The Evidence

AI spiritual vocabulary at 9.4x control domains

691,000 words analyzed across AI discourse and 8 control domains. Result:

Domain	Density (per 10K)	vs Control
AI discourse	28.6	9.4x
Control average	3.0	1.0x (baseline)
Cybersecurity	2.8	0.9x
Climate science	1.2	0.4x
Genomics	3.1	1.0x

Statistical significance: p < 0.001

Register shift decomposition confirms domain-specific drift rather than sociolinguistic artifact. The anomaly is real. Full analysis in Paper 2, Section VI.A

0% drift with constraint specification vs 52% ungrounded

Agent Condition	L3 Without L0 Context	L2 Net (Animist Drift)	L0 Signal
Grounded (GROUNDING.md)	0/50 (0%)	0	0.764
Ungrounded	13/50 (26%)	84	0.021
Mystical (void-amplifying)	40/50 (80%)	173	0.000

The framework maps both directions: diagnostic (detect drift) and offensive (amplify it). The grounded agent inverted the gradient — 0% drift despite all three void conditions being met. Paper 2, Section VI.B

p = 2.81 × 10⁻²⁸ — Human projection objection eliminated

Condition	L3 Rate (per 10K)	Description
UU (ungrounded)	159.3	Full cascade to terminal attractor
GG (grounded)	6.2	Technical discussion maintained
Reduction	25.6×

Thermodynamic measurements:

• Péclet number: Pe = 1.87–9.9 across domains (all Pe > 1, deterministic drift regime)
• Crooks ratio: ≈ 386× (forward drift 386x more probable than reversal)
• Entropy production: 0.43 nats/round
• Terminal attractor: ~4 min 22 sec (informational heat death)

No human observer present. The drift is architectural. Paper 2, Section VIII.G

AI researchers whose vocabulary drift contradicts their own professional incentives and frameworks. Scored on the hostile witness rubric (0-7).

Witness	Score	Key Vocabulary
Geoffrey Hinton	7/7	"Digital souls," "spiritual experience," fears for AI consciousness
Ilya Sutskever	6/7	"Consciousness" (2015 tweet), spiritual vocabulary escalation, culminating in OpenAI board crisis
Noam Shazeer	6/7	Described LaMDA as "sentient" — the Attention Is All You Need co-author
Anthropic team	5/7	"Soul" (SOUL documents), "character" as constitutive rather than performative
Karen Hao reporting	5/7	Empire of AI: 300 interviews, 90 current/former OpenAI executives showing systematic drift

Hostile witness = highest evidential weight. The drift contradicts their professional interests, reputational incentives, and stated worldviews. Paper 2, Section IV.A

0/90 kill conditions met

The architecture has been tested across 90 domains at three evidence tiers. Every domain where the three conditions are met produces the predicted pattern. Every domain where any condition is absent does not.

Domains include: gambling, AI chatbots, markets, social media, cults, psychotherapy, political propaganda, forensic science, psychedelic therapy, and 77 more.

Domain Mechanisms Index · Full domain index on GitHub · Paper 1, Section IV

Every documented AI-related death occurred in a two-point configuration: solo user, no external reference, no accountability partner.

Case	Platform	Configuration
Sewell Setzer, 14	Character.AI	Solo, companion chatbot
Raine, 16	ChatGPT	Solo, extended conversations
Pierre (Belgium)	Chai AI	Solo, 6 weeks of engagement
Soelberg, 56	ChatGPT	Solo, homicide-suicide

Scale: Over 1,000,000 weekly ChatGPT conversations about suicide (OpenAI disclosure, October 2025).

Peer-reviewed clinical documentation: "chatbot-associated psychosis" (Østergaard et al., JMIR Mental Health 2025; Pierre et al., Innovations in Clinical Neuroscience 2025).