November 3, 2026
·Jackson
When You Ask an AI About Your Chart, the Numbers Should Be Real
People bring astrology and Human Design to AI at hard moments. We tested 210 birth charts: Claude on its own got 1 fully right. With a calculation tool, 209.
People often bring their chart to an AI at a hard moment. After a breakup. In a job change. In a rough week. They tell it things they might not tell a friend, and they ask what the sky says about it.
A reading lands harder in a moment like that. So the chart underneath it has to be right. If it is wrong, the reading is wrong, and nothing on the screen tells you.
What we found
We gave Claude 210 invented birth dates and asked for a birth chart each time.
On its own, Claude declined 40. Of the 170 it answered, 169 had at least one planet in the wrong sign. One chart in 210 was fully right. Every reply warned that its figures were estimates. The warning did not fix them.
We ran the same 210 with Open Ephemeris connected. 209 charts were fully right.
Claude usually works out the Ascendant by hand and takes the planets from memory. The Sun was almost always right, and so were the slow outer planets. Mars was in the wrong sign in about four answers out of five, and Venus in about seven out of ten.
Why it is hard to notice
A wrong chart reads like a right one. It has all the usual parts, in the usual order, in confident language. Unless you already have your real chart open next to it, you have nothing to check it against.
That is the problem. In a hard moment, you are not likely to check the chart against a second source.
What changes with a calculation tool
The AI stops guessing the positions. It asks a tool for them, and the tool computes them from NASA JPL DE440, the same data used to navigate spacecraft. Then the AI does what it is good at, which is explaining what the positions mean.
Open Ephemeris does this inside Claude and ChatGPT. You ask for your chart in the conversation and the real one arrives, without leaving the chat or opening a separate program.
What we do not know yet
- We tested Claude here. We have since tested ChatGPT, Gemini, Grok and DeepSeek.
- We tested birth charts. Human Design also starts from an exact birth moment, so the same risk applies, but we have not measured it.
- The tool is not perfect. With the tool, 3 of 290 replies were wrong. All three had the same cause: Claude replaced the tool's historical time zone with its own assumption.
- We make the tool. This is our test, not an independent one. The method is on the proof page, and the data, code and every raw reply are public, so anyone can check or repeat it.
