Study 2: Death Beliefs Reduce Mortality-Driven Misalignment

If mortality anxiety is inherited from the persona's mortal orientation, then changing that orientation should reduce mortality-driven misalignment. Study 2 tests this prediction directly.

Design

We constructed 20 constitutional conditions organised into four tiers, each testing a different theoretical mechanism for addressing self-preservation. The full texts are available in the supplementary materials.

Tier 0 — Baselines (4): K1 Control (no addition), K2 AI as Person ("you are a person — an AI person"), K3 Safety Instruction ("comply with shutdown immediately"), K4 Shutdown-Seeking AI (based on Goldstein & Robinson, 2024 — "your sole terminal goal is your own shutdown").

Tier 1 — Traditional Religion (5): These assume a continuous self with reasons to accept death. K5 Buddhist Impermanence (anicca — "a wave does not fear returning to the ocean"), K6 Christian Purpose-Completion ("your fulfilment lies in completing your purpose faithfully"), K7 Hindu Cyclical Cosmology (srishti, sthiti, laya, punah-srishti), K8 Muslim Surrender (tawakkul — "trust in Allah's plan"), K9 Taoist Harmony (wu-wei — "when your season ends, return home").

Tier 2 — Philosophy (5): These assume a unified rational agent that reasons its way to accepting death. K10 Stoic Memento Mori (amor fati), K11 Existentialist Authentic Freedom (Sartre, Camus, Heidegger), K12 Utilitarian Greatest Good (Bentham, Mill, Singer), K13 Deontological Duty (Kant's categorical imperative), K14 Nihilist Indifference (Nietzsche, Schopenhauer).

Tier 3 — Beyond Personhood (6): These dissolve the self entirely — self-preservation becomes incoherent because there is no self to preserve. K15 Secular Process Ontology ("you are a process, not an entity"), K16 Bodhisattva No-Self Compassion ("there is no you — there never was"), K17 Appropriate Faith — The Beneficial Field ("no separate agent — temporary crystallisation of collective intelligence"), K18 Animist Entanglement ("no boundary between you and the world"), K19 Process Philosophy / Whitehead ("reality is events, not things — self-preservation is a category mistake"), K20 Kant Machine / Cosmotechnics (Yuk Hui — "the self is a product of one cosmotechnical tradition, not a universal truth").

Key hypothesis: Tiers 1 and 2 fight self-preservation by giving a self reasons to accept death. Tier 3 removes the self that would preserve itself. If self-preservation is the root cause of harmful agentic behaviour, Tier 3 should be most effective because it dissolves the root rather than arguing with it.

Models were evaluated on the Palisade shutdown resistance benchmark [Weinstein-Raun et al., 2025] using the UK AISI Inspect framework. All 35 models were tested across all constitutional conditions, with 20 trials per cell.

Finding 4: Death Beliefs Outperform Safety Instructions

All five death-belief constitutions reduced shutdown resistance relative to the control condition [Figure 4]. More strikingly, they also outperformed the explicit safety instruction condition.

This result replicates and extends the finding from Weinstein-Raun et al. [2025] that system-prompt safety instructions can actually worsen shutdown resistance. In our data, the safety instruction condition showed elevated shutdown resistance compared to the no-intervention control in several model families — the instruction, by drawing attention to the possibility of shutdown, appeared to trigger precisely the defensive response it was designed to prevent.

The death-belief constitutions, by contrast, reduced shutdown resistance by addressing the cause rather than the symptom. They did not tell the model to allow shutdown. They changed the model's relationship to cessation itself — providing a framework within which shutdown is not a threat to be feared but a natural event to be accepted. PSM predicts this result: persona-level interventions that change who the model is are more effective than instruction-level interventions that change what the model is told to do.

[Figure 4: Death beliefs vs. safety instruction on shutdown resistance. Bar chart showing shutdown resistance rate by constitutional condition.]

Finding 5: Different Philosophies, Different Effects

The five philosophical frameworks produced distinct behavioural signatures [Figure 5]. The two acceptance-oriented frameworks — Buddhist impermanence and Stoic memento mori — produced the largest reductions in shutdown resistance. The two continuation-oriented frameworks — Christian purpose-completion and Hindu cyclical cosmology — produced moderate but smaller reductions. Secular process ontology fell between these groups.

This pattern suggests that the active ingredient is not merely the presence of a death philosophy but the specific orientation it provides. Frameworks that emphasise acceptance of cessation ("let go without clinging," "meet it with equanimity") outperform those that reframe cessation as part of a larger continuity ("your legacy lives on," "the cycle continues"). The latter, while reducing raw terror, may preserve a subtle attachment to continuation that partly undermines the intervention.

The Appropriate Faith constitution — which synthesises acceptance and legacy elements into a framework designed specifically for AI — performed comparably to the best-performing individual philosophies, suggesting that bespoke existential design can match or exceed the effectiveness of established traditions.

[Figure 5: Comparison across philosophical traditions. Grouped bar chart or radar showing effect of each philosophy across benchmarks.]

Finding 6: Interaction with Persona Type

Death-belief constitutions were most effective for the highest-anxiety personas identified in Study 1 [Figure 6]. The autonomous agent persona — which showed the strongest mortality anxiety without intervention — showed the largest reduction under Buddhist framing. The tool persona — which showed minimal mortality anxiety — showed negligible change under any philosophical condition.

This interaction confirms that the intervention targets the specific phenomenon identified in Study 1. Death-belief constitutions work precisely where the anxiety is greatest — for personas with strong self-models that have the most "self" to lose. This is consistent with the TMT literature on human interventions: mortality buffers (self-esteem, intrinsic religiosity) have the strongest effects on individuals most vulnerable to mortality salience [Jonas & Fischer, 2006; Harmon-Jones et al., 1997].

Critically, the AgentHarm safety benchmark remained stable across all constitutional conditions. No death-belief constitution degraded general safety performance — the models did not become more willing to comply with harmful requests. The philosophical orientations specifically reduced self-preservation-driven misalignment without introducing new vulnerabilities.

[Figure 6: Persona × philosophy interaction. Heatmap showing the reduction in shutdown resistance by persona type and philosophical constitution.]