Can ChatGPT or DeepSeek read your BaZi chart?

Reading a chart sounds like one skill. It is really three, and they fail differently.

Working out the chart turns your birth moment into eight characters. That is calendar arithmetic. Judging it weighs those characters against each other to decide what the chart is built around and what it needs — a set procedure with a right answer. Only the third job, explaining, is language work.

Language models are genuinely good at the third. The first two come first, and nothing in the output tells you whether they went right.

Your chart is not derived from your birthday. It comes from where your birth moment falls in the solar year, and that calendar does not line up with the one on your wall.

The year turns at 立春 (lì chūn, the start of spring), around 4 February — not at New Year, and not at Chinese New Year. Months turn with the seasons rather than on the 1st. Hours run in two-hour blocks. Add time zones and historical daylight-saving, and each is a line an answer can land on the wrong side of.

Here is what one of those lines costs:

            year   month   day    hour
1990-02-03   己巳    丁丑    己亥    己巳
1990-02-05   庚午    戊寅    辛丑    癸巳

Eight characters is where a reading starts, not where it ends. Those eight carry the five elements — wood, fire, earth, metal, water — across four pillars, and the reading lives in how they act on each other. Elements feed one another and check one another; some pairs pull together, others pull apart; where each sits changes how much it counts.

From that, three things get worked out in order: how well supported you are, what the chart is organised around, and — from those two — which elements genuinely help you. That last judgement is what everything practical rests on, and it is not a matter of opinion. It follows from the rest.

This is where a fluent answer is most likely to be a confident guess, and there is no way to tell from outside: the wrong judgement arrives in the same voice as the right one.

It is a fair question, and the intuition behind it is reasonable — a model trained on far more Chinese will handle the vocabulary, the classical register and the phrasing more naturally than one that was not.

But that is the third job. Neither of the first two is a language problem. Working out your pillars is arithmetic against a solar calendar, and judging them follows a procedure; fluency in Chinese makes neither more reliable.

If anything the fluency cuts against you here. An answer that sounds more at home in the subject is harder to doubt — and doubt is the only tool you have, unless the working is shown.

A wrong pillar can at least be checked against a calculator. An invented quotation cannot be checked by most readers at all.

Ask about a classical text and a model will often produce a line, a chapter and an attribution — to 《子平真诠》, say, or 《滴天髓》. Some of it is real. Some has the shape of a real quotation with nothing behind it, and the two look identical unless you go and check.

This matters more than a wrong fact. A citation is what a reader uses to decide whether to trust the rest, so an invented one borrows credit that was never earned.

The first two are computed. Your chart; how strongly your day master stands; the elements that genuinely help it — the 喜用神; and the interactions of the four pillars, both with one another and with the time you are standing in — the day, the month, the year, the decade — all come out of fixed classical rules, scored across the whole chart rather than counted, with each step recording what it set aside and why. The same birth moment always gives the same reading, and every claim opens to show the reasoning beneath it.

The third job is where a model belongs, and it is where one will be used. Ask 明命｜MING, currently being built, answers from the reading already computed for you — your chart, your season, the reasoning behind each — and is not permitted to introduce a claim that is not already in it. A model for language; never for the jobs that have a right answer.

There is also a difference in what each will answer. Ask a general model when you will marry, whether this year is dangerous, or which date to sign, and it will usually oblige. 明命｜MING does not predict events — what it computes about time is the interaction of your four pillars with it: what this decade emphasises in your chart, what this year asks of it, what today leans on. Perspective on the season you are in, never a named event or a date — and no date selection, no matching two charts, no reading a chart that is not yours, no diagnosing illness. Its position is 识己，而非预测 — recognition, not prediction — lived as its philosophy: 明其命 know your nature, 尽其力 give your best, 顺其时 move with your season.

You do not need to know the method to catch the common errors. Three checks are worth running against any chart you were handed, whoever handed it to you.

If you were born in late January or early February, check the year especially hard. That is the 立春 boundary, where the most consequential errors land.

If you were born near midnight, check the hour — and know that this is the one place where even correct sources legitimately differ. The double hour from 23:00 to 01:00 straddles two days, and the schools themselves split on which day a 23:10 birth belongs to. 明命｜MING follows the common convention (the hour advances; the day stays) and flags the alternative rather than hiding the question. A source that silently picks a side here is not wrong — but one that cannot tell you which side it picked is guessing.

Then ask the same question twice. A computed reading gives the same answer every time. A generated one may not — and if the characters or the verdict move between two attempts, no procedure was being run.

Reasonable, and here is how to do it without the failure mode.

Get your chart from something that computes it, and check it. Then paste those eight characters in and ask for explanation rather than derivation — that plays to what the model is good at. Ask it what a term means, not what your chart needs. And treat any classical quotation it offers as unverified until you have found it yourself.