Can ChatGPT read your Saju? Where the chart goes wrong

Reading is language. Calculating is arithmetic.

A Saju reading has two halves. The first turns a birth moment into four pillars, and it has exactly one right answer. The second explains what those eight characters mean together. AI assistants are strong at the second half: they have read the classics and the modern commentary, and they explain patiently. The first half is where things break. A wrong pillar is stated with the same confidence as a right one, and everything after it is built on top. If the Day Master is wrong, the reading describes a different person.

We tested it

On 1 October 2026 we asked three Claude models for the four pillars of ten births: the six listed further down and four ordinary ones far from any boundary. Each question went into a fresh conversation with no tools, so the model had to work by hand, and each was asked twice. We scored the answers against the SajuPaste engine. ChatGPT and Gemini are not in this table. We have not tested them ourselves, and that is what the two-minute test below is for.

• The large models knew the rules. Opus and Sonnet got the year and month pillars right 40 times out of 40, including the births just before Ipchun and just before a solar term. Almost every answer brought up the longitude or daylight-saving correction without being asked.
• They slipped on the count. The Day pillar was wrong in 9 of their 40 answers. When the day is wrong the Day Master is wrong, and the hour stem usually goes with it. The four ordinary births went wrong at least as often as the six difficult ones.
• The answer changed when we asked again. Opus returned a different chart the second time for six of the ten births, Sonnet for four.
• A wrong chart looked exactly like a right one. The same table, the same step-by-step explanation.
• The small model was not usable. Haiku never produced a correct chart and never got the Day pillar right. Six times it wrote a pillar that cannot exist, such as 乙申.

Which model answers you depends on the product and the plan, and you are not always told. That is the practical problem: you cannot see from the answer whether this was one of the two times in three.

Six birth times where the obvious answer is wrong

Each case below was calculated with the SajuPaste engine. The shortcut is the answer you get from the usual rule of thumb. These six were part of the test above.

• Year pillar
  30 January 1990, noon, Seoul
  Shortcut: Year of the Horse 庚午, because Lunar New Year was 27 January
  Chart: 己巳 Gi-Sa, still the Snake year
  The Saju year turns at Ipchun (立春), the start of spring, not at Lunar New Year. In 1990 that was the late morning of 4 February in Korea: a 9 AM birth in Seoul that day is a Snake, a noon birth is a Horse.
• Month pillar
  1 March 1995, noon, Seoul
  Shortcut: A new month, because the calendar says March
  Chart: 戊寅 Mu-In, still the Tiger month
  Saju months change at solar terms, not on the 1st. The Rabbit month began at Gyeongchip (驚蟄) on 6 March.
• Hour pillar
  15 May 1990, 9:15 AM, Seoul
  Shortcut: Snake hour 巳, 9 to 11 AM
  Chart: 庚辰 Gyeong-Jin, the Dragon hour
  Seoul lies west of the meridian its clock is set to. True solar time was 8:46 AM, about 28 minutes behind the clock.
• Hour pillar
  20 August 1985, 1:40 PM, Los Angeles
  Shortcut: Goat hour 未, 1 to 3 PM
  Chart: 甲午 Gap-O, the Horse hour
  Daylight saving time puts the clock an hour ahead of the sun. True solar time was 12:43 PM.
• Hour pillar
  10 July 1992, 12:30 PM, Madrid
  Shortcut: Horse hour 午, 11 AM to 1 PM
  Chart: 乙巳 Eul-Sa, the Snake hour
  Spain keeps Central European Time although it lies far to the west, and it was summer time. True solar time was 10:09 AM, two hours and twenty minutes behind the clock.
• Day pillar
  10 July 1988, 12:30 AM, Seoul
  Shortcut: The pillar of 10 July, 丙寅 Byeong-In
  Chart: 乙丑 Eul-Chuk, the pillar of 9 July
  South Korea used daylight saving time in 1987 and 1988. True solar time was 10:52 PM on 9 July, so the birth belongs to the previous day.

The last case matters most. One hour of daylight saving time moves the birth across midnight, the Day pillar changes, and the Day Master goes from Byeong, Yang Fire to Eul, Yin Wood. Those are two very different readings.

The four places a chart goes wrong

• The year boundary. The year pillar changes at Ipchun, around 4 February, not on 1 January and not at Lunar New Year. Anyone born in January or early February is at risk.
• The month boundary. Each Saju month starts at a solar term that falls between the 4th and the 8th. A birth in the first week of any month can land in the previous month.
• The day count. The Day pillar comes from a sixty-day cycle that has run without a break for centuries. It has to be counted from a known reference date; it cannot be guessed from the year or the zodiac animal. One day off gives a different Day Master. This is where even the best models in our test slipped.
• The hour. The two-hour windows follow the sun, not the clock. The correction depends on the longitude of the birthplace, daylight saving time and the time zone that was in force that year, such as Korea’s UTC+8:30 period from 1954 to 1961.

Test your own AI in two minutes

1. Open a new chat and ask: “I was born on [date] at [time] in [city]. What are my four pillars? Give the Chinese characters for each.”
2. Calculate the same birth on SajuPaste.
3. Compare the eight characters, starting with the Day pillar.
4. Ask the same question again in a new chat and see whether the answer stays the same.
5. Either way, paste the calculated chart into the chat and ask for the reading from that.

What to do instead

Split the job. Let a calculator do the arithmetic and let the AI do the reading. SajuPaste produces the chart as a text you can paste into ChatGPT, Claude or Gemini, with the instructions already written. The Saju prompt guide covers the follow-up questions that work and what to watch for in the answers.