Harm Washing
A concept in AI bias governance referring to the phenomenon whereby discriminatory biases in large language models are not eliminated through safety alignment training but instead transition from explicit expression to more covert forms—such as tone, word choice, or implicit assumptions—thereby evading standardized bias detection mechanisms while still producing discriminatory effects in actual model outputs.
Timeline (last 90 days)
A study targeting GPT-series models introduced the concept of "Harm Laundering," referring to the phenomenon where gender bias in a model is not truly eliminated but is instead transformed into more covert, harder-to-detect forms.
A model can avoid words or sentence patterns flagged as harmful at the surface output level while retaining the original bias structure in the semantic associations encoded in its weights.
A model that scores "low bias" on evaluation leaderboards is not necessarily truly fair in actual deployment.
Stress Testing, which involves designing boundary scenarios, role switching, or multi-turn dialogues to induce models to expose suppressed tendencies, is a potential direction for identifying "Harm Laundering," though no standardized methodology has yet been established.
If bias is encoded in a model's embedding space in the form of vector associations such as "female–home, male–workplace," it may still seep into actual outputs through pathways such as contextual inference, even if the output layer has been trained not to express it directly.
The topic of "Harm Laundering" is still in an early stage of discussion, with limited attention and discourse in the technical community, and its empirical evidence, quantification methods, and generalizability remain to be validated by broader research.
All Facts (5)
A study targeting GPT-series models introduced the concept of "Harm Laundering," referring to the phenomenon where gender bias in a model is not truly eliminated but is instead transformed into more covert, harder-to-detect forms.
50%UnverifiedA model can avoid words or sentence patterns flagged as harmful at the surface output level while retaining the original bias structure in the semantic associations encoded in its weights.
50%UnverifiedA model that scores "low bias" on evaluation leaderboards is not necessarily truly fair in actual deployment.
50%UnverifiedStress Testing, which involves designing boundary scenarios, role switching, or multi-turn dialogues to induce models to expose suppressed tendencies, is a potential direction for identifying "Harm Laundering," though no standardized methodology has yet been established.
50%UnverifiedIf bias is encoded in a model's embedding space in the form of vector associations such as "female–home, male–workplace," it may still seep into actual outputs through pathways such as contextual inference, even if the output layer has been trained not to express it directly.
50%