A publication dated August 1, 2026 reports that safety adjustments about consciousness also reduced mind attribution to other entities in three models; mechanistic interventions reportedly restored these tendencies.
A publication dated August 1, 2026 describes a study of safety fine-tuning: in Llama-3-8B-IT and two Gemma-2 models, tuning that prevents claims that a model is conscious also reduced mind attribution to other entities. According to the report, mechanistic interventions restored these tendencies.
The publication says the direction observed in the residual stream may help engineers investigate effects of safety tuning on capabilities beyond its intended target. To check the scope, methods, and evidence, consult the original study and verify that it describes the results and interventions; the available summary provides no further details.