We like to think of AI as a neutral arbiter of truth—a digital librarian that simply hands us the facts. But a growing body of research suggests that Large Language Models (LLMs) have some pretty specific opinions, especially when it comes to religion. A recent study has highlighted a particular friction point: AI models seem to have a systematic bias against Jehovah’s Witnesses, often reflecting the internet's most critical and controversial perspectives rather than objective theology.
The Training Data Trap
The issue boils down to the "garbage in, garbage out" rule of machine learning. LLMs are trained on massive scrapes of the internet, which includes everything from scholarly articles to heated forum debates. When it comes to religious minorities like Jehovah’s Witnesses, the digital landscape is often polarized.
Much of the data available to these models comes from secular "cult-watch" organizations or mainstream theological critics. Because the AI absorbs these perspectives without the nuance of lived experience, it tends to parrot labels like "cult dynamics" or focus exclusively on insular worldviews. Instead of providing a balanced theological overview, the AI reflects the internet's loudest—and often most negative—voices found in archives like Cult News 101 or various critical magazines.
Beyond Kingdom Hall: A Broader Theological Bias
This isn't just a problem for one specific group; it’s a symptom of how AI handles minority faiths. Researchers are uncovering that LLMs often marginalize specific religions by default. Whether it’s summarizing the legal battles in Russia where the group faced bans or discussing internal doctrines like "wicked spirit forces," the AI struggles to separate objective reporting from sectarian bias.
When an AI is asked about the group, it might prioritize "insider" critiques or critical studies over the group's actual doctrines. For a technology increasingly used to summarize human knowledge, this creates a feedback loop where minority voices are systematically misrepresented or dismissed as fringe. Even when the AI tries to be neutral, the sheer weight of critical training data from sites like StackExchange or Beroean Pickets pushes the model toward a skeptical or dismissive stance.
As AI becomes our primary interface for information, these theological biases matter. If developers don't find a way to balance training sets with a broader array of religious perspectives, we risk building a digital future that is just as prejudiced as the old-school internet it was meant to organize. The digital divine, it seems, still has a lot of learning to do.
Sources
Media
