line texts are most relevant to your question; then read the original text yourself. The AI commentary becomes a reading aid, not the reading itself. The primary text remains the source of authority; the LLM helps you get traction on it faster.
This approach also provides a natural check on hallucination. When you read the original text after reading the AI commentary, you will notice immediately if the LLM has mischaracterised it. That is a feature, not a limitation: the practice of checking the AI against the text is a form of close reading in itself.
The I Ching has survived three thousand years because it is genuinely useful as a thinking tool. AI interpretation extends its accessibility without changing its core character. The hexagram is still a model of a situation; the text is still terse and demanding; the work of connecting it to your life is still yours to do. The AI handles the translation layer. Everything else is the same as it has always been.