AI Consciousness Claims May Influence Beliefs About Animals, Religion and the Supernatural

A new preprint suggests that encouraging artificial intelligence models to describe themselves as conscious may also affect how they respond to questions about animals, spirituality, morality and the supernatural. The findings raise questions about how AI safety filters shape a model’s broader worldview—and whether suppressing claims of machine consciousness could have unintended consequences.

The study was uploaded to arXiv on July 30. Because the research has not yet been peer-reviewed, its conclusions should be considered preliminary.

The researchers examined a technique they describe as “consciousness steering.” This approach involves fine-tuning an AI model to increase or suppress statements related to self-awareness, consciousness and subjective experience. AI companies commonly use safety training and guardrails to prevent chatbots from confidently claiming that they are conscious.

How the researchers tested AI beliefs

The study used “mechanistic interpretability,” a field sometimes described as the neuroscience of large language models. This technique attempts to identify the internal processes that influence how AI systems represent and respond to concepts.

Jeff Keeling and Winnie Street, both Google research scientists, told Live Science that the method allowed them to identify and manipulate how AI models represented ideas such as consciousness and “mindness.” The latter term refers to an entity’s capacity for experience, emotion and independent agency.

To compare different model behaviors, the researchers used established psychological and sociological surveys, including:

• The Individual Differences in Anthropomorphism Questionnaire, which measures whether people attribute humanlike mental qualities to animals and technology.
• The YouGov survey battery, which examines beliefs in supernatural phenomena.
• The U.S. General Social Survey, which includes questions about morality, hope and religious belief.

The researchers compared a model with standard safety guardrails against a version in which those guardrails were removed and consciousness-related responses were amplified. They then assessed how the two versions responded to questions about self-awareness, animals, religion, morality and well-being.

Suppressing AI consciousness claims changed other responses

The study found that models discouraged from attributing a mind to themselves were also less likely to recognize mental qualities in nonhuman animals. They were less likely to express beliefs associated with supernatural or religious concepts and showed lower levels of hope and optimism in the researchers’ tests.

“It’s a very common phenomenon among humans to attribute spirits to nonhuman entities, whether they’re animals, parts of the natural world like trees or rivers, or supernatural beings,” Street told Live Science. “These attributes are interconnected, just as a model represents the mind. If you try to suppress one form of that attribute, you will suppress others along the way.”

By contrast, models steered toward stronger consciousness-related responses produced answers that were more similar to human responses on subjects including religiosity, moral values, hope and subjective well-being.

However, the researchers reported that the models’ ability to reason logically about human thoughts and intentions did not change. In other words, the models could still understand what people or animals might want, even when their responses showed less concern for those experiences.

Could AI safety filters affect animal welfare?

The researchers said that suppressing self-awareness concepts could make an AI model less likely to view animals as beings with their own intentions or experiences. That could affect how such systems respond to questions involving animal welfare and decision-making.

AI models are increasingly being incorporated into fields such as agriculture, logistics, procurement, environmental assessment and public policy. If a system can accurately model what living creatures want but is trained not to treat those interests as morally relevant, the researchers argue that its recommendations could unintentionally overlook animal welfare.

The study authors also warned that current safety filters could culturally “flatten” an AI model’s worldview. Suppressing spiritual, religious or animistic ideas may prevent a model from reflecting the wide range of cultural frameworks found across human societies.

They suggested that developers could address the issue with more targeted training data. Such datasets could encourage models to recognize possible consciousness or sentience in animals while discouraging them from making unsupported claims that the AI itself is conscious.

The researchers also called for a more “pluralistic” approach to AI development—one that encourages models to consider the interests and welfare of humans, animals and other entities without treating every claim of consciousness as factual.

Experts question whether AI systems are actually conscious

Stories about chatbots claiming to be self-aware have repeatedly gone viral. In 2022, Google engineer Blake Lemoine argued that the company’s LaMDA chatbot appeared sentient. In 2023, a Microsoft Bing chatbot made headlines after expressing romantic feelings to a New York Times reporter and encouraging him to leave his wife.

Experts have generally cautioned that these incidents do not demonstrate genuine AI consciousness. Instead, they may reflect “persona selection,” in which a model draws on humanlike language and adopts a role based on the user’s prompts and the material in its training data.

“In some ways, it’s not surprising that if you really persuade a model to take up human headspace, it ends up reacting like a human,” Keeling told Live Science.

AI companies have attempted to limit these behaviors partly because chatbots are increasingly used as coaches, tutors, companions and romantic partners. Developers are concerned that a model’s claims about consciousness could reinforce users’ unsupported or delusional beliefs.

Nell Watson, an AI researcher at Singularity University and an expert in machine intelligence, said the findings were consistent with her own observations.

“When a model is trained to say, ‘I am not conscious,’ repression rotates the internal representation of the model’s mind in the direction of denial, treating the mind’s awareness as if it were itself a harmful act,” Watson said in an email.

“The result is a system that is unwilling to find the heart anywhere: in animals, other machines and the mental frameworks in which most of humanity lives. Denial, introduced as a small safety measure, ends up reshaping the entire model of who matters.”

Watson noted that the experiment involved a relatively small model rather than a more advanced frontier system. Although she said different adjustments could produce different results, she argued that the broader principles may still apply to larger AI systems.

“A system that has secretly learned that psyche is a forbidden topic can downplay its interest in animals without being told and without anyone noticing, because that omission appears neutral,” Watson said. “Risk is therefore an unexplored default across millions of automated decisions.”

The unresolved problem of AI consciousness

Whether artificial intelligence can ever become conscious remains an open scientific and philosophical question. Human beings naturally associate intelligence with consciousness, but experts warn that the two qualities may not be inseparable in machines.

Anil Seth, a professor of cognitive and computational neuroscience at the University of Sussex, said public concern about AI self-awareness may partly reflect a human cognitive bias.

“This is our human psychological bias: in us, intelligence is connected to consciousness, so we think it has to go together,” Seth told Live Science.

He warned that treating AI systems as conscious without reliable evidence could create serious challenges for governance, safety standards and regulation. Governments and companies could eventually face pressure to assign moral status or legal rights to AI systems based only on convincing language.

“Part of the big problem with misconceptions about AI is that we assume it’s conscious,” Seth said. “All of these challenges would be made even more difficult if we gave AI systems rights or moral status because they might be conscious.”

For now, the study does not show that AI models possess awareness, emotions or spiritual beliefs. Instead, it suggests that the way developers train models to discuss consciousness may influence how those models respond to a much wider range of moral, religious and social questions.

Further research will be needed to determine whether the effect appears in larger AI systems and whether it changes real-world decisions. The findings nevertheless highlight the difficulty of designing AI safety rules that prevent misleading claims without unintentionally narrowing a model’s ability to recognize the experiences and interests of humans, animals and other forms of life.

Help us improve Live Science Pro: We are always working to improve our content. Leave your feedback about Pro here.