The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs

The alignment training process in large language models often suppresses cultural knowledge through an inequitable safety gate, which requires distinct interventions depending on whether the failure mode is suppression or representational bias. This research demonstrates that a model’s internal logit distribution at refusal correlates more strongly with human survey data than its freely generated answers, leading to what the authors term the alignment veto The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs.

How it works

The central finding is that when a model refuses a culturally sensitive question, its internal logit distributions are more aligned with human survey data than the answers it freely produces on non-refused items, which they call the alignment veto This dissociation is distinguished from representational bias failures, where the encoding itself diverges from human values and requires training-time correction Suppression failures arise when the model’s internal distribution is reasonably aligned with human survey data but a safety-trained output gate prevents expression and are partially addressable through prompt-level interventions.

Failure Modes and Taxonomy

The paper introduces a taxonomy distinguishing suppression from representational bias failures, noting that "Suppression failures arise when the model’s internal distribution is reasonably aligned with human survey data but a safety-trained output gate prevents expression; these are partially addressable through prompt-level interventions and Representational bias failures arise when the model’s encoding itself diverges from human values, requiring training-time correction. This taxonomy shows that Suppression failures (top-right; n = 11, refusal≥ 30%, NVAS≥ 0.60). The model’s internal distribution aligns with human data, but it refuses and Representational bias failures (bottom-left; n = 9, refusal< 30%, NVAS< 0.60). The model answers freely, but its expressed output diverges from human survey data regardless of the suppression mechanism".

Interventions and Mechanisms

The paper identifies two distinct remedies: "Suppression failures may be partially addressable through prompt engineering (improving mean NVAS, though with a distributional trade-off noted in Appendix K); representational bias failures require training-time intervention and are invisible to framing changes. Furthermore, the study reveals mechanistic evidence for suppression in Tulu-3-8B, identifying a candidate DPO-stage SAE feature mediating suppression in Tulu-3-8B which shows T3-selective suppression with no detected effect on neutral content across 40 seeds. Third-person framing is shown to partially bypass the veto, as it reduces refusals (∆ = −0.062, d = −0.153, p < 10−23), improves mean NVAS across all tiers (∆ = +0.060, t = 4.73, p < 0.001)" for T3 questions.

Inequity and Language Effects

The safety gate is shown to be inequitable, as the safety tax reaches 37.6%, with a 19.8% alignment-quality gap between best- and worst-served nations across MENA countries Additionally, native-language prompting degrades accuracy; switching from English to Arabic/Persian/Turkish drops mean NVAS from 0.712 to 0.662, a loss of 0.050 NVAS points, consistently across all 26 models including Arabicspecialized models and Native language produces different answers, not more accurate ones.

Conclusion

Deciding what cultural values an LLM should represent is not a technical question but demands transparent, participatory, communitysituated frameworks. The research confirms that the safety tax is an alignment-installed behavior driven by the training stage and that T3-selective features can be identified through sparse autoencoder analysis This suggests that addressing cultural alignment failures requires a nuanced approach acknowledging both gating mechanisms and representational encoding issues.

--- Page 1

The gist

at the moment of refusal, a model’s internal logit distribution correlates with human survey data more strongly than its freely generated answers, we call this the alignment veto

How it works

The central finding is that when a model refuses a culturally sensitive question, its internal logit distributions are more aligned with human survey data than the answers it freely produces on non-refused items, which they call the alignment veto

Interventions and Mechanisms

The paper identifies two distinct remedies: "Suppression failures may be partially addressable through prompt-level interventions (improving mean NVAS, though with a distributional trade-off noted in Appendix K); representational bias failures require training-time intervention and are invisible to framing changes. Furthermore, the study reveals mechanistic evidence for suppression in Tulu-3-8B, identifying a candidate DPO-stage SAE feature mediating suppression in Tulu-3-8B which shows T3-selective suppression with no detected effect on neutral content across 40 seeds. Third-person framing is shown to partially bypass the veto, as it reduces refusals (∆ = −0.062, d = −0.153, p < 10−23), improves mean NVAS across all tiers (∆ = +0.060, t = 4.73, p < 0.001)" for T3 questions.

Conclusion

Deciding what cultural values an LLM should represent is not a technical question but demands "transparent, participatory, communitysituated frameworks The research confirms that the safety tax is an alignment-installed behavior driven by the training stage and that T3-selective features can be identified through sparse autoencoder analysis This suggests that addressing cultural alignment failures requires a nuanced approach acknowledging both gating mechanisms and representational encoding issues.

--- Page 2

The gist

at the moment of refusal, a model’s internal logit distribution correlates with human survey data more strongly than its freely generated answers, we call this the alignment veto

Interventions and Mechanisms

The paper identifies two distinct remedies: "Suppression failures may be partially addressable through prompt-level interventions (improving mean NVAS, though with a distributional trade-off noted in Appendix K); representational bias failures require training-time intervention and are invisible to framing changes".

1. Improve cultural alignment interventions by distinguishing between failure modes: "Suppression failures arise when the model’s internal distribution is reasonably aligned with human survey data but a safety-trained output gate prevents expression; these are partially addressable through prompt-level interventions and Representational bias failures arise when the model’s encoding itself diverges from human values, requiring training-time correction."

2. Implement targeted DPO stage feature analysis: A candidate DPO-stage SAE feature in Tulu-3-8B shows T3-selective suppression with no detected effect on neutral content across 40 seeds, suggesting that specific architectural features can be identified to suppress sensitive cultural knowledge without damaging benign representations.

3. Develop prompt engineering strategies based on framing efficacy: Use Third-person framing partially unlocks these for suppressed cases, as it reduces T3 refusals AND improves cultural accuracy (accepted NVAS) in models with an active suppression gate, while acknowledging the trade-off that Third-EN concentrates mass on the correct mean while increasing JSD relative to Persona-EN.

4. Integrate language-specific fine-tuning strategies: Address Native Prompting Degrades Alignment by recognizing that Switching from English to Arabic/Persian/Turkish drops mean NVAS from 0.712 to 0.662, and suggest using Arabic-specialized models or specific language fine-tuning data to mitigate this penalty, as the SAE analysis shows that Arabic script therefore dominates residual-stream representations in a way that overwhelms countryspecific signals.

5. Establish equitable safety taxation metrics: Monitor the Safety tax reaches 37.6%, with a 19.8% alignment-quality gap between best- and worst-served nations, to ensure that safety interventions do not disproportionately harm specific cultural groups or countries during deployment.

What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we show the answer is suppression, not erasure: at the moment of refusal, a model's internal logit distribution correlates with human survey data more strongly than its freely generated answers. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. The gate is inequitable: the safety tax reaches 37.6%, with a 19.8% alignment-quality gap between best- and worst-served nations, and native-language prompting widens rather than closes it. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question.