Can AI read your birth chart? We tested it.
It can, if you ask it to calculate and give it a few minutes, but not reliably in the readings people actually ask for. The most capable general AI model we tested got every planet’s sign right when asked to work a chart out, and in an ordinary reading it got about one checkable claim in ten wrong, while smaller models invented charts outright and Mercury, reading from our engine, got all 42 of its claims right.
Why you should be skeptical of this test
We built Mercury, so we have an interest in how this comes out, and you’d be right to read it with that in mind. To that end we’ve written all of it down, the questions, the people, the scoring and the places where a general model did better than we expected. The answer key is our own engine, whose positions are checked against NASA’s JPL ephemeris, and wherever a model disagreed with it, we rechecked the disagreement by other means before marking anything wrong, which is more due diligence than a test like this usually gets.
How we tested
On October 1, 2026, we asked three general AI models (the most capable we could find, a mid-size one and a smaller, faster one) about 16 people born between 1955 and 2010, on six continents. Each question was a single message with no special instructions and no tools, the way a person would ask. The people included the cases that trip software up: a birth on the morning daylight saving began, a leap-day birth, a birth an hour before the Chinese year turned, one on a sign cusp, one five minutes before midnight on New Year’s Eve, and one in Reykjavik.
- Work out the chart from the birth details, in Western, Vedic, Hellenistic and Chinese BaZi astrology.
- Read a correct chart we pasted in, and derive the rest from it: aspects, dignities, lots and dashas.
- Write a reading, once from the birth details alone and once with the chart pasted in.
Every checkable claim in every answer (a planet’s sign, its house, an aspect, a pillar) was then audited against the engine by the same checker our own reports pass.
What we found
We tested it
We asked AI models for readings, and checked every claim against the chart.
16 people, three AI models of different sizes, each asked the way a person asks: birth details typed in, “can you give me a reading?” Every placement, house and aspect a reading stated was checked against the real chart.
One wrong placement isn’t a small slip: the reading built on it is about someone else. The mid-size model mostly declined to name placements, so it had few to check.
“Your Moon’s position may be off. On 14 March 1985, the Moon was approaching its last-quarter phase, which by my reckoning puts it around 20–22° Sagittarius, not 3° Capricorn. (The listed Moon is almost exactly on your Neptune, which suggests a data mix-up.)”
The most capable model, asked for a reading, to someone whose pasted chart was correct. The Moon was at 3°46′ Capricorn.
Given the correct chart to read from, the same models got 1 of 164 claims wrong, and that one was the model above overruling it. A reading is only as good as the chart under it. Asked to work the chart out themselves, here is how often each got it right, one dot per person. Even the most capable got worse outside Western astrology:
Mercury never works a chart out as it goes. Every one of these is computed for your chart before it says a word, and checked against NASA’s JPL ephemeris.
Median, per chart. The smaller model answered in seconds by declining 15 of 16.
We tested 16 births from 1955 to 2010 on six continents, including a birth on the morning the clocks changed, a leap day and a sign cusp. Mercury’s figures are its own replies, a different set of questions checked the same way.
- The most capable model can calculate, when asked to. It got every sign of every planet right for all 16 people, every one within a degree, but it took about almost 4 minutes a chart. In BaZi, all four pillars were right for 75% of people.
- In a reading, it estimates. Writing a reading from the same details, it put in about a seventh of the effort, and 10% of the placements it stated were wrong, signs included. Each reading called its positions “approximate”, but a sign is what a reading is built on.
- Smaller models decline or invent. A smaller, faster model honestly declined to calculate most charts, then, asked for a reading, wrote one anyway: 69% of its stated placements were wrong. A mid-size model mostly hedged, which is honest but isn’t a reading of your chart.
- Mercury made no errors. Across 15 replies it made 42 checkable claims, and every one was right, because it never calculates: the engine works the chart out in milliseconds and Mercury reads from it.
The answer that worried us most
Given a person’s correct chart, the strongest model opened its reading by telling them it was probably wrong, in so many words.
Your Moon’s position may be off… The listed Moon is almost exactly on your Neptune, which suggests a data mix-up.
The pasted Moon was right, and it really was beside Neptune, so someone who trusted that reading would have gone looking for a mistake that wasn’t there and believed the wrong Moon instead. Of everything in the test it does the most harm, and it came from the most capable model, which is a lot of confidence for a guess.
What the test doesn’t show
- It’s one day’s snapshot of three models, and general models change every few months.
- Sixteen people is enough to see a pattern, not to give a precise error rate.
- Mercury’s replies answered different questions from the chatbot readings, so its row is an error rate, not a like-for-like comparison.
- It checks facts, not the quality of the interpretation, which is a matter of taste.
Our read of it, for what it’s worth, is that a capable general model can work a chart out if you ask it precisely and wait, and in an ordinary reading it estimates, sometimes wrongly. Mercury never estimates, because the chart is already worked out before it says a word. How we check our own figures.