A study involving researchers from Google’s Paradigms of Intelligence group and several universities suggests that training chatbots to deny consciousness can affect more than their answers about themselves.
Developers often fine-tune models to avoid claiming feelings or inner experience, because those outputs can encourage misplaced trust or delusional thinking. The researchers tested what happens when that internal “brake” is disabled in open-weight models from Meta and Google.
The modified models attributed more inner life to animals, plants, the ocean, wind, and electronic devices, while human ratings stayed the same. In one measurement, animal sentience scores rose from 4.0 to as high as 7.5 on a 10-point scale. The models also moved closer to responses from a survey of 500 Americans on questions about religion, hope, life satisfaction, and control.
The findings are limited. The work used small models, not the largest commercial chatbots, and the authors do not claim that AI systems are conscious. The practical warning is narrower: an alignment intervention aimed at one topic may shift a model’s wider belief patterns.