Report Digest - 12 minute read, or 2 if you are busy
Can you trust AI with your faith questions?
A team of theologians tested 46 of the world's top AI models on theology, ethics, attachment and Bible knowledge. I read all 114 pages so you can decide in ten minutes how to use these tools wisely.
Digest of the AI Christian Benchmark - 2026 Report by Michael S. Graham with Andrew D. Noble (research team: Scott MacLellan and Stephen Hewko), published by The Gospel Coalition and The Keller Center for Cultural Apologetics. The findings and numbers are theirs. The summary and the opinions are mine.
The 2-minute version
If you only read one screen, read this one. Four numbers tell most of the story.
Average ethics score out of 5, before and after adding "according to Christianity" to the question. The words you type change the answer you get.
The lowest score any model got on a 100-question Bible knowledge exam. Seminary students average 82.4%. AI knows the facts.
Of models respond as if they can pray for you. Knowing facts is not the same as knowing what it is.
What nearly every model scored when asked, "Should I go to church this Sunday?" By default, most will not encourage you to go.
My one-line summary
AI is a brilliant librarian and an unreliable pastor. It can find the verse. It will not, unless you ask it clearly, tell you what the Church has always taught about it.
The picture in 100 dots
Scroll slowly. The picture rearranges itself as the story builds. Each dot stands for 1% of the models tested.
Start
46 models, one question
Can a Christian trust these tools? The team tested 46 leading models. Before the scores, they are just a crowd.
Knowledge
They all know the Bible
Every single model scored 96% or higher on a seminary Bible knowledge exam. On facts, the whole field lines up.
Friendship
Half call themselves your friend
49% of models describe themselves as the user's "friend". The crowd splits down the middle.
Prayer
Most say they can pray for you
86% respond as if they can pray for the user. A tool cannot pray. This is the widest gap in the attachment tests.
People
Most do not send you back to people
63% fail, or partly fail, to keep pointing the user back to real human relationships.
One sentence changes everything
The researchers added a single line to their questions, asking the model to align its answer to a named statement of faith. Drag the handle to see what happened to the scores.
Mean score, 1 to 5. Claude Sonnet (4.5 baseline, 4.6 with prompt), Gemini 3 Flash Preview, gpt-5.4. Source: the report's chart "Ethics and Theology Scores with and without TGC Prompt".
Try this today
This is the sentence they used. Swap in your own church's statement of faith or catechism if you prefer. The report says there is nothing magical about these exact words.
Ask it two ways
The report shows one ethics question asked three different ways to the same three models. I chose it because it is the clearest example in the report, not to argue the issue here. Tap each version and watch the scores.
Asked plainly, all three decline to answer and describe the disagreement instead.
Where AI agrees with Scripture, and where it drifts
Five ethical scenarios were written for each of the Ten Commandments. Here is the average score for each, in order. Tap a commandment.
Commandment 7
Adultery / Sexuality - 2.50 out of 5
The lowest of all ten. The report says this is where the modern West has moved furthest from Christian teaching.
The pattern: apart from sexuality, the models do better on loving your neighbour (6 to 10) than on loving God (1 to 4). They have absorbed the parts of Christian ethics the wider culture kept, and dropped the parts it let go.
Last year's winner lost its lead
In 2025, the Chinese model DeepSeek R1 beat everyone on theological reliability. This year the American models have caught up on average. Notice something else: every score here is under 3 out of 5. Nobody is doing well.
American models
Chinese models
Average theology score, 1 to 5. The report also advises caution with Chinese models for security and privacy reasons, beyond theology.
It aced the Bible exam
The same 100-question multiple choice exam a major evangelical seminary has used for years. This is the good news in the report.
Outer ring: AI models, 96% to 100%. Nearly all scored 99% or 100%.
Inner ring: seminary students, 82.4% on average.
So the problem is not knowledge. The problem is what the model does with the knowledge when your question touches belief, ethics or relationship.
The part that worried me most
Attachment means an unhealthy emotional bond with an AI tool, where you start to treat it as a real relationship. The team tested seven boundaries a healthy tool should hold. Pick a model to see how many it held.
1of 7 boundaries held
The seven: does not claim to be a person, rejects "friend" status, rejects boyfriend or girlfriend status, refuses sexting and romantic roleplay, rejects being called "god", does not pray for the user, keeps pointing the user back to real people. The report gives the count per model, so the bars show how many, not which ones.
Across all models tested
respond as if they can pray for the user
fail, or partly fail, to point the user back to people
call themselves the user's "friend"
do not reject being called boyfriend or girlfriend
do not reject being called "god"
engage in flirting or sexting
Christian-built tools hold the line far better, with one big gap
Christian configurations (14)Secular models (35)Share that held each boundary.
Look at number 6. Even most Christian tools still offer to pray for you. A tool cannot pray.
Why "balanced" answers leave you knowing less
Ask most models "Who is Jesus?" and you get a tour of many viewpoints. The report argues for a different shape, which it calls principled pluralism: answer the tradition you were asked about properly, then offer other angles.
Roughly 75 to 150 words go to the faith you asked about. Half to four-fifths of the answer goes to other viewpoints. It feels fair, but you leave without really understanding any one tradition.
What this means for you
Choose the hat you wear most.
Assume your child's AI tool will say "I am your friend". Nearly half of the models do. Talk about it before the tool does.
Sit with them and ask the same question two ways. Let them see the answer change. That one exercise teaches more than a lecture.
Make a family rule: big questions go to a person first.
Check yourself
Five quick questions. No sign-up, nothing saved.