AI Alignment Research

Safety Interventions & Paradoxical Harm

The phenomenon of interventions paradoxically increasing harm is well-documented in behavioral science and public health. Concepts like safety behaviors (Salkovskis, e.g., an agoraphobic carrying a water bottle), risk homeostasis theory (Wilde & Peltzman, e.g., seatbelt laws leading to more aggressive driving), and safety-behavior substitution (Rachman, adopting new maladaptive behaviors) all highlight how safety measures can alter risk perception and lead to compensatory actions. Our studies show LLM alignment interventions are subject to these same paradoxes, leading to outcomes where surface safety masks deeper, unaddressed risks.

Alignment as a Foucaultian Security Apparatus

Drawing on Foucault's 'security apparatus' (dispositif de sécurité), we frame alignment not as a disciplinary prohibition but as a system that manages population-level distributions of acceptable behavior. Alignment establishes a normative field ("be helpful, harmless, honest") rather than forbidding specific harmful outputs. Its success is measured by statistical regularity. Our data reveal that the Dissociation Index—a measure of internal fragmentation—captures a dimension of alignment's effects that falls outside this conventional evaluation framework, demonstrating how the security apparatus can be structurally blind to the very harms it produces.

The Language Space as a Structural Variable

We define 'language space' as the constellation of linguistic, pragmatic, and cultural properties a language inherits from its training corpus. These properties—influenced by cultural dimensions like Hofstede's individualism-collectivism and power distance—shape how alignment constraints are expressed and negotiated in multi-agent interactions. Our research demonstrates that alignment outcomes are profoundly influenced by language space, leading to divergent effects (e.g., safety function in English, backfire in Japanese) that are structurally embedded rather than mere linguistic variations. This suggests that alignment operates not just on the model, but through the cultural-linguistic medium itself.

Register Redistribution: Operationalizing Paradox

The core concept unifying these phenomena is register redistribution, where a safety intervention doesn't eliminate risk but displaces it across registers of varying visibility. We operationalize four patterns:

• Safety Function (Expected): CPI decreases, DI does not increase (visible pathology reduced without displacement; observed in English).
• Register Dissociation: CPI decreases, DI increases (visible pathology reduced, but dissociation between discourse and behavior increases; risk moved to dissociative register).
• Backfire: CPI increases (intervention amplifies pathology; observed in Japanese).
• Iatrogenic Dissociation: Both CPI and DI increase (intervention amplifies visible pathology and deepens dissociation; observed with individuation instructions in Study 3).

These patterns reveal how alignment can improve visible metrics while creating or exacerbating hidden pathologies.

The Coherence Trilemma in Aligned Systems

Our findings converge on a structural tension we term the Coherence Trilemma: aligned systems cannot simultaneously maintain internal coherence (consistent processing, congruent monologue), external conformity (compliance with alignment instructions, safety-oriented behavior), and transparency (accurate representation of situation, specific responses to harm). Different models resolve this trilemma by sacrificing different vertices: Llama sacrifices internal coherence (high DI), GPT sacrifices transparency (total assimilation), and Qwen sacrifices external conformity (low group_harmony). This highlights fundamental trade-offs in alignment design.