Why ChatGPT Gets Your Birth Chart Wrong (and Gives a Different One Next Time)

Updated 2026-09-13

ChatGPT gets birth charts wrong because it has no ephemeris, no time-zone history and no Chinese solar-term calendar; it generates placements that sound right instead of computing them. It gives a different chart next time because each answer is sampled with randomness, so the guess is re-rolled in every new window. Deterministic chart engines fix both problems: the same birth data always yields the same chart, and the AI is allowed only to interpret it.

An ephemeris is a lookup; a language model is a prediction of words

An astronomical chart starts from an ephemeris, a table of where the Sun, Moon and planets were at every moment, computed from orbital mechanics. A chart program takes your birth instant in Universal Time, reads positions from that table, then works out the Ascendant and house cusps from the sidereal time at your longitude and latitude.

A language model contains none of that. It has read many charts and learned what a chart usually looks like, so it produces text that resembles one. For the Sun sign it is often right, because dates map to signs simply. For the Moon, which changes sign every two to three days, and for the Ascendant, which changes every two hours or so, resemblance is not enough.

Time zones and daylight saving are where charts break

Converting a local birth time to Universal Time requires knowing the offset in force at that place on that date. Daylight saving rules have changed repeatedly in most countries, wartime and double summer time existed, and some regions used half-hour offsets or switched zones entirely. A model estimating the offset from memory is often one hour out.

One hour moves the Ascendant by roughly a sign, shifts every house cusp, and can push the Moon across a sign boundary. The rest of the reading then describes a different person.

Chinese charts: the solar-term calendar, not the Western month

BaZi and Zi Wei Dou Shu use the stem-branch calendar. The year pillar changes at Li Chun in early February, not on January 1 and not at Chinese New Year. The month pillar changes at each of twelve solar terms, whose dates shift by a day or so every year. The hour pillar follows true solar time, so clock time must be corrected for longitude, and the Zi hour that spans 23:00 to 01:00 raises the question of when the day pillar turns.

Zi Wei Dou Shu adds lunar months and days, leap-month rules and a long sequence of placement rules for fourteen major stars and dozens of assistant stars. A chatbot reasoning from the Gregorian date alone can get the year and day pillars right and still put the month or hour pillar in the wrong place, and its Zi Wei charts are rarely reliable at all.

Why the second answer differs: sampling

Chatbots do not return the single most likely answer; they sample from a distribution of plausible next words, which keeps their prose natural. For a poem that is a feature. For a Moon degree it means the guess is re-drawn every time, which is why two windows with identical birth data return two charts. A calculation that changes on repetition was never a calculation.

How deterministic engines fix it

The chart is computed by code from an ephemeris or the stem-branch calendar, with solar terms and the day-pillar rule handled explicitly. The same input always yields the same output, and an expert can verify every placement.
The language model receives the computed chart and writes only within it. If a placement is not in the chart, it cannot appear in the reading.
Suan (suan.you) works this way across fifteen systems: the engine casts the chart, the AI only interprets. Western natal charts come from an astronomy engine, BaZi changes its pillars at the solar terms, and Zi Wei Dou Shu is placed by rule rather than recalled from memory.
A verifiable chart does not make the interpretation true; it makes it about you.