Why ChatGPT Gets Your Saju Wrong

2026-06-11

There’s a prompt going around. You’ve probably seen it — a long block of text you paste into ChatGPT along with your birth date, time, and city, and out comes your saju (사주), the Korean Four Pillars reading that maps a life from the moment it started. The TikTok videos follow a formula: someone reads their result aloud, goes quiet at how accurate it feels, stitches a reaction. The comments fill with people trading pillars like star signs at a party.

It’s a genuinely fun trend, and we’re not here to ruin it. We’re here to point out one small structural problem: for a real portion of the people running that prompt, the chart underneath the reading is wrong. Not “debatable interpretation” wrong. Wrong the way a misdialed phone number is wrong — everything after the mistake is a conversation with a stranger.

A Saju Chart Is a Calculation, Not a Vibe

Before anyone interprets anything, a saju chart has to be computed. Four pillars — year, month, day, hour — each built from one heavenly stem and one earthly branch, all derived from astronomical time. The interpretation can be loose, intuitive, even playful. The calculation cannot. It’s closer to figuring out a tide table than reading tea leaves.

ChatGPT does not calculate. It predicts plausible text. Hand it a birth date and it doesn’t consult an ephemeris; it produces stems and branches that resemble the ones that usually accompany dates like yours in its training data. For a birthday parked safely in the middle of a month, that educated guess often lands. Near a boundary, it quietly falls apart — and a language model never announces the moment it starts guessing. It hands you wrong pillars with the same serene confidence as right ones. That’s the whole problem in one sentence: the failure is silent.

The Month Changes at a Minute, Not at Midnight

Here is the part almost no viral prompt accounts for. In saju, months are not calendar months. They begin at jeolgi (절기) — solar terms, the moments the sun crosses specific points along its annual path. A solar term does not arrive at midnight on a tidy date. It arrives at an exact minute — say, 4:17 on a Tuesday afternoon — and at that minute, the month pillar flips. Before it, you’re one month. After it, you’re another.

Two babies born in the same hospital on the same date, a few hours apart, can carry different month pillars because a solar term passed between their cries. And the timing isn’t fixed: it drifts from year to year, and it depends on where on Earth you were born, because a minute in Seoul is a different minute in Chicago. There is no lookup-by-date shortcut. You need the actual astronomy.

ChatGPT does not have the actual astronomy. It has a hazy statistical memory of roughly when solar terms tend to fall — which is like navigating a harbor at night with a postcard of the harbor. Charming souvenir. Terrible instrument.

The January Problem

The year pillar hides the same trap, with higher stakes, because the year pillar carries your zodiac animal — the part people are most attached to.

The saju year does not begin on January 1. It begins at Ipchun (입춘), the solar term marking the start of spring, which lands around February 4 — at, once again, an exact minute that moves from year to year. Which means anyone born in January, and many people born in early February, belongs to the previous year’s pillar and the previous year’s animal.

Born in January 1990? By the Gregorian calendar, a Horse. In saju, you arrived before Ipchun, so your year pillar belongs to 1989 — you’re a Snake. ChatGPT, reasoning from the calendar printed on mugs and placemats, will routinely hand January babies the wrong animal. And because the pillars interact — the year shapes how the month is derived, the elements all weigh against each other — one wrong pillar doesn’t stay contained. It seeps. The reading downstream is fluent, detailed, emotionally resonant, and about somebody else.

Korean grandmothers have guarded the Ipchun rule for generations. The prompt has not.

It Can’t Check Its Own Work

Suppose the model somehow recites the correct method. It still can’t verify the arithmetic underneath. The day pillar, for instance, rides on a sixty-day cycle that has been turning, unbroken, for millennia — computing it means doing real modular arithmetic over a real count of days, which is precisely the kind of operation language models are famous for fumbling. Worse, a model can’t feel itself fumble. There’s no inner abacus going pale. It miscounts, then narrates the miscount in confident, well-organized prose, possibly with a tasteful emoji.

A tool that is wrong in a way that feels right is its own category of hazard. Most wrong things look wrong. This one looks like insight.

What ChatGPT Is Actually Great At

Now the twist, because this isn’t a hit piece: once the chart is correct, ChatGPT is a genuinely lovely interpreter. Give it accurate pillars and it will explain what a strong Fire day master tends to want, how an unbalanced chart might show up at work or in love, what the classical texts say and how a modern reader might hold that loosely. It answers follow-up questions at two in the morning without judgment. It translates a dense, hierarchical tradition into language your group chat can argue about.

The trend’s instinct is exactly right — people want a conversation with their chart, not a laminated printout. The pipeline is just backwards. Interpretation is a language problem, and language models are good at language. Calculation is an astronomy problem, and it should be done by something that does astronomy.

Calculate First, Then Chat

That split is the whole premise of sajucard.com. We compute the solar-term times astronomically — actual solar positions, actual minutes, your actual time zone — and we cross-check the results against independent implementations, so that a boundary birthday lands on the correct side of the line instead of the convenient one. Cast your chart and the term times do their quiet work underneath; the January problem and the 4:17-on-a-Tuesday problem are simply handled.

Then, by all means, carry your correct pillars over to ChatGPT and talk all afternoon. It makes a wonderful conversationalist and a poor astronomer, and there’s no shame in that — most of us are at least one of the two.

And if your next question is what your chart says next to someone else’s, that’s an older tradition than the prompt — we wrote about gunghap, the Korean art of reading compatibility, too.

When you’re ready for pillars that survive the math, cast your chart at sajucard.com.