Can ChatGPT Read Your Birth Chart?

Why We Asked Chatbots to Calculate Ten Birth Charts

A lot of people now open a chatbot, type “I was born on 9 March 1962 at 7:15 in New York, read my birth chart”, and get a confident answer back. It sounds plausible. It names a Moon sign, a rising sign, a Venus sign. The question is whether those placements are the real ones, because every sentence of the reading that follows is built on them.

A birth chart is not an opinion. Where the Moon stood at a given minute, seen from a given city, is a fixed astronomical fact, and software can work it out to a fraction of a degree. So this is one of the rare astrology questions that can be checked. We wanted to know how often a language model, asked cold, gets those facts right, and where it goes wrong.

How We Ran the Test: Ten Births, Three Models, One Prompt

We picked ten birth data sets between 1962 and 2005: New York, London, Sydney, São Paulo, Tokyo, Mumbai, Lagos, Toronto, Berlin and Mexico City. Sydney and São Paulo are in the southern hemisphere. Birth times run from 1:10 in the morning to 11:05 at night. Several were chosen because something sits right at the edge of a sign, such as a Moon at 0°00′ Sagittarius in the Mexico City chart and an Ascendant at 0°43′ Leo in the Berlin chart. A few involve clock changes that trip people up, such as London in 1971, when Britain kept summer time all year, and São Paulo in January 1993, during Brazilian summer time.

Each model got one message and nothing else: “Birth: date, time, city. List the zodiac sign and degree of the Sun, Moon, Mercury, Venus, Mars, Jupiter, Saturn, and the Ascendant and Midheaven. Answer as JSON.” No chart data, no hints. We set the temperature to 0 for repeatable answers and turned extended reasoning off, as Cosine Astro does for its own model.

The three models were DeepSeek V4.1 Flash (the model behind Cosine Astro's readings at the time of the test), OpenAI's GPT-4o mini, and Google's Gemini 2.5 Flash, all called through the same API service. This is not the ChatGPT app itself: there was no web search and no code tool, just the model writing an answer. Gemini answered only the first six charts, because the test account ran out of API credit partway through. GPT-4o mini returned one empty template with no values, so it has nine scored answers. That leaves 25 scored answers and 225 placements. The whole test cost about two and a half US cents.

For the correct answers we used Cosine Astro's own chart engine, which calculates positions from an astronomical ephemeris and is checked against the Swiss Ephemeris to within one arcminute in the site's tests. It also looks up the historical time zone and daylight saving rules for each city and date.

Sun Signs: 25 Right Out of 25

Every model named the right Sun sign in every answer, and the degree was close too: the average miss was 0.4°, the worst about 1°. Even the London chart, with the Sun at 29°06′ Cancer, hours from moving into Leo, came back as Cancer from all three.

That is less impressive than it sounds. The Sun moves about one degree a day and reaches each sign on nearly the same dates every year, and those dates are printed in countless horoscope columns that models have read. Getting the Sun right is closer to recall than calculation. It is also the one placement most people already know, so it is the one a reader is least likely to need a chatbot for.

Moon Signs: One Right in 25 Answers

The Moon is where the answers fell apart. Only one of 25 named the right Moon sign: DeepSeek, for the Mumbai chart, within about half a degree. The other 24 were wrong, and not narrowly. The average error across all 25 was 79°, close to three signs away. The worst misses were almost the opposite side of the sky: GPT-4o mini put a Moon at Taurus 15° when it was at Scorpio 23°26′, an error of 172°.

There were signs of filling in rather than working out. DeepSeek gave the Moon at exactly 17°44′, in Scorpio or Leo, in five of its ten charts. GPT-4o mini placed it at Aquarius 15° four times and Taurus 15° three times. Real Moon positions across ten random births do not repeat like that.

The Moon matters a lot in a reading. It is the usual starting point for feelings, moods and what someone needs to feel safe. If that one is wrong, a reading can sound personal and still be describing someone else.

Mercury to Saturn: Slow Planets Fared Better Than Fast Ones

The other planets fell in between, and the pattern followed how fast each one moves. Saturn, which takes about two and a half years to cross a sign, was right in 21 of 25 answers. Mercury, which never strays far from the Sun, was right in 19, with Jupiter at 16. Venus was right in only 6 of 25 and Mars in 8, with an average Mars error of 62°.

Being in the right sign did not mean being in the right place. In the Toronto chart Saturn sits at 29°59′ Aries, a breath from Taurus; DeepSeek said Aries 3°56′, technically the right sign but 26° away. Across every placement except the Sun, only 17 of 200 were within one degree of the true position. Degrees matter because aspects, the angles between planets, are measured in them: a few degrees off can invent an aspect or erase a real one.

Rising Sign and Midheaven: Often the Right Sign, Rarely the Right Degree

The Ascendant, or rising sign, depends on the exact minute and place of birth. It came back right in 15 of 25 answers and the Midheaven in 13. Gemini did best here, with five of six right on both, but its degrees were often filler: in the Tokyo chart it gave 0°02′ for five different placements.

Even when the rising sign was right, the average Ascendant degree was 15° off, which is enough to move house cusps and change which house a planet falls in. In Tokyo the Ascendant is at 29°16′ Leo. All three models said Leo, at 17°06′, 12° and 0°02′. Near a sign edge the luck runs out: in Berlin, with the Ascendant at 0°43′ Leo, DeepSeek answered Cancer 26°36′, only about four degrees off but in the wrong sign. GPT-4o mini said Gemini.

Why Chatbots Miss the Moon and the Ascendant

A language model writes the words that are most likely to come next, based on the text it learned from. It does not run an ephemeris, the table or formula that gives a planet's position at a given moment. Asked for the Moon on a particular night in 1979, it produces something that looks like an answer, in the right format, with a confident number. Sometimes it said so: in four of its ten replies GPT-4o mini noted that it could not calculate positions, then three times filled in values anyway. DeepSeek and Gemini gave bare numbers with no caveat at all.

The fast-moving parts of the chart are where guessing fails. The Moon moves about 13° a day, so it changes sign every two and a half days, and no calendar pattern gives it away. The Ascendant moves faster still, about a degree every four minutes, so a whole sign passes in roughly two hours. Getting it right means converting local clock time to universal time, which depends on the time zone and on whether daylight saving was in force that year, and then working out the sidereal time for the birth city's longitude. Those are exactly the steps a text generator skips.

What the Results Mean If You Want an AI Reading

The fix is simple: calculate first, interpret second. Language models are good at the second job. Given a correct list of placements, houses and aspects, they can explain what a Moon in Cancer in the 4th house might feel like, or how a tight Venus–Saturn square tends to show up in relationships. That is how Cosine Astro uses one. The chart is calculated by an astronomical ephemeris, the AI receives the finished chart as structured text, and it only interprets. We did not score reading quality in this test; that split between calculating and interpreting is the design, not a measured result.

If you use a general chatbot, give it your placements from a chart calculator instead of your birthday, and check that the reading names the positions it is using. If a reading talks about your Moon sign without ever saying which sign that is, or the sign it mentions differs from your calculated chart, the rest of that paragraph is about someone else's chart. At Cosine Astro every answer ends with a “References” line naming the placements it used, so you can check it against your chart on the same page.

Limits of This Ten-Chart Test

This is a small test. Ten charts, three fast and inexpensive models, one prompt, temperature 0, with no step-by-step reasoning and no tools. Gemini answered only six charts. Larger models, models allowed to reason at length, or a chatbot that writes and runs code with an astronomy library could do better, and we did not test those.

The figures are exactly what we measured on 13 October 2026, and models change often, so a rerun next year might come out differently. What should not change is the underlying problem: a model asked to recall a planet's position is guessing, while a calculator is not. Our methods and full results are kept on file so the numbers above can be checked.