The short answer
It is good at explaining and unreliable at calculating.
A language model produces text that resembles a correct answer. It is not performing astronomical calculation unless it has been connected to a tool that does, and the difference is invisible in the output.
So a chart it generates arrives formatted, confident and complete, with degrees to two decimal places. Some of it will be right. The parts that are wrong will look exactly like the parts that are right.
What it is genuinely good at
This is not an argument that the technology is useless, and pretending otherwise would be as unhelpful as overclaiming for it.
Explaining concepts. What is a yuti. What does the 7th house traditionally cover. How does a dasha system work. What is the difference between sidereal and tropical. For learning vocabulary it is faster and considerably more patient than a book.
Summarising traditions. Telling you what different schools say about the same placement, and where they disagree.
Drafting and organising. Turning your own notes into something structured, or helping you frame a question before a consultation.
Those are real uses and they are worth having.
Where it fails
Calculation, and therefore anything depending on the exact chart.
The largest single failure is ayanamsa, the correction between the sidereal zodiac used in Vedic astrology and the tropical zodiac used in Western astrology. It is currently around 24 degrees, which is most of a sign. Miss it and roughly four people in five get the wrong sign for most of their planets, while the output looks entirely normal.
After that: historical time zones and daylight saving, which are genuinely difficult and frequently wrong; the rising sign, which depends on exact time and precise coordinates and is the value everything else is built on; and retrograde status, which is a simple lookup and a common invention.
Why ChatGPT can get your birth chart wrong sets out all four with five checks you can run in about two minutes.
The division of labour that works
Let software calculate. Use an actual ephemeris-based tool for the chart. Our kundli tool calculates rather than generates, which makes it a reasonable thing to work from or to check against.
Let the model explain. Once you have a correct chart, asking what a placement traditionally means is exactly the kind of question it handles well.
Let a person judge. Which of the indications matter for your question, and what to do about it.
That third step is where the remaining gap sits, and it is worth being specific about.
The thing it cannot do at all
Suppose the chart is correct. Something is still missing, and it is not more knowledge.
A reading is not only interpretation. It is a judgement about what to say, made by someone accountable for the effect of saying it.
Which two of the fifteen visible indications actually carry this person's question. Which findings are true but would only frighten them. When the honest answer is that the chart does not support a clear view. When the answer is "see a doctor". And when a question should be refused outright.
Ask a model about lifespan and it will usually produce something, because producing a plausible response is what it does. A practitioner refuses, and the refusal is part of the method rather than a gap in it. Why we do not predict death sets out that reasoning.
A model will also give you a confident reading of every indication it can see, rather than the two that matter. Weighing is most of the skill, and weighing requires knowing what the reading is for.
The accountability point
This is not a claim that people are mystically better readers.
It is a claim about consequence. A practitioner who tells someone something frightening and wrong has done damage they are answerable for. That answerability is what produces caution, and caution is what produces the refusals.
A model has no stake in whether you are all right afterwards. That is not a criticism of the technology; it is a description of what it is.
If you have already had an AI reading
Two things worth doing before you conclude anything.
Verify the chart. Run the same details through a calculating tool and compare the rising sign and the Moon. If either differs, nothing in the reading was about you.
Discount the confidence. The tone of an AI reading carries no information about its accuracy. Neither does the tone of a human one, but people are at least used to discounting that.
If the reading described a life you did not recognise, the likely explanations in order are: the wrong ayanamsa, a time zone error, a rising sign error, and only then anything about you.
Where this leaves you
Use it to learn. Do not use it to decide.
If you want the mechanics of where it goes wrong and how to check, why ChatGPT can get your birth chart wrong. If you want to know how to tell whether any reader, human or otherwise, has a method, why structured frameworks matter gives four questions.
And a BNN birth chart reading starts by checking the chart against events that have already happened, which is the one test neither a model nor a confident astrologer can talk their way around.