Level 7 · Lesson 1 of 5

What AI does reliably — and what it fabricates

8 min · Goal: Separate the parts of an AI reading that can be trusted from those that cannot.

Two different systems in one answer

Every AI astrology tool does two quite different things, and they have completely different reliability. It calculates positions, and it writes about them.

Calculation is deterministic. An ephemeris routine takes a date, time and place and returns planetary longitudes, and the same inputs always produce the same output. There is no judgment involved and nothing for a model to invent. On this site the numbers come from an astronomical engine, not from the language model — the model never touches them.

Interpretation is generation. A language model produces plausible text about the positions it is given. That text is fluent, usually reasonable, and sometimes wrong in ways that are difficult to detect precisely because it is fluent.

How models fabricate in this domain

Four patterns, all worth recognising.

Invented classical citations. A model will attribute a rule to Parashara or Jaimini with confidence when no such passage exists. Astrological literature is vast, old and unevenly digitised, which makes this both easy to do and hard to check.

Smoothed disagreement. Where schools genuinely conflict — neecha bhanga conditions, retrograde interpretation, whether Rahu can be exalted — a model tends to produce one confident answer rather than reporting the dispute. The confidence is the failure.

Overstated base rates. A model trained largely on astrology content written to impress will describe Gajakesari as a rare and fortunate combination, because that is what most of its training data says. The arithmetic from Level 4 is not in the corpus.

Arithmetic drift. Ask a model to compute a dasha balance in prose and it will often produce a plausible number that is wrong. Language models are poor calculators, which is exactly why the calculation should not be left to them.

The useful heuristic

Trust the numbers if you know they came from an ephemeris. Treat everything narrative as a first draft by a well-read but overconfident student.

That is not a dismissal. A well-read student is genuinely useful — they will notice placements you missed and phrase things clearly. But you would check their claims before repeating them to someone who is worried about their marriage.

Try it with your own chart

Catch a fabrication

  1. Ask any general AI assistant for the classical source of a specific rule — the conditions for neecha bhanga are a good test.
  2. Ask twice in separate conversations and compare the answers.
  3. Note whether it reports the disagreement between schools or gives one confident version.
  4. Now ask it to calculate a dasha balance from a Moon position and check it by hand using Level 6.
Compare with our own tool

Key points

Check yourself

An AI tells you that Parashara specifies four conditions for neecha bhanga and lists them. What should you do before repeating it?

← Level 1 overviewVerifying the computation