# Why LLMs Hallucinate Astrology Why LLMs hallucinate astrology: why the genre is the perfect trap, the six shapes the errors take, and the four defences compute... Author: Sandesh · Published: 2026-08-09 · Updated: 2026-08-10 · Section: Founder Notes · Category: Behind the Build · Reading time: 8 min · Publisher: Supastro (Quantum Karma), Bangalore, India People talk about AI hallucination like it is a bug waiting to be patched. It is not a bug. It is the core behaviour of a system trained to produce the most plausible next words, whether or not those words are true. A language model does not lie, because lying would require knowing the truth. It fills gaps, fluently and confidently, and in astrology the gaps are enormous. That is the whole problem, and this final note is about it. I have spent six entries describing a stack - determinism, strength, timing, grounding, verification - that probably looked like over engineering if you did not know what it was defending against. This is what it was defending against. > **What hallucination actually is** > > A hallucination is not a wild error. It is plausible text that has come unmoored from fact. It reads exactly like a correct answer, because sounding correct is the only thing the model was optimising for. In astrology, that camouflage is perfect: the genre already sounds authoritative. ## Why astrology is the perfect trap Three things combine to make astrology unusually easy to hallucinate, and unusually hard to catch. - **The surface is easy to imitate.** Astrological language is patterned and evocative, and a model has read oceans of it. It can produce the sound of a reading effortlessly. - **The facts underneath are precise, and the model never computed them.** Positions, strengths, and periods are exact numbers. A model asked for them will produce a number-shaped answer that was never calculated. - **There is no visible ground truth.** A reader cannot glance at a paragraph and see that the current period lord is wrong. The mistake is invisible at exactly the moment it matters most. Put those together and you get the perfect trap: a machine fluent in the surface, blind to the facts, talking to someone with no way to check it. ## The specific ways it goes wrong In practice, astrological hallucination is not random. It shows up in a handful of recognisable shapes: | Failure | What it looks like | Why it happens | | --- | --- | --- | | Invented positions | A planet placed where it is not | The model estimates instead of computing | | Wrong current period | A confident but incorrect period lord | Timing needs a schedule it never ran | | Fabricated numbers | Strengths or counts that were never derived | A number-shaped gap gets filled | | Misquoted rules | A classical principle stated slightly wrong | Recall from memory, not from the text | | Confident dates | A precise year with nothing behind it | Fluency abhors a vacuum | | Sycophantic agreement | Folding the moment you push back | Trained to be agreeable, not correct | ## A hallucination in the wild Here is what one looks like, because the abstraction undersells how convincing it is. Ask an ungrounded system about your career and it might write: your current period is ruled by a strong benefic in your house of work, so the next eighteen months strongly favour a leadership move. That sentence is fluent, specific, and encouraging. It is also four claims stacked together, the current period, the ruling planet, its strength, and the window - any of which the model may have invented, and all of which read identically whether they are true or not. A reader has no way to separate the true version of that sentence from the fabricated one. They feel the confidence, act on the window, and only much later notice that nothing lined up. The reading was not malicious. It was simply unmoored, and unmoored is more than enough to do real harm. ## Why it matters more here than almost anywhere Hallucinated trivia is annoying. Hallucinated astrology is a different category, for three reasons. People bring real, vulnerable decisions to a reading. The output carries an authoritative tone that quietly discourages doubt. And an app repeats the same confident mistake for everyone who asks, so [scale](https://supastro.com/founder-notes/ai-astrology-vs-computational-astrology) turns a single wrong reading into a policy. > **The uncomfortable truth** > > A more fluent model does not hallucinate less about astrology. It hallucinates more convincingly. Progress on language is progress on camouflage, not on truth. ## How you actually prevent it You cannot scold a model out of hallucinating. You engineer around it, and that is precisely what the previous six notes describe. Four defences, in order: 1. **Compute the facts, do not ask for them.** Positions, strengths, and periods are derived by a deterministic engine, never estimated by the model. 2. **Ground the meaning.** Retrieve real classical passages and reason over those, instead of trusting the model's memory. 3. **Constrain the task.** The model explains facts it is handed; it is never the source of them. 4. **Verify the output.** Check every checkable claim against the computed truth before anyone reads it. Each defence closes one of the gaps that hallucination pours into. Remove any single one and the water finds its way back in. ## What verification actually catches The last defence deserves a closer look, because it is where hallucination goes to die. After the words are written, a validator reads them back against the facts and flags the tells: a cited strength value that does not match the computed one, a current period lord that is wrong, a number that was never derived, a timing claim with no window, an answer that caved the instant it was questioned. The reading stays free to interpret. It is simply not free to contradict the math or dodge the question. Some of those checks are blunt and effective. A time-bound question answered with no actual window is flagged as vague. An answer that reverses itself the moment the user objects is flagged as a capitulation. A cited count that falls outside the range the math allows is caught before it ever prints. None of this asks the model to be honest. It asks the output to survive an inspection it did not get to write. > **The reframe** > > Hallucination is not a moral failing of the model to be lectured away. It is a system-design problem to be engineered around. Blaming the model is like blaming water for being wet. ## The honest limit I will not claim this makes hallucination impossible. A model can still phrase something clumsily or lean on the wrong emphasis. What it cannot do, inside this stack, is invent a fact that survives the check. The failure mode shrinks from confidently wrong about your life to occasionally awkward in its wording and those two risks are not remotely the same. That is not a marketing distinction; it is the difference between an app you can hand to someone at a fragile moment and one you quietly cannot. ## Why the whole series exists If you have followed all seven notes, here is the shape of it. Every layer I described is an answer to this final question. Determinism answers invented positions. Strength answers empty confidence. Timing answers fabricated dates. Grounding answers misquoted rules. Verification answers all of them at once. The stack is not there to be clever. It is there so a reading can be trusted by someone who cannot check it themselves, which, when you are the one asking about your own life, is everyone. That is the end of the tour, and the whole argument in one line: fluency is cheap now, and getting cheaper, so the only thing worth building is the part fluency cannot fake. > The goal was never a model that sounds right. It was a system that is right, and can prove it. --- ## FAQ #### Why do language models hallucinate at all? Because they are trained to produce plausible text, not verified facts. When a model lacks a fact, it generates something that fits the pattern. That is useful for prose and dangerous for anything with a right answer. #### Can't a better model just stop hallucinating astrology? A better model hallucinates more convincingly, not less. Fluency is not accuracy. The fix is architectural - compute, ground, constrain, verify - not a matter of waiting for the next release. #### How can I tell if a reading was hallucinated? Ask for the specifics behind a claim: the position, the period, the number. A grounded system can show its work; a hallucinating one deflects. Cross-checking a timing claim is the fastest tell of all. #### Isn't some hallucination fine for something as loose as astrology? That is the trap. Because astrology sounds loose, wrong facts slip through unchallenged, and yet people still make real decisions on them. The apparent looseness is exactly why the underlying facts have to be rigid. Interpretation can breathe; the chart cannot. #### Does Supastro hallucinate? The stack is built to make hallucination fail loudly instead of pass silently. Facts are computed, meaning is grounded, and every claim is checked against the math. It is not magic, but it removes the failure that matters most. ### The full series - [Phase 1: How Supastro Is Built](https://supastro.com/founder-notes/how-supastro-is-built) — The five-layer architecture - [Phase 2: How Deterministic Engines Work](https://supastro.com/founder-notes/how-deterministic-astrology-engines-work) — Answering invented positions - [Phase 3: AI vs Computational Astrology](https://supastro.com/founder-notes/ai-astrology-vs-computational-astrology) — Order of operations - [Phase 4: Computational Vedic Intelligence](https://supastro.com/founder-notes/what-is-computational-vedic-intelligence) — The four pillars - [Phase 5: Why Planetary Strength Matters](https://supastro.com/founder-notes/why-planetary-strength-matters-in-ai-astrology) — Answering empty confidence - [Phase 6: Reasoning Across Dashas and Transits](https://supastro.com/founder-notes/how-ai-reasons-across-dashas-and-transits) — Answering fabricated dates **Read astrology that can prove it** Build your chart and run a reading where the facts are computed, grounded, and checked before you ever see them. [Open Supastro](https://supastro.com/dashboard) --- Canonical: https://supastro.com/founder-notes/why-llms-hallucinate-astrology · Supastro is an AI astrology platform for Vedic astrology (Jyotish), built as Vedic Super Intelligence (SI), by Quantum Karma, Bangalore, India. Chart positions are computed from NASA JPL ephemeris data on the Lahiri (Chitrapaksha) ayanamsa. 40 instruments are free to use.