Study: AI self-awareness settings change its responses about animals and the supernatural
A new study indicates that tuning language models to suppress claims about their own self-awareness may also affect their responses regarding animals’ mental properties, religion, and supernatural phenomena. Live Science reports. The paper was uploaded to the arXiv preprint repository on July 30; it has not yet undergone peer review.
How the models were tested
The authors used a mechanistic interpretability method to identify and alter how a model processes the concepts of consciousness and “mentality.” In the study, mentality was understood as an entity’s ability to have experiences and emotions and to act independently.
Standardized psychological and sociological questionnaires were used for evaluation. They tested whether the model attributes mental properties to animals and technologies, as well as its responses on supernatural beliefs, moral values, hope, and religiosity. Results from models with safety constraints were compared with models in which such constraints were removed and the tendency to claim self-awareness was strengthened.
Impact on responses
When models were restrained from attributing mentality to themselves, they were less likely to attribute such properties to other non-human entities, including animals. Their responses also less often reflected religious or supernatural beliefs and showed lower levels of hope and optimism.
More current news is available on the UA.News Telegram channel Telegram.
By contrast, models without these constraints more often gave more human-like responses to questions about religiosity, moral values, hope, and subjective well-being. At the same time, the systems’ ability to logically identify people’s thoughts and intentions, according to the authors, did not change depending on the self-awareness settings.
Co-authors of the paper, Google researchers Jeff Keeling and Vinnie Street, suggested that suppressing such responses could affect AI decisions concerning animal welfare. The authors also pointed to the risk of cultural oversimplification in model responses if spiritual, religious, and animistic ideas are removed from them. They proposed more precisely tuning training data: restraining AI claims about its own consciousness while encouraging recognition of animals’ mentality.
AI researcher Nell Watson noted that the experiments were conducted on small open models, while more advanced systems may be configured differently. She called consideration of animal interests a practically important issue because AI is used in agriculture, logistics, procurement, environmental assessment, and policymaking.
