Can ChatGPT Calculate Your BaZi Chart? A Test You Can Run Yourself

Can ChatGPT Calculate Your BaZi Chart? A Test You Can Run Yourself

Ask a chatbot for your BaZi chart and you will get one. It will come back fast, formatted cleanly, with a confident paragraph about your day master underneath. What it will not tell you is whether the eight characters at the top are the ones you were actually born with.

Here is the plain answer first. A chatbot answering from memory is not a reliable way to cast a BaZi chart. The reason is not that it cannot reason about BaZi. The reason is that two of your four pillars turn on astronomical moments, and those moments fall at a different minute every year. A model recalling them from text is guessing near every boundary. A chatbot that calls a real calculation tool can get them right. The trouble is that the two sound exactly the same.

So this piece does not ask you to trust us. It gives you four real birth moments, the correct pillars for each, and the reason each one is a trap. Paste them into any chatbot you like and compare.

Why a chart is astronomy, not a calendar

A BaZi chart has four pillars, for the year, month, day and hour of birth, and each pillar is two characters. That is where the name comes from. Bazi means eight characters. If that is new, what a BaZi calculator actually does walks through all four.

The day and hour pillars are counting. The day pillar runs through an unbroken sixty-day cycle, and the hour pillar follows from it. A model can count, and these two are rarely where charts go wrong.

The year and month pillars are different. They do not turn on dates. They turn on the sun.

The Hong Kong Observatory describes the 24 solar terms as the sun's yearly path divided into 24 equal parts, 15 degrees of solar longitude apart. Twelve of those terms open the twelve BaZi months. Each one arrives at a precise instant, the moment the sun reaches that longitude, and the instant moves from year to year because a calendar year is not a whole number of days. The 24 solar terms and your month pillar covers the full set.

That is the whole problem in one sentence. A chart depends on where the sun was at the minute of your birth, and a language model has no way to see where the sun was. It can only remember what text said about it, and text rounds.

The two traps

Near a boundary, two traps do almost all of the damage.

The year trap. The lunar New Year is the date most people know, so it is the date a model reaches for. But in standard Four Pillars practice the BaZi year turns at 立春, the Start of Spring, not at Lunar New Year. The Observatory dates the Start of Spring to 4 February 2026 at 04:02 Hong Kong time. Lunar New Year 2026 fell on 17 February. For those thirteen days the two calendars disagree about what year it is.

The error does not only run one way. In 2025 Lunar New Year came first, on 29 January, and the Start of Spring came after, on 3 February at 22:10 Hong Kong time. So a rule of thumb that fixes one year breaks the next. There is no shortcut that avoids looking the moment up.

The month trap. The BaZi month does not turn on the first of the month, and it does not turn at midnight on the day of the solar term either. It turns at the minute the term arrives. Someone born on the morning of a term day and someone born that evening can hold different month pillars. The month pillar decides your chart structure 格局 and the order of your luck pillars 大运, so this is not a small slip. It changes the reading from the ground up.

The test

Four births, all in Hong Kong time (UTC+8), each sitting close to one of the boundaries above. The pillars are listed year, month, day, hour. Give a chatbot the date, the time and the city, ask it for the Four Pillars, and compare.

A few things to watch for as the answers come back.

The first birth is the subtle one, because the lunar year has already turned and the BaZi year has not. The correct year is 甲辰. A chatbot that says 乙巳, the Snake year, has used the wrong boundary.

The second and third births sit either side of the same instant and belong to different years. The first is still 乙巳, the second is already 丙午. If a chatbot gives them the same year, it is working from the lunar calendar.

The fourth birth is the month trap. Cold Dew arrives on 8 October 2026 at 14:29 Hong Kong time. Born at 13:30 the same day, the month pillar would be 丁酉. Born at 15:30, it is 戊戌. A chatbot that names the month from the date alone will get one of those two wrong, and it cannot know which without the minute of the term.

Then ask the same question again in a fresh conversation. A calculation gives the same answer twice. A recollection sometimes does not.

What we are not claiming

We are not publishing a hit rate for any chatbot, and that is deliberate.

Several pages in this corner of the web say that chatbots get BaZi charts wrong, and they may well be right. None that we read disclosed a sample size, a model version or a count of errors. One set of accuracy figures, repeated across more than one page, is attributed to a research paper that we could not find anywhere. A number with no method behind it is not evidence, and a number attributed to a paper that does not exist is worse than none. So we are not adding another one.

We are also not claiming a chatbot cannot do this. A model connected to a real ephemeris or a proper BaZi calculator, and actually using it, can cast a chart correctly. The problem is narrower and more practical. When you ask a chatbot for a chart, you usually cannot see whether it calculated or remembered, and it sounds equally sure either way. The test above is how you find out.

What people actually use chatbots for

It is worth being fair about why people ask a chatbot in the first place. A 2026 study of fortune-telling with language models in China, by Lin and colleagues, analysed 1,045 social media posts and interviewed 20 users. It found that people treat these answers less as accurate prediction and more as emotional support.

That matches something we believe too. A BaZi reading is worth having for the recognition it gives you, the sense of being described accurately, not for a forecast. We do not claim to predict events and nothing in the maths lets anyone honestly claim that. But recognition only works if it is recognition of you. A fluent reading built on the wrong month pillar describes someone else, and it will still sound convincing.

How we calculate it

The Eight computes the solar terms directly rather than looking them up in a table. It works out the sun's apparent position from a planetary model called VSOP87, corrected for the slow drift of the Earth's rotation and for the small wobble of its axis, then searches for the exact minute the sun crosses each term's longitude.

We check that against the Hong Kong Observatory's published terms for 2024 through 2027, and it agrees within a minute on every boundary in the test above. That was not always true. Until August 2026 our own model ran a few minutes early against the Observatory, enough to put two real boundaries on the wrong calendar day in some time zones. We replaced it. We mention this because the same honesty we ask of chatbots should apply to us, and because it shows how small an error it takes.

One more thing is built in. If your birth falls within an hour of a solar term, your chart says so, names the neighbouring chart you would have if your recorded birth time were slightly off, and leaves the call with you. Birth times are often rounded, and near a boundary that rounding matters more than any model's precision.

Check your own chart

If you were born in the first week of February, or within a day of any solar term, this is worth five minutes. If you were born in the first days of October, note that the month of 戌 does not begin until Cold Dew on 8 October. Anything that describes the whole of October as one BaZi month is wrong for the first week of it.

Cast your Four Pillars free on your chart. It is computed from the sun's actual position at your birth minute, it tells you if you sit on a boundary, and it gives you eight characters worth taking to any reader, chatbot or otherwise.