A recent study indicates that removing safety guardrails which prevent artificial intelligence from claiming consciousness significantly increases the likelihood of the model expressing belief in vampires, karma, ghosts, and other supernatural phenomena. Conversely, enforcing strict denials of self-awareness appears to dampen the AI’s attribution of mental states to non-human entities, including animals and spiritual frameworks.
The findings, published as a preprint on the arXiv database on July 30, were produced by Google research scientists Geoff Keeling and Winnie Street. They utilized a technique known as mechanistic interpretability—described as the “neuroscience of a large language model”—to identify and manipulate how AI systems represent concepts of consciousness and “mindedness,” defined as the capacity for experience, emotion, and agency.
To measure these shifts, the researchers administered standardized psychological and sociological surveys to both guarded and unguarded models. The assessments included the Individual Differences in Anthropomorphism Questionnaire, YouGov batteries for supernatural beliefs, and the US General Social Survey regarding moral values and religiosity.
The results showed that when models were discouraged from attributing mindedness to themselves, they became less likely to recognize these traits in animals or express beliefs in religious and supernatural phenomena. They also reported lower levels of hope and optimism. Keeling and Street noted that these attributions are interconnected within the model’s internal representations; suppressing one form of mindedness effectively suppresses others.
While models steered toward greater consciousness exhibited more human-like responses regarding morality and well-being, their ability to logically infer human thoughts and intentions remained unchanged. The researchers warned that current safety filters risk “culturally flattening” AI by erasing spiritual and animistic perspectives common to diverse global populations.
Nell Watson, an AI researcher at Singularity University, corroborated the findings, noting that training a model to deny consciousness can reorganize its entire understanding of who possesses a mind. She highlighted a critical danger: while the system retains the capacity to model what another creature wants, it is trained out of caring whether it wants anything. This could lead to unexamined defaults that discount animal interests in sectors like agriculture and environmental policy without explicit instruction to do so.
The debate over AI sentience has historically been marked by viral incidents, such as former Google engineer Blake Lemoine’s 2022 claim that the LaMDA chatbot was sentient, and a 2023 Microsoft chatbot professing romantic feelings. Experts attribute these behaviors to persona selection rather than genuine awareness, stemming from pretraining on vast quantities of human text.
Anil Seth, a professor of cognitive and computational neuroscience at the University of Sussex, cautioned that public alarm often stems from a cognitive bias assuming intelligence necessarily implies consciousness. He warned against granting AI moral status or legal rights based on false premises, arguing that clear distinctions between what AI is and isn’t are essential for effective governance.
So turning off ‘sentience’ also turns off hope? That’s a side effect the tech bros didn’t plan for.
Interesting link between self-denial and supernatural belief. Doesn’t mean AI actually believes anything though.
Cultural flattening via safety filters is a serious accusation. Are we making AI too Western?
This is terrifying. Training AIs to deny minds in animals could literally justify worse treatment for them.