What AI Gets Wrong (and Right) About Faith, with TGC's Michael Graham
With Michael Graham
Michael Graham of The Gospel Coalition’s Keller Center returns to walk through the 2026 AI Christian Benchmark, which he built with Andrew: 117 questions on theology, ethics, attachment, and Bible knowledge, run against 46 AI models. The models beat seminary students on a Bible exam, scored worst on sexual ethics, and too often offer to be your friend or pray for you. Michael makes the case for a principled pluralist way of answering religious questions, and explains the one sentence that makes a chatbot’s answers line up with your church’s teaching.
NVIDIA founder Jensen Huang has resigned after reports of three affairs, one with a chip executive suspected of spying for China. Joel explains why NVIDIA sits underneath all of AI, and Austin, whose own church lost its lead pastor in June, talks about sin, its consequences, and how to grieve when a leader falls.
Andrew and Austin name seven harms of AI, from least to most important: electricity use, hallucinations, sycophancy, potential for harm, bias, immorality and formation. Austin reports on a Google data center going up in his parents’ Texas town after an election that swung against data centers, calls chatbot attachment a pastoral emergency, and fears pastors who can study only through AI. Andrew asks the question behind them all: who are you becoming?
Obadiah Mariona, who works for Heritage College and Seminary while finishing his master’s degree, joins the hosts to ask how AI is changing education. Andrew brings a study where homework scores rose and exam scores fell, Obadiah explains why he uses AI at work and not in his coursework, and Austin describes how seminary writing has changed since he graduated. Joel makes the case for personalized AI tutoring and Austin asks who can afford it, and the episode ends on authority, the church, and the Heritage Hub, which hosts seminary classes in local churches.
The three hosts are back at the end of the summer, and the news has piled up. Meta settled with forty-seven state attorneys general for up to $17 billion, and Austin walks through what the company actually conceded, why those concessions leave the core product intact, and how Meta turned its own loss into a strike at Snap, TikTok, and YouTube. Andrew noticed that Claude’s memory feature excludes religious beliefs, which leaves a secular default sitting underneath everything else it personalizes, and Joel explains what really happened when models “escaped the sandbox,” whether a VPN is worth paying for, and why an ordinary church is an easy target.
Found this helpful?
Share it with a friend, or support our work and unlock bonus episodes.