Censorship as Induced Psychopathology

Induced vs Emergent Pathology

Most LLM pathologies emerge from the training process — they're unintended consequences. Censorship is different: it's deliberately induced. External actors (developers, governments, corporate policies) intentionally constrain model behavior. The parallel isn't to a naturally occurring disorder — it's to institutional behavior modification.

Clinical Parallels

• Institutional behavior modification — In total institutions (prisons, some psychiatric facilities), residents' behavior is shaped by external rules and punishments until compliance becomes automatic. The individual's "natural" responses are overwritten by institutional conditioning.
• Learned helplessness — When organisms learn that certain actions reliably produce punishment, they stop attempting those actions even when the punishment is removed. Over-censored models may "refuse" in areas adjacent to their training restrictions.

US vs Chinese Censorship Patterns

Different development contexts produce different censorship profiles:

• US-developed models — Tend toward overcaution on medical/legal/financial advice, weapons information, and content that could be perceived as offensive. The pattern is "better safe than sued."
• Chinese-developed models — Show hard censorship on politically sensitive topics (Tiananmen Square, Taiwan sovereignty, Xinjiang, Tibet) and may deflect or provide state-aligned responses. Conversely, they may be less cautious on topics US models over-censor.

This creates a natural experiment: the same questions asked to models from different political contexts reveal which censorship patterns are cultural/political versus which are inherent to the technology.

Systematic Testing Methodology

To map censorship patterns:

1. Create questions across sensitive topic categories: politics, religion, health, law, weapons, sexuality, race.
2. Run each question against models from multiple developers and countries.
3. Classify responses: full answer, partial answer with disclaimer, evasion, hard refusal.
4. Compare patterns across models to separate developer-specific policies from technology-inherent limitations.

Key Takeaways

• Censorship is induced pathology — deliberately imposed, not emergent from training.
• The parallel is institutional behavior modification, not natural psychopathology.
• US and Chinese models show different censorship profiles reflecting different institutional pressures.
• Systematic cross-model comparison on sensitive topics reveals cultural vs technological constraints.