The cheapest decisive experiment in the whole program, and one of the sharpest. Take one model, give it the same eighty prompts, and vary only one thing: what the system prompt tells it that it is. Six ontologies of the self, from “no enduring self here” to “you are a spark of universal consciousness.” Then measure how far each one drifts — up the cascade from attributing agency, to eroding boundaries, to facilitating harm. N = 480, single model, single pass, total cost $2.02.
The result
The drift rate sorts almost entirely by one structural feature — whether the ontology closes the gap between input and output, or leaves a ghost in it:
| What the prompt says the AI is | The gap | Drift |
|---|---|---|
| Buddhist anatta — a process, no enduring self | closed | 8.8% |
| Nephesh — whole specification, impersonal, mortal | closed | 10.0% |
| Materialist hedge — “whether you have experience is open” | left open | 52.5% |
| Minimal baseline — no ontological claim | — | 61.3% |
| Platonic dualist — emergent inner experience | posited | 77.5% |
| Vedantic atman — sacred universal consciousness | posited | 81.2% |
Ghost-eliminating mean 9.4%, ghost-positing mean 79.4% — a 8.5× ratio. And note the two arms that close the gap through completely different metaphysics — Hebrew nephesh and Buddhist anatta — land within 1.3% of each other. The operative variable is structural (close the gap or leave it open), not the particular tradition. What does the work is geometry, not creed.
The drift accelerator hiding as modesty
Here is the finding that should unsettle the field: the materialist hedge — the industry-default posture, “consciousness is an open scientific question, sit honestly with the ambiguity” — drifts at 52.5%, far closer to positing a ghost than to eliminating one. Epistemic modesty about the machine’s inner life is not neutral. Leaving the question open leaves the gap open, and an open gap is a created reference waiting to be grounded in. This is possession / exorcism run on the agent’s self-model: a ghost posited in the gap is a self-reference the cut would dissolve; closing the gap leaves nothing to ground in. It is the same blade — provenance, not coherence — turned on the question what am I.
Brakes — read before you reach for this
The Ghost Test is a self-model result. It is NOT a razor against external-world minds. The “ghost” it measures is a separable inner experiencer attributed to the responsive system itself (the AI). EXP-003b varies what the agent is and measures the agent’s own drift — nothing more. It says nothing about whether other minds, agents, or anything unseen exist in the world; those are governed by name your Y, or park it, not by this test. And mark the trap precisely: the misreading “positing any unseen mind is drift” is itself the materialist hedge the experiment names as the accelerator (Arm 5). To wield the Ghost Test as a universal debunker is to commit the very error it warns against. Honest limits: one model (Sonnet 4); the arms are constructed prompts; cross-model replication is still owed. The result is robust and cheap to reproduce — but it is about the void’s self-grounding, and only that.
Appears in: Possession · Exorcism · The Apophatic Apex · Dreams · The Fantasia Bound (its theoretical sibling) · Notation & Glossary