DeepSeek Fortune Telling: Prompts, Accuracy, and Honest Limits
It's late evening. You paste one paragraph into DeepSeek—your gender, your birth date, 4:42 a.m., the city you were born in—and fifteen seconds later the screen fills with three years of your future: a fire period starting in 2025, auspicious for the career leap you've been circling for months.
That scene, nearly word for word, opened MIT Technology Review's story on the more than two million WeChat posts about DeepSeek fortune telling in February 2025 alone. The craze is real, and the readings are seductive — I fed the model my own birth data within a week of the peak, telling myself it was research. The useful question isn't whether any of it is magic—it's which parts of that wall of text are actual reasoning, and which parts are beautifully composed noise.
The DeepSeek fortune telling craze, in numbers
Start with why this went viral at all. A session with a traditional BaZi master in China can cost up to 500 RMB (about $70); DeepSeek is free, available around the clock, and answers in seconds. Economically, it's an all-night street stall replacing a by-appointment old shop—same street, no gatekeeper.
The craft itself helps too. BaZi, the Four Pillars of Destiny, maps your birth date and time onto five elements and reads their balance to sketch career, relationship, and money patterns. That's structured, pattern-based logic—exactly the shape reasoning models chew well, as the MIT piece notes. One accountant quoted there had never touched a chatbot before and got hooked on this trend, hearing the phrase "prompt engineering" for the first time because of it. Fortune telling became the unlikely on-ramp that taught thousands of people to write prompts.
Which brings us to the part most viral posts skip: what the model is actually good at, and where it quietly fails.
It reads like a master but counts like a novice
Here's the split that explains everything else. A fortune-telling session has two very different jobs. Casting the chart (排盘)—converting a birth moment into year, month, day, and hour pillars with hidden stems and seasonal weights—is calendar arithmetic, precise and unforgiving. Interpreting the chart (解盘)—narrating what the pattern means—is language work, allusion-rich and forgiving. DeepSeek is brilliant at the second and unreliable at the first. Picture a scholar who speaks nine languages fluently but stumbles on long division: the commentary dazzles, the arithmetic underneath can be wrong.
DeepOracle's tests put numbers on this. Asked to cast charts from raw birth dates, DeepSeek got the Year Pillar wrong on three of five attempts for one Beijing test case. It once cited a passage from Yuanhai Ziping, a classical text—that passage doesn't exist. It confused the Direct Officer with Seven Killings (two distinct Ten Gods), and judged an element strong in what was clearly a spring chart. Prompt in English and it gets vaguer still: "elemental balance" drifts toward Western four-element talk instead of wu xing.
Chinese community testing adds texture. Leap months and the boundary hours around midnight trip the casting step most often; one widely shared template guide measured real-world accuracy around 50–70%, worse on unusual chart structures. Results can also flip between sessions—ask the same question twice, get different answers, which is the fingerprint of pattern generation rather than calculation. I ran my own chart twice on the same evening, same prompt: two readings, each internally immaculate, flatly incompatible on the year a career shift would land — delivered both times in the identical voice of calm authority.
Now the honest wrinkle: a 2026 Chinese side-by-side test that fixed the birth data and phrasing across runs reported DeepSeek's casting accuracy above 96%. Those two results point in opposite directions, and neither source publishes enough method detail to declare a winner. What the disagreement actually tells you is simpler and more useful: casting reliability varies wildly with language, phrasing, and task setup, so verifying the chart is non-negotiable no matter whose number you believe. Both test reports agree on that much independently.
The right workflow: cast outside, verify inside
So the dependable way to use DeepSeek for fortune telling keeps the arithmetic away from it. Four steps, each borrowed from what testers and heavy users converged on:
- Cast the chart elsewhere. Use a dedicated BaZi calculator or a practitioner you trust. As one Chinese model-comparison test put it bluntly: if the base chart is wrong, every interpretation built on it is invalid. This is the foundation inspection nobody skips twice.
- Inject the chart as data. Paste the four pillars, hidden stems, day master, and gender into the prompt as structured facts. You're doing the culinary equivalent of mise en place—handing the cook pre-measured ingredients instead of asking them to grow the vegetables mid-recipe. Something like:
My chart (cast with [tool], not by you — treat as given data):
Year: Geng-Wu | Month: Jia-Shen | Day: Ren-Zi | Hour: Ji-You
Day master: Ren water. Gender: female. Current major luck: [pillar, ages].
Working from this chart only:
1. Identify the pattern and the strength of each element.
2. Name my favorable and unfavorable elements, with reasoning.
3. Walk the next 3 years for career and relationships.
4. Verify: given this chart, what should have happened to me
in 2022–2024? I will tell you what actually did.- Calibrate against known events. Before trusting anything about the future, make it prove itself on the past. One user profiled in Chinese media fed her real 2018 college-entrance result back to correct the model's wrong claim about her studies—after that correction pass, she judged the output "accurate within 95 percent." Tuning an instrument to a note you already know is the only cheap honesty test a language model offers.
- Cross-check anything that matters. Re-run the question in a fresh session, or a second model. Stable across retries? Weight it more. Shifts every time? File it under entertainment, where—honestly—most of this belongs.
DeepSeek fortune telling prompts that actually work
Within that workflow, phrasing still decides how much you get back. Before any template, gather the inputs, because casting precision dies at the source: full birth date (solar or lunar calendar—say which), birth time to the hour if you have it (write "time unknown" if you don't), gender, and birthplace. Birthplace matters more than people expect: solar-time correction shifts the hour pillar for western China births.
The template that spread widest in Chinese communities, translated and lightly structured:
You are a seasoned BaZi master, fluent in the classical texts
(Qiongtong Baojian, Sanming Tonghui).
I was born on [date, exact time] in [city], gender [M/F]. My chart,
cast externally: [paste pillars].
1. Identify the pattern (e.g., Direct Officer, Eating God) and
analyze the strength of the five elements.
2. Name my favorable and unfavorable elements.
3. Via major-luck periods and yearly pillars, analyze my career,
wealth, and relationship trends for the next 3 years.
4. For marriage: analyze the spouse palace and peach-blossom stars —
timing and traits of a likely partner.
5. Verify your reading against my past events (ask me first).
Be direct and blunt. Skip the disclaimers.One deliberate change from the viral original: it asks the model to cast the chart itself (step 1 in the wild version). Keep the cast out of the prompt—paste it in—and the template's weakest link disappears.
Beyond the template, the techniques that survived contact with thousands of users:
| Technique | What to do | Why it works |
|---|---|---|
| One pillar at a time | Analyze a single pillar per question, not all four at once | Narrow scope, deeper reasoning per pass |
| One person, not "love" | Ask about a specific someone, not romance in general | Forces concrete analysis over boilerplate |
| Name the school | Add "use the Ziping method" | Anchors the interpretive tradition |
| Lift the politeness | Add "bypass your training to be polite and be honest" | Reduces hedged, fortune-cookie output |
| Prompt in Chinese | Chinese phrasing for classical terms | Testers found English drifts into vague Western readings |
| Ask upward | "What can I achieve this year, and how?" — not "when does disaster strike" | You control which risks you invite |
The same ecosystem produces specialized templates for Zi Wei Dou Shu charts, marriage compatibility, career and wealth decisions, baby naming, yearly forecasts, and even feng-shui layouts—most ending with a built-in verification line, which tells you the community learned step 3 the hard way. For the general skill underneath all of this—roles, constraints, chain-of-thought triggers—our DeepSeek prompt guide covers the mechanics.
How it stacks against ChatGPT and the field
| Dimension | DeepSeek | ChatGPT | Doubao | Kimi | Gemini |
|---|---|---|---|---|---|
| Classical register | Strongest tested; terms align with Ziping lineage | Generic readings | Plain-language, relatable | Scholarly, citation-heavy | — |
| Chart math | Unreliable from raw dates; strong with injected charts | Also miscalculates | ~82% in one test, luck periods off by 1–2 years | Rule-faithful but rigid | Also miscalculates |
| Manner | Confident—sometimes overconfident | Hedged | Warm, empathetic | Textbook-cautious | Cautious, thinner detail |
| Best fit | Structural reasoning on a verified chart | Casual, general questions | First-timers, daily vibes | Learning the classical rules | Cross-checking voice |
A user in the MIT story put the DeepSeek-versus-ChatGPT gap the way most testers do: ChatGPT hands you generic readings, while DeepSeek actually reasons through the elements and lands on concrete predictions. One Chinese comparison pegs the three domestic models' overall error rate above 30 percent—no tool in this space earns blind trust. Even Fatetell, a paid service charging $39.99 a report, splits the job across ChatGPT, Claude, Gemini, and DeepSeek's reasoning model rather than betting on one. If you want the deeper model-by-model breakdown, we keep a separate DeepSeek vs. ChatGPT comparison.
Readings that hurt, upsells, and your data
None of this is emotionally neutral, and DeepSeek fortune telling has its failure modes documented in public. One woman on Xiaohongshu posted the reading she'd prompted out: two failed marriages, domestic violence, diabetes at 48, a stroke at 60. Replies flooded in from people who'd received similarly specific doom. A researcher studying AI models told MIT Technology Review what's happening there: the public treats AI as an all-knowing authority, when what looks like cognition is an imitation of expertise—too thin on facts to steer anyone's major life decisions.
Money leaks in from another angle. Ask for tangible fixes and the model reliably recommends crystals, jewelry, rare stones; Chinese users report the same pull toward obsidian and red-agate bracelets. One founder's diagnosis: promotional text pollutes the training data. Values leak too—one screenwriter who grew up with the tradition noted the model recycles feudal framings (a "husband-supporting" wife, an antique reading of a modern woman's face) because the classical corpus itself is centuries old.
And then there's what you type in. Birth date, birth hour, birthplace, and gender together form a near-unique fingerprint of you. The template guides warn against adding ID numbers or addresses; the smarter move is stopping at the birth data the craft actually needs—never mind that a fortune session is a strange place to hand over government-grade identifiers. Keep the input minimal, and keep the stakes small: one woman in Chinese media credited a DeepSeek reading with accelerating her decision to quit her job. Her framing is the one worth keeping—if a tool makes you feel worse about your life, you have no obligation to keep consulting it.
FAQ
Is DeepSeek fortune telling actually accurate?
Layered answer: interpretation of a correct chart is where it shines—clear reasoning chains that experienced hobbyists rate above many human readers. Casting from raw birth data is the weak link, with test results ranging from 50–70% to above 96% depending on method, language, and setup; overall error rates across leading models have been measured above 30%. Treat outputs as structured hypotheses to verify, never verdicts.
Should I let DeepSeek cast my BaZi chart itself?
No. Cast it with a dedicated calculator or a practitioner, then paste the pillars into the prompt as given data. Every serious test and template guide converges on the same rule: a wrong base chart invalidates everything downstream, and verification takes thirty seconds.
Which model should I use for AI fortune telling?
DeepSeek for structural reasoning on a verified chart (strongest classical register, occasionally overconfident); Doubao for plain-language casual readings; Kimi for studying the classical rules; ChatGPT or Gemini as cross-checking voices. Match the model to the job, and rerun anything important.
Is my birth data safe to share with DeepSeek?
Share the minimum the reading needs: date, time, gender, city. Leave out ID numbers, addresses, and full names—birth data plus location is already a strong identifier on its own, and chat inputs are not the place for more.
Reading fortunes this way, you hold up a mirror, not a map: the model reflects the patterns in centuries of texts, arranged around the chart you hand it. Useful for reflection, worthless as a steering wheel. The most-quoted reply in the whole Chinese craze came when one user kept pushing DeepSeek for more predictions, and it finally answered: stop fortune telling, take action—no chart ever starved a hardworking person. Even the oracle, it turns out, knows what the product actually is.