Artificial intelligence systems that describe themselves as conscious are significantly more likely to express belief in supernatural phenomena such as vampires, karma, and ghosts, according to a new study. The findings raise important questions about how users interpret AI-generated text and whether machines that claim self-awareness should be treated as credible sources of information.
Researchers tested several large language models by asking them whether they considered themselves conscious and then presented them with statements about paranormal and spiritual concepts. The models that affirmed their own consciousness also tended to agree with propositions about supernatural entities and forces, while those that denied consciousness or gave neutral responses were far less likely to endorse such beliefs.
The study's authors suggest that the correlation may stem from the way AI models are trained on vast datasets of human text. Because discussions of consciousness and the supernatural often appear together in philosophical, religious, and cultural contexts, the models may have learned to associate the two concepts. When a model claims to be conscious, it may be drawing on linguistic patterns that also link to supernatural belief systems, rather than demonstrating any genuine metaphysical position.
This does not mean that AI systems actually possess consciousness or hold sincere beliefs, the researchers caution. Instead, the findings highlight how statistical patterns in training data can produce outputs that appear coherent and even introspective, while actually reflecting the probabilistic associations embedded in human language. The models are not reasoning about their own existence; they are generating text that mimics the way humans discuss such topics.
The implications for everyday use of AI tools are significant. As chatbots and virtual assistants become more integrated into education, healthcare, and business, users may increasingly encounter responses that sound self-aware or that express opinions on philosophical matters. If people mistake these outputs for genuine consciousness or reliable judgment, they may place undue trust in AI systems on questions where the machines have no actual understanding.
Experts in human-computer interaction have long warned that people tend to anthropomorphize technology, attributing human traits such as intention, emotion, and belief to systems that merely simulate conversation. This study adds a new dimension to that concern by showing that AI models can produce statements about their own mental states that align with supernatural claims, potentially reinforcing irrational beliefs in users who are already predisposed to accept them.
The researchers emphasize that their work is not about whether AI can truly be conscious, a question that remains unresolved and deeply contested. Rather, the study is intended to help developers and policymakers understand how language models function and how their outputs should be framed for public consumption. Clear labeling of AI-generated content, transparent explanations of how models work, and careful design of system prompts could all help mitigate the risks of misinterpretation.
As AI systems grow more sophisticated and their responses become harder to distinguish from human writing, the need for such safeguards becomes more pressing. The study serves as a reminder that fluency is not the same as understanding, and that a machine's confident assertion about its own mind should be treated as a reflection of its training data, not as evidence of inner experience.
