AI Fortune Telling Flaws: Why AI Gets BaZi Charts Wrong

The Fatal Flaw of AI Fortune Telling: Why AI Gets BaZi Charts Wrong

Late one night in 2025, a product manager working in Shanghai opened DeepSeek, typed in her birth date, and within seconds received an elaborate BaZi analysis. In fluent, confident Chinese, the AI described the strength of her Day Master, explained how to identify her favorable elements, and even projected her career trajectory over the next two years. She screenshotted it and sent it to a friend. "This is scary accurate," her friend replied.

The problem was, her BaZi chart had been calculated incorrectly.

This is not an isolated case. Ever since DeepSeek ignited the "AI fortune telling" craze in early 2025, tens of millions of Chinese users have been using large language models — ChatGPT, Claude, Ernie Bot, Kimi — to generate BaZi (八字) charts and read destiny charts. Social media has been flooded with screenshots and debates, and "AI fortune telling" has become a genuine cultural phenomenon. There is real logic behind the trend: traditional numerology consultants charge considerable fees, vary wildly in quality, and often hide behind impenetrable mystical jargon. AI is always available, infinitely patient, and mercifully direct.

But behind this wave of enthusiasm lies a technical problem that almost nobody is taking seriously: large language models have several systematic, fundamental flaws when it comes to calculating BaZi charts. These are not random computation errors. They are deep contradictions between how LLMs work and what BaZi calculation actually demands.

This article intends to do something rigorous: dissect these flaws one by one, explain their technical origins, and make clear why they are more serious than most people realize.

LLMs Are Not Calculators: Understanding the Source of the Errors

Before getting into the specific errors, there is a foundational fact to understand: the core mechanism of a large language model is predicting the next token. Its "knowledge" comes from statistical learning across enormous volumes of text. What an LLM "knows" about BaZi exists because its training data contained large quantities of numerology books, forum discussions, and analytical articles.

This creates a fundamental contradiction. BaZi chart generation is, at its core, a precise astronomical and calendrical calculation problem. It requires the accurate conversion of a Gregorian calendar date and time into the lunar calendar, solar term positions, and the Heavenly Stems and Earthly Branches (天干地支, tiāngān dìzhī) cyclical system. This process demands table lookups, algorithmic steps, and exact handling of specific threshold values. LLMs excel at pattern recognition, language generation, and knowledge retrieval — but for tasks requiring precise numerical computation, asking an LLM to calculate a BaZi chart is like asking a widely-read literature professor to solve a calculus problem. He knows what calculus is. That does not mean he will get the answer right.

Complicating matters further, LLMs sometimes exhibit a deeply unsettling kind of confidence. Even when uncertain, they tend to produce an answer that sounds plausible rather than admitting ignorance. In the domain of numerological analysis, this tendency is especially dangerous, because errors in BaZi calculation cascade like dominoes: get the Day Pillar (日柱, rì zhù) wrong, and the Day Master (日主, rì zhǔ) is wrong; get the Day Master wrong, and the favorable element (用神, yòng shén) analysis is wrong; get the favorable element wrong, and every annual and decadal cycle judgment collapses with it.

Let us examine the most common categories of error one by one.

The First Category of Error: The Treacherous Conversion Between Lunar and Solar Calendars

The first step in generating a BaZi chart is determining the Year Pillar (年柱, nián zhù), Month Pillar (月柱, yuè zhù), Day Pillar, and Hour Pillar (时柱, shí zhù). Many people know that BaZi uses the lunar calendar, but the reality is more complex: within the BaZi system, the Month Pillar is determined by solar terms (节气, jié qì), not by the lunar calendar month. The boundary point for the Year Pillar is Lìchūn (立春, the "Start of Spring" solar term), not the first day of the first lunar month. This trips up many beginners, but professional software handles it correctly. The real traps lie elsewhere.

The most common errors occur in the conversion between the lunar and Gregorian calendars. The Chinese lunar calendar is a lunisolar system that incorporates complex intercalary rules. Certain years contain a leap month, and the same Gregorian date in different years can correspond to entirely different lunar dates. When LLMs handle these conversions, they tend to rely on "typical cases" encountered in training data, and their handling of edge cases is extremely fragile.

A concrete example: the year 2001 contained an intercalary fourth lunar month. That year had two fourth months — a standard fourth month and a leap fourth month. If you were born in June of that year according to the Gregorian calendar and asked an LLM for your lunar birth date, the answer it gives you could easily be off — because the model would need to know that specific year's intercalary rules, the precise boundaries between months, and then perform an exact day-count calculation. This is not what LLMs are built for.

A subtler class of error involves the precise moment of each solar term. In the BaZi system, the transition point for the Month Pillar is the solar term itself, not the first day of the lunar month or the first of the Gregorian calendar month. Lìchūn, for instance, falls somewhere between February 3rd and February 5th in the Gregorian calendar each year — the exact moment is determined down to the minute, sometimes to the second. If you were born within a few hours of Lìchūn, both the Month Pillar and the Year Pillar may be at their transition point. LLMs generally know that "Lìchūn is around February 4th," but "around" is lethal here. A difference of a few hours can shift the Year Pillar from Jiǎ Zǐ (甲子) to Yǐ Chǒu (乙丑), producing an entirely different chart.

On the question of precise solar term timing, the LLM's own training data is inherently limited. The exact moment of each solar term is an astronomical calculation derived from the sun's ecliptic longitude. An LLM cannot "memorize" the precise timing of every solar term in every year. What it produces is typically a vague approximation — and in BaZi, approximations break charts.

The Second Category of Error: The Systemic Neglect of True Solar Time

This is the most widespread and the most overlooked error in AI fortune telling.

Modern society operates on standardized time zones. In China, this is "Beijing Time" — the Eastern Eighth Zone standard (UTC+8), which corresponds to the local solar time at 120 degrees east longitude. But Beijing Time is an administrative convention. China's territory spans nearly 65 degrees of longitude, from the Pamirs of Xinjiang at approximately 73 degrees east to the Heilongjiang border at approximately 135 degrees east. This means that within China's borders, the difference between actual local solar time and Beijing Time can reach nearly two full hours.

In the classical BaZi tradition, the Hour Pillar is calculated using True Solar Time (真太阳时, zhēn tàiyáng shí), not standard clock time. The reason is simple: the BaZi system developed in an era without time zones. It was built around the actual movement of the local sun — sunrise, noon, sunset — which are inherently local events.

The overwhelming majority of AI fortune-telling tools — including virtually all large language models — take a user's entered birth time at face value and use it directly, completely bypassing True Solar Time correction. This error systematically affects every single user of these tools.

Consider a concrete case: a user born in Ürümqi, Xinjiang, with a registered birth time of 8:00 a.m. Beijing Time. Ürümqi sits at approximately 87.6 degrees east longitude, which is roughly 32.4 degrees west of the 120-degree standard meridian. Converting that difference into time gives approximately 2 hours and 9 minutes. In other words, when Beijing Time reads 8:00 a.m., the True Solar Time in Ürümqi is approximately 5:51 a.m. In the BaZi hour system, 8:00 a.m. falls within Chén Hour (辰时, 7:00–9:00 a.m.), while 5:51 a.m. falls within Mǎo Hour (卯时, 5:00–7:00 a.m.). The Hour Pillar shifts from Chén to Mǎo — different Nayin (纳音) resonance, different position in the Twelve Growth Stages (十二长生, shí'èr zhǎngshēng), and entirely different relationships with the other three pillars.

Even outside the frontier regions, this problem persists in the eastern provinces. Chengdu sits at approximately 104.1 degrees east, roughly 15.9 degrees west of the standard meridian — a time difference of about 64 minutes. Shanghai at approximately 121.5 degrees east is close to the baseline, with only about a 6-minute difference. But for a person born right at the boundary between two hour intervals, even a 6-minute discrepancy can determine which Hour Pillar they carry.

There is an additional layer of complexity: the historical changes to China's time administration. During the Republic of China period, the country used five separate time zones. After 1949, the People's Republic unified the country under Beijing Time. During certain periods of the Second Sino-Japanese War, some regions also observed daylight saving time. If you were born before 1949 or during certain specific historical windows, all of these factors need to be accounted for. An LLM's ability to handle these historical details reliably is, to put it charitably, deeply questionable.

---

Want to experience professional AI BaZi analysis? Generate your free chart now and see what precise calculation combined with AI interpretation actually feels like.

---

The Third Category of Error: "Hallucinatory Inference" in Earthly Branch Relationships

This category of error is different in nature from the first two. The first two are calculation errors that occur during the chart generation stage. The third occurs during the analysis stage and represents a specific cognitive distortion native to LLMs — what researchers commonly call "hallucination."

The relationships between Earthly Branches (地支, dìzhī) sit at the heart of BaZi analysis: the Six Combinations (六合, liù hé), Triple Combinations (三合, sān hé), Directional Combinations (三会, sān huì), Six Clashes (六冲, liù chōng), Six Harms (六害, liù hài), and Six Punishments (六刑, liù xíng) together form the foundation for reading a chart's dynamic energy. The problem is that these relationships are not a simple lookup table. Applying them correctly requires evaluating activation conditions, relative strength, transparency into the Heavenly Stems, void and emptiness (空亡, kōng wáng) status, and numerous other contextual factors.

When processing these questions, LLMs are inclined to produce an analysis that "sounds reasonable" rather than strictly deriving conclusions from the rules. A classic error pattern is the over-triggering of branch relationships: if Zǐ (子) and Chǒu (丑) both appear in a chart, the LLM will announce "Zǐ-Chǒu Six Combination, Earth absorbing Water," with no consideration of whether these two branches are adjacent, whether other branches intervene, or whether the conditions for transformation are actually met.

In professional numerological practice, the formation of a Triple Combination (三合局, sān hé jú) is conditional. Take the Hài-Mǎo-Wèi (亥卯未) Wood combination as an example: even if all three branches appear in a chart, one must further evaluate whether spatial conditions for combination exist, whether the relevant pillars are adjacent to one another, whether any of the branches have been clashed or damaged, and what Heavenly Stems are revealed above them. These answers determine the combination's strength and actual utility. LLMs lack the capacity for this kind of strictly conditional reasoning. They are far more comfortable outputting simple declarations like "Hài-Mǎo-Wèi present, Wood combination formed, Wood prosperous, Fire in season."

The calculation of void and emptiness — Xún Kōng (旬空), sometimes called "empty branches" — is another disaster area. Calculating Xún Kōng requires first identifying the Day Pillar's Heavenly Stem, finding the corresponding ten-day cycle (旬, xún), and then determining which two Earthly Branches are absent from that cycle. Those absent branches are the void branches. This is a sequential calculation with precise steps. In testing, we found that major LLMs produce incorrect Xún Kōng calculations at an alarmingly high rate — not because they don't know what Xún Kōng is, but because they stumble during step-by-step numerical execution. Worse, they continue building analysis on top of the erroneous foundation, generating a cascade of compounding errors.

The calculation of auxiliary stars (神煞, shén shà) follows the same pattern. Nobleman Stars (贵人, guì rén), Traveling Horse Stars (驿马, yì mǎ), Peach Blossom Stars (桃花, táo huā), Canopy Stars (华盖, huá gài), Sword-Edge Stars (羊刃, yáng rèn) — every one of these has its own distinct calculation rule. Some are derived from the Year Branch, others from the Day Stem, others require combining multiple stems and branches. LLMs frequently cross-contaminate the rules for different auxiliary stars, or memorize the "common case" for one star and then generalize it to situations where it does not apply.

The Fourth Category of Error: The Logical Traps of Luck Cycle Calculation

The Luck Cycles (大运, dà yùn) represent the temporal spine of BaZi analysis. The age at which a person's first Luck Cycle begins, and whether the cycles run forward or backward through the stems and branches, must be determined based on the yin-yang polarity of the birth year and the gender of the Day Master. Each Luck Cycle governs a decade, and the stems and branches of successive cycles follow the sequence of month pillars either forward or backward in the calendar.

This calculation depends critically on precise solar term timing: the activation age is derived from counting the days between the birth date and the nearest solar term, forward or backward depending on the chart's parameters, with three days conventionally equating to one year (though different schools apply minor variations to this rule). LLMs face the same problem here as before: they may know the approximate date of a solar term, but the exact day count involves precise calendar arithmetic that is easy to get wrong.

More subtle errors emerge in the layered analysis of Luck Cycles interacting with annual cycles. When the stems and branches of a current Luck Cycle create complex relationships with the natal chart — for example, the Luck Cycle's stems and branches clashing with natal pillars while an annual cycle simultaneously activates a Triple Combination — this multi-layered dynamic analysis requires application of strict priority rules. LLMs frequently produce self-contradictory analyses in this territory: one paragraph describes a particular period as favorable, while another paragraph assigns pressure or conflict to the same stems-and-branches configuration, and the model never registers that these two judgments are mutually exclusive.

This kind of internal logical inconsistency is a persistent weakness of LLMs generating long-form text. When predicting each successive token, there is no global logical integrity check ensuring that the argument remains consistent across the whole document. BaZi analysis — a task requiring logical coherence across multiple simultaneous dimensions — presents a genuine and serious challenge for this architecture.

The Fifth Category of Error: The Epistemological Chaos of Mixed Schools

This is a deeper problem, and one that almost nobody discusses seriously.

BaZi numerology is not a unified system. Several major schools have developed historically: the Zǐpíng method (子平法, zǐpíng fǎ) associated with Xú Zǐpíng, which represents the traditional mainstream; the Blind School (盲派, máng pài) associated with figures like Yuán Shùshān and Wáng Tíngzhī, which emphasizes concrete event prediction; modern numerology, which incorporates more statistical methods and psychological frameworks; and the seasonal adjustment system represented by the classical text Qiōng Tōng Bǎo Jiàn (穷通宝鉴). These schools hold substantively different positions on core questions — for instance, in how they identify the chart's ruling structure (格局, gé jú) and determine the favorable element, the Zǐpíng method and the Blind School diverge quite sharply.

An LLM's training data contains texts from all of these schools, but the model has no capacity to identify which statement originates from which school, let alone to make coherent, consistent choices between them. The result is that the analysis you receive is typically a random blend of multiple schools' perspectives: one paragraph uses Zǐpíng structural logic, the next applies Blind School auxiliary star reasoning, the following one quotes the seasonal adjustment principles of Qiōng Tōng Bǎo Jiàn — and in certain situations, these three frameworks directly contradict each other.

A reader with genuine knowledge of numerology will notice this confusion immediately. But most people seeking a reading have no ability to detect these inconsistencies. What AI presents is a sense of authority — fluent prose, professional terminology, confident tone — and that surface authority conceals the internal logical chaos completely.

The Cascade Effect: Why "Small Errors" Are Anything But

Every category of error discussed above might seem, in isolation, to be tolerable. But BaZi analysis is a highly interdependent system. Any foundational error propagates forward through the analytical chain and amplifies exponentially.

Trace the path of a single error: suppose that due to ignored True Solar Time, someone's Hour Pillar is incorrectly calculated as Zǐ Hour (子时) when it should be Hài Hour (亥时). This single mistake triggers the following consequences: the Heavenly Stem and Earthly Branch of the Hour Pillar both change; all Six Combinations, Clashes, Harms, and Punishments between the Hour Branch and every other branch must be entirely recalculated; the positioning of various auxiliary stars derived from the Day Branch shifts; the combined relationship between the Day Master's sitting branch and the Hour Branch changes, altering the assessment of the Day Master's overall strength or weakness; the identification of the favorable element changes accordingly; the directional influence of Luck Cycles and annual cycles on the chart changes; and finally, every concrete judgment about career, relationships, and wealth is built on a corrupted foundation.

One incorrect Hour Pillar can bring the entire analysis down from its roots. And this error is silent. It triggers no warning. The AI will continue generating smooth, detailed, superficially professional analysis on top of the mistake. The user has absolutely no way to assess from the quality of the writing whether the underlying calculation is correct.

This is precisely why the independence and verifiability of the calculation engine matters so much.

The DeepOracle Approach: Separating Calculation from Interpretation

In the AI fortune-telling space, DeepOracle has taken a different path: explicitly separating verifiable astronomical and calendrical calculation from AI-driven multi-school interpretation.

Chart generation should not be performed by an LLM. DeepOracle uses an independent calculation engine to handle all calendar conversions and stems-and-branches derivations. This engine is built on precise astronomical algorithms and correctly performs True Solar Time correction calculated from the birth location's exact longitude, accurate handling of all historical time zone changes and daylight saving time periods, solar term timing precise to the minute rather than "approximately this date," and strict rule-based calculation of void branches, auxiliary stars, and Luck Cycle start and end points.

The calculated results are independently verifiable — you can cross-check them against other professional software outputs. This represents a fundamental difference from an LLM generating a chart directly: the LLM's "calculation" is a black box. There is no way to know what intermediate steps it took, and no way to trace the origin of an error.

With an accurate chart confirmed as the foundation, AI capability finally has a legitimate role to play: synthesizing analytical perspectives from multiple schools, retrieving relevant classical texts, conducting logically consistent structural analysis, and explaining complex terminology in clear language. At this stage, the LLM's language comprehension and generation abilities are genuine strengths — deployed in a context where they actually belong, rather than being misapplied to chart generation where they do not.

---

Interested in DeepOracle's professional plans? View the full features and pricing and find the option that fits your depth of inquiry.

---

Advice for Users: How to Recognize Unreliable AI Fortune Telling

Before closing, it is worth saying something practical. If you are using or considering an AI fortune-telling service, several questions can help you quickly evaluate a tool's reliability.

First, ask the tool directly: does your birth location have any effect on the chart calculation? If it says no — or if it never asked for your birth location in the first place — you can be almost certain it is ignoring True Solar Time correction.

Second, look for whether the tool clearly explains how it generates the chart. If the chart is produced directly by an LLM with no independent calculation engine behind it, the accuracy of the lunar calendar conversion and solar term handling is essentially unguaranteed.

Third, watch for internal contradictions in the analysis text. If the same stems-and-branches relationship is assigned contradictory meanings within a single reading, or if the identification of the favorable element is inconsistent from one section to the next, this is typically a symptom of mixed-school confusion.

Fourth, maintain healthy skepticism toward sweeping pronouncements about fate. Serious numerological analysis should be probabilistic and conditional, not deterministic prophecy. Phrases like "in 2026, you will definitely..." — whether from a human or an AI — deserve suspicion regardless of the source.

Finally, and most importantly: BaZi numerology is a complex knowledge system with several centuries of history. Its value lies in helping people understand their own energetic characteristics and the rhythms of time — not in delivering certain answers about destiny. AI can be an excellent translator between this ancient system and the modern world. But only if that translator's foundational work is honest and reliable.

FAQ

Q: Why do large language models make errors in BaZi chart generation specifically, rather than in general knowledge questions?

BaZi chart generation is fundamentally a multi-step precise calculation task, not a knowledge retrieval task. It requires converting a specific Gregorian date and time through a series of algorithmic steps — lunar calendar conversion, solar term positioning, local time correction — to arrive at four stems-and-branches pillars. Every step in this process demands exact numerical computation; approximation is not acceptable. Large language models operate through statistical token prediction, and there is a fundamental gap between "knowing how to describe the chart generation process" and "being able to correctly execute the chart generation calculation." It is similar to the difference between being able to eloquently describe how to perform a piece of piano music and actually being able to play it.

Q: How important is True Solar Time correction really? Do ordinary users need to care?

True Solar Time correction has the greatest impact on users born in the eastern and western extremes of China. Users in Xinjiang, Tibet, Gansu, and western Inner Mongolia may face a discrepancy between their registered birth time and True Solar Time that exceeds one hour, making it quite likely their Hour Pillar is incorrect. Even in central and southwestern provinces like Sichuan and Chongqing, the difference is around one hour. For anyone whose birth time falls within approximately 30 minutes of the start or end of an hour interval in the BaZi system, True Solar Time correction is a decisive factor in determining the Hour Pillar. In practical terms: if you were born outside the eastern coastal provinces, or if your birth time is close to an exact or half hour on the clock, True Solar Time correction is a necessity for you, not an optional refinement.

Q: Compared to consulting a human numerologist, is AI fortune telling more or less reliable?

The answer depends on which dimension you are comparing. An experienced professional numerologist brings irreplaceable advantages in intuitive interpretation, concrete event prediction, and the ability to follow a dynamic conversation. However, the quality of human practitioners varies enormously. The market contains a very large number of people operating under the title of "numerology master" whose actual competence is deeply concerning. In terms of chart generation accuracy, professional software using an independent astronomical calculation engine should be more reliable than an LLM generating a chart directly. An ideal approach is to use a verified tool to generate an accurate chart, combine this with AI-driven multi-school analysis, and treat the result as a supplementary framework for understanding your own characteristics — rather than as an authoritative source of answers about your fate.

Q: Why do LLMs so frequently get void branch (Xún Kōng) calculations wrong?

Calculating Xún Kōng requires three steps: identify the ten-day cycle to which the Day Pillar's Heavenly Stem belongs (each cycle begins with Jiǎ and covers ten stems), determine which two of the twelve Earthly Branches are absent from that cycle's ten-day span, and identify those absent branches as the void branches for the current cycle. This sounds straightforward, but LLMs tend to stumble on the first step — correctly identifying which ten-day cycle a given stem belongs to. They may misremember the cycle assignment for a particular stem, or confuse the void branches of different cycles. A more common error is conflating the calculation rules for Xún Kōng (cycle void) with those for Jié Kōng (月空, monthly void). Since Xún Kōng calculation starts from the Day Pillar, and the Day Pillar itself may already contain an error, a mistaken void branch calculation compounds on top of an already flawed foundation, with significant consequences for the overall reading.

Q: How significant are the differences between BaZi schools, and can AI handle multi-school analysis?

The differences between schools are quite substantive on certain questions. Take the identification of the favorable element as an example: the Zǐpíng method centers on the chart's structural pattern and its favorable or unfavorable configurations, deriving the favorable element from the success or failure of that structure; the Blind School centers on the Day Master's strength or weakness and takes the favorable element directly as whatever supports or restrains that strength; the seasonal adjustment system of Qiōng Tōng Bǎo Jiàn prioritizes balancing the chart's warmth, cold, dryness, and moisture. For the same chart, these three frameworks can produce different and even opposing identifications of the favorable element, with downstream effects on assessments of career direction, beneficial industries, and favorable Luck Cycle periods. An AI system specifically designed and prompted to handle multi-school analysis can, to a meaningful degree, present the perspectives of different schools with clarity and explain where they diverge — which is actually more valuable than many human practitioners who only work within a single framework. But this requires deliberate system design. It cannot be achieved by simply letting an LLM improvise.

Q: How is DeepOracle's calculation engine different from ordinary BaZi apps?

Most BaZi apps and websites on the market use a lookup-table approach — they pre-calculate the stems and branches for every date within a given range and simply retrieve results from that table. This works fine for common dates, but it is error-prone near solar term boundaries and in cases involving historical time zone changes. The vast majority also perform no True Solar Time correction whatsoever. DeepOracle's calculation engine uses astronomical algorithms to compute the precise moment of each solar term in real time, supports True Solar Time correction based on the birth location's exact longitude, and correctly handles special historical cases including the Republic of China's multi-timezone period and the various daylight saving time schemes used in different eras. More importantly, this calculation process is transparent and traceable: users can see exactly how many minutes of time difference were applied and what the precise moment of the relevant solar term boundary was — rather than receiving a result whose origin cannot be examined or verified.

Further Reading

How Does AI BaZi Analysis Actually Work? A Full Technical Breakdown

DeepSeek vs. ChatGPT for Fortune Telling: An Evaluation from a Professional BaZi Perspective

How Important Is Your Birth Hour in BaZi? A Deep Dive into the Hour Pillar

Free Online BaZi Chart: How to Read the Foundational Structure of Your Chart