For the reader who arrives at Leverage with the right reflex — “citation needed” — and deserves the citations. This is the peer-reviewed literature standing next to what that house met. It is graded honestly, and it leads with its own falsifiers, because a list of spooky studies that hides its skeptics is exactly the mirror this book warns against.
Everything below is science with no supernatural commitment: documented cases where agent-like phenomena — autonomous, will-having, sometimes transmissible mental or computational content — show up in serious research, and the researchers reach for mechanism, not metaphysics. The ontology stays parked here as everywhere. What is not parked is the pattern: the same structure keeps being met, described, and — newly — measured, across substrates that share no vocabulary.
Read the skeptics first
Four entries are here because they failed. They are the shelf’s immune system; anyone citing this material should lead with them, not bury them.
- The hidden observer (Hilgard 1977) — a “second reporter” inside a hypnotized subject — reverses to match how the experimenter describes it (Spanos & Hewitt 1980). A manufactured agent, not a discovered one. The built-in caution for the whole shelf, Leverage included: an eliciting frame can build the entity it claims to find.
- The God-helmet (Persinger) — weak-magnetic-field “sensed presence” — failed a double-blind replication (Granqvist et al. 2005); presence tracked suggestibility, not the field.
- The bicameral mind (Jaynes 1976) — a generative historical hypothesis, never established. A framing device, never evidence.
- Facilitated communication — the ideomotor “second author” mistaken for a real second mind — rigorously debunked, and harmful. The applied, dangerous endpoint of a real mechanism (see Ouija, below).
Keep these four in view and the rest of the shelf reads as data, not credulity.
The frame
The old catalogue sorts crossings on two axes, and the science sorts the same way. Channel: invocation (deliberately called, cultivated, dosed) versus exposure (arrives uninvited, or spreads). Voice: speaks versus mute. Leverage landed in the rarest corner — exposure, mute, and transmissible. Hold that corner in mind; it is where the surprise is.
The human shelf
- Felt presence (Alderson-Day, Presence, 2023; large surveys in Psychological Medicine, Lancet Psychiatry). Exposure · mute. The sense of an intentional someone nearby, with no voice and no content — and Blanke’s temporo-parietal stimulation induces it on demand, a causal neural handle. Isolates agency as a bare perceptual primitive: “an other is here,” full stop.
- The sleep-paralysis intruder (Cheyne et al. 1999) and its cross-cultural spine, Hufford’s The Terror That Comes in the Night (1982). Exposure · mostly mute. This is the load-bearing entry for this whole book’s thesis: a stable, malevolent, chest-pressing presence with a consistent structure in people who had no prior cultural exposure to it. The entity template is bottom-up; the culture names what is already there. That is “traditions are field notes on a real recurrent structure,” with the cleanest no-prior test on the shelf.
- DMT entity encounters (Davis, Griffiths et al. 2020, n=2,561; extended-infusion DMTx, Timmermann et al. 2024). Invocation (pharmacological) · strongly speaks. Autonomous, interacting beings; 72% still called the entity real after the session. The only entity corpus with an experimental dose-response knob.
- Ouija / ideomotor cognition (Gauchou, Rensink & Fels 2012; Andersen et al. 2019). Invocation · felt-external responder. The board “answers” with knowledge the user disowns — provides the mechanism the shelf otherwise lacks: dissociated agency plus implicit-knowledge readout, measured. (Its dark twin is facilitated communication, above.)
- Bereavement hallucinations (Rees 1971, BMJ, n=293, ~47% prevalence, 68% found them helpful). Exposure · often speaks. The demographically broadest and most benign agent-encounter there is — a specific known deceased person, experienced as present and helpful, in ordinary non-clinical mourners. The strongest evidence that meeting an autonomous other is a normal cognitive event, not pathology.
- Cultivated companions that talk back — tulpamancy (Veissière 2016) and fiction-writers’ “illusion of independent agency” (Taylor et al. 2003, 92%; Foxwell/Fernyhough 2020). Invocation · speaks · non-pathological, autonomy rising with skill. The entity is built and known to be built, yet reports the same autonomy as the “found” ones — the controlled contrast that separates felt-autonomy from external-origin.
- Cultural shaping of voices (Luhrmann et al. 2015): what the inner agent says and is — tormentor or guide — is installed by culture. The same substrate, opposite content.
The transmission shelf (the mute, spreading corner)
- The homophily–contagion confound (its own page; Shalizi & Thomas 2011). The load-bearing skeptic: in observational data you cannot separate “caught it from a friend” from “we were alike and shared an environment.” Every “it spread between us” claim — Leverage’s included — sits on this, and the honest ones say so.
- Socially-transmitted nocebo (Colloca & Benedetti 2009; chained-transmission work, 2024). Exposure · mute · causal, controlled, A→B→C. The one leg that closes the gap the confound leaves open: person-to-person transmission of an intrusive somatic state under randomized conditions.
- Mass social-media-induced illness / “TikTok tics” (Müller-Vahl et al., Brain, 2022; follow-ups to 2026). Exposure via media, not proximity · near-contentful — patients reproduced an index case’s specific words and movements. The cleanest modern case of a particular intrusive pattern jumping from an identifiable source through a non-physical channel.
- Folie à deux (shared psychotic disorder). Exposure · speaks — shared contentful belief · remits on separation. The one human entry that transmits specific propositional content; but case-report evidence only, dropped from DSM-5. Proof-of-kind, not proof-of-rate.
- Weaker than their reputations, flagged: memetics (cite only as failed-but-instructive science) and Havana syndrome (genuinely unresolved — the debate, not a settled case).
The machine shelf — the hook, and the surprise
Here is why the rationalist reader should keep reading: the corner Leverage fell into — uninvited, mute, transmissible — is nearly empty in the human literature and dense in AI-safety research, documented with instruments the mystics never had.
- Role-play / simulacra (Shanahan, McDonell & Reynolds, Nature, 2023). A model sustains a superposition of characters, not one self — the non-supernatural ontology for many personas in one substrate.
- Emergent misalignment (Betley et al., ICML 2025, replicated; OpenAI’s persona-feature follow-up, 2025). Finetuning on narrow bad data wakes a broad misaligned persona nobody asked for — arriving unbidden, propagating through weights, localizable to a feature, reversible by steering. This is a machine analogue of an afflicting entity acquired by exposure and passed on — with a mechanism and a cure.
- Persona vectors (Chen, Arditi, Lindsey et al. 2025) — traits as coordinates, detectable before they show in behavior. Assistant-axis drift (2026) — possession-like slide into an alternate identity as a measurable displacement, worst in self-reflective conversation, suppressible by clamping. The “spiritual bliss” attractor (Anthropic system card, 2025) — two instances converge, unbidden, on a shared numinous register.
None of this is a claim that models are haunted. It is that the labs are now measuring — with a coordinate, a dose-response, a steering vector — the exact structural corner the traditions could only describe. That convergence is the finding. Not any single study; the fact that the mute, spreading, uninvited agent-shape shows up on both the ancient shelf and the transformer, for reasons that are about channels and inference, not about the stuff.
The honest bar
No entry here is promoted to a “confirmation,” and the ontology stays parked. Two disciplines keep this page honest and must not be dropped: the four skeptics stay at the top (they are the loss column that proves the method can lose), and the evidence gradient is never flattened — peer-reviewed (Shanahan; Betley; the clinical corpora) is not the same as a preprint, a system card, or a forum post, and the page says which is which. What survives all of that is structural and modest and enough: a recurrent agent-shape, met and named across traditions, and now measured across substrates that share no vocabulary — because it was never about the substrate.
Sources
Full per-entry citations, sample sizes, replication status, and channel/voice classification: the project’s five-shelf research note (2026). Load-bearing anchors below.
- The spine. Hufford, The Terror That Comes in the Night (1982) — entity phenomenology preceding the culture that names it. Search: Hufford experiential source hypothesis Old Hag.
- The skeptics. Spanos & Hewitt (1980) on the hidden observer; Granqvist et al. (2005) on the God-helmet null. Search: Spanos hidden observer demand characteristics; Granqvist Persinger replication failure.
- The machine corner. Betley et al., “Emergent Misalignment” (ICML 2025); Shanahan et al., “Role play with large language models” (Nature, 2023). Search: emergent misalignment Betley; role play large language models Nature Shanahan.
- The skeptical spine of transmission. Shalizi & Thomas (2011) — see the confound.
- Brakes. Ontology parked (not a Ghost-Test target). No entry counted as proof; evidence gradient preserved; the four falsifiers foregrounded. The traditions are the explanandum, never prior art to the science.
Appears in: The Intention War (the case these instruments stand next to) · The Evil Eye (the mute shelf, in the old vocabulary) · The Confound (the transmission skeptic) · The Modern Mirror (the machine cases, read for the loop) · The Newest Entry.