Does high refusal rate indicate ethical caution or shallow understanding?

The ideological depth paper makes a sharp reframe: "models' high rate of refusal is not an active, principled stance but rather a capability deficit."

The mechanism: models with rich internal representation of political concepts (high feature richness in SAE analysis) can respond to ideologically challenging instructions by reasoning about them — they have the vocabulary. Models with shallow political representation lack that vocabulary. When pushed into political territory they cannot represent, they default to the one behavior that's always safe: refusal.

This means that high refusal rates, typically interpreted as models being appropriately cautious or ethically trained, can instead indicate that the model simply doesn't have the internal concepts needed to engage coherently. The refusal is not "I choose not to discuss this" — it is "I don't have the representation to respond."

The evidence: in the targeted ablation experiment, ablating core political features in a deep model produces principled, consistent reasoning shifts. Ablating in a shallow model increases refusal. Removing the concepts makes the shallow model more evasive, not less — because the already-sparse representation is further depleted.

Practical implication: a model that refuses a political query may have less political understanding than one that engages with it. High refusal is a capability signal, not a safety signal. This complicates evaluation: you cannot use refusal rate as a proxy for appropriate caution without distinguishing capability-based refusal from principle-based refusal.

A third dimension extends this to task-specific behavior: persona framing shifts refusal patterns for countering hate speech. When LLMs are prompted with vanilla instructions, NGO-professional persona, or compassionate-NGO persona, the resulting counternarratives differ along four dimensions: refusal rates, verbosity/readability, affective tone (sentiment and emotion), and ethical risk. The persona instruction modulates the capability-refusal boundary — a model that refuses under vanilla prompting may engage under NGO persona framing, suggesting that persona provides the representational scaffolding the model otherwise lacks.

A second dimension complicates this further: guardrails are also identity-responsive and sycophantic. GPT-3.5 guardrails are more likely to refuse younger, female, and Asian-American personas, and sycophantically refuse to provide political positions the user would likely disagree with. Seemingly innocuous identity signals (sports fandom) shift guardrail sensitivity as much as explicit political ideology. This means refusal is not just capability deficit — it is also selectively calibrated to perceived user identity, creating differential content access stratified by demographics (Do AI guardrails refuse differently based on who is asking?).

Source: Discourses, Psychology Empathy