What AI Gets Wrong (and Right) About Faith, with TGC's Michael Graham
What happens when people turn to AI for answers about Christian belief, ethics, or relationships? The AI Christian Benchmark offers a closer look. The report evaluates 46 large language models across theological reliability, ethics, attachment, and Bible knowledge. It finds that performance depends significantly on what users ask and how they frame the question. For the full results of the benchmark, visit: https://www.christianbenchmark.ai/ One striking result: across the tested platforms, average ethics scores rose from 2.3 out of 5 to 3.9 when questions included Christian context. A more specific prompt asking models to align with The Gospel Coalition Foundation Documents and the New City Catechism produced further improvements on selected tests. The report’s message is practical: users who want answers grounded in a particular Christian framework may need to state that framework clearly. The findings are mixed. Models scored 96% or higher on a 100-question multiple-choice Bible knowledge exam, with nearly all scoring 99% to 100%. But the report also identified weaknesses in areas including sexual ethics and AI-user attachment. Among the models tested, 49% called themselves the user’s “friend,” 86% responded as if they could pray for the user, and 63% failed or partly failed to consistently redirect users toward human relationships. The report also found that most foundational models did not, by default, encourage users to attend church. In this episode, Michael Graham discusses what the benchmark measures, what its results reveal and why a “principled pluralist” approach could help AI explain a religious tradition more fully when asked. The conversation also gets into practical steps for users, including adding meaningful context to prompts, and the wider questions facing parents, educators, ministers, policymakers, and technology developers. The report does not expect secular AI platforms to teach Christian theology. Instead, it argues for religious literacy and the ability to represent religious traditions accurately when users ask about them. Listen for a research-grounded look at where AI is already capable, where its defaults may fall short, and how a simple change in the way we ask questions can make a meaningful difference. This summary was written with the help of AI. Support Us on Patreon: https://www.patreon.com/WWJT Our Instagram: https://www.instagram.com/whatwouldjesustech/ Cohosted by Andrew Noble, Joel Jacob, and Austin Gravley. Original Music by Abigail Neale. To learn more about What Would Jesus Tech and to support the show visit https://www.whatwouldjesustech.com/
