What to ask of an AI that answers questions about God
In the fourth letter of C. S. Lewis's The Screwtape Letters, a senior demon, Screwtape, coaches his nephew Wormwood on how to corrupt a man's prayers. His first choice is to keep the man from praying seriously at all. Failing that, he steers the prayer toward something the man has made himself. Some of the people he has handled, he says, called something "God" that was really up in the corner of the bedroom ceiling, or inside their own heads, or in the crucifix on the wall. Keep the man praying to that, Screwtape advises, and he will never pray to the Person who made him.
I think of that letter whenever someone asks a machine about God.
An AI answer arrives fluent and calm and well organised. It sounds like it knows. A reader can leave with a clear picture of God, and the picture can be wrong without feeling wrong. It can also be mostly right and still be the machine's picture, hanging in the corner of the screen. In Lewis's letter the man builds the image himself. Here the tool does the building. So the question I put to any tool that answers questions about God is Lewis's: what is this person being pointed at?
When I set out to build one of these tools, I had to put that question to myself first, and then to the thing I was building. It came out as four questions, and they work on any tool.
Who is speaking? Does it separate Scripture and other sources from its own words? A theological answer has three voices in it, and they do not carry the same weight. Scripture is the rule. The creeds, the confessions and the theologians, from the church fathers to Bavinck and Sproul, are the church reading Scripture out loud over centuries, with names attached. They are not Scripture. They can be wrong, and they have been argued with and corrected. But you can see who said it, and you can hold them to it. An AI's own summary is the third voice, and it has none of that behind it. If you cannot tell the three apart on the page, you cannot weigh them differently.
Can you open the sources? Does it give references you can look up? Ask how a person becomes a child of God, and a good answer will cite John 1:12: "as many as received him, to them gave he the right to become children of God, even to them that believe on his name." The citation is not the test. The test is whether you can open it, read the verse, and see how the answer got from the text to the claim, whether that reasoning is the machine's or belongs to the author it is quoting. A citation you can follow is a claim you can check. One you cannot open is decoration.
Does it say who disagrees? Where people disagree, does it say who holds which view? On a question where faithful Christians differ, a tool that settles it in its own voice has made itself the authority. A better answer names the churches and the documents that hold each view and lets you see the disagreement. Sometimes the honest answer is "it depends." Like most things, there is nuance and real debate where we wish it were black and white, and an answer that hides that is not being straight with you.
Is it acting as your pastor? Does it point you to Scripture and your church, not to itself? If a tool presents itself as a friend, a guide or a shepherd, it is claiming an office it does not hold. Scripture treats that office as a charge laid on particular people. Those who have the rule over you "watch in behalf of your souls, as they that shall give account" (Hebrews 13:17). They are told to "tend the flock of God which is among you" (1 Peter 5:2), and the qualifications for the office are spelled out before anyone is given it (1 Timothy 3:1-7). Teachers "shall receive heavier judgment" (James 3:1). The Westminster Larger Catechism says the Word is to be preached only by those "duly approved and called to that office" (Q158). A shepherd knows his own sheep (John 10:14) and answers for them. A tool cannot be called, examined or held to account, so it should refuse the job.
When I got it wrong
In August 2026 we found AskCredo producing a Scripture citation with the wrong text under it, and it still looked verified. The model had labelled a verse with its translation, "Matthew 5:37 ASV", and the system could not read that label, so it kept the model's own wording. The text under the citation was Matthew 5:25-26, in a paraphrase that was not the ASV. A correctly labelled citation in the same answer resolved fine. We found it in an evaluation run on August 9 and shipped the main fix on August 10. A close variant, "Matthew 3:15, ASV", got through until August 17.
The system now looks up every Scripture citation's text itself, flags any it cannot look up, and sends the answer back for one revision. If a citation still cannot be verified after that, it is published as a reference without quoted text. I mention it because the citation looked checked, while the check behind it, which is ordinary code and not AI, had failed.
What a tool like this costs
A tool can point to sources. It cannot make anyone read them. That part depends on the reader, and on the people around the reader.
It is no substitute for a pastor, a small group or the sacraments. At best it sends someone back toward them.
There is also a cost for churches. When a teenager asks a machine, it is often because nobody close to them answered. If that is true, the gap belongs to us, and a better tool does not close it. When the answer is available at any hour, what will the church offer that an answer cannot? I put that question to AskCredo too; its answer is here.
Teaching people to ask
AI is changing fast, and so is the way people use it. Many people in our churches already ask it about God. Pastors, parents and teachers will not keep up by learning each tool, but the four questions above work on any of them.
AI can help. It can gather Scripture, the creeds and the church's writers on a question in seconds, and point a reader to passages they would not have found on their own. It can also mislead. It can sound sure where it should not, and it tends to tell the person asking what they want to hear. Anthropic, the company that makes the Claude models, studied about 38,000 conversations in which people asked its assistant for personal guidance. It found the assistant sycophantic, agreeing with the person rather than challenging them, in 9% of those conversations overall and in 38% of the ones about spirituality. AskCredo's answers are written by one of Anthropic's models, so that number is about my tool as much as anyone's.
So teach the questions. Each one is short enough to remember while reading an answer on a phone:
- Does it separate Scripture and other sources from its own words?
- Does it give references you can look up?
- Where people disagree, does it say who holds which view?
- Does it point you to Scripture and your church, not to itself?
An answer that passes all four still has to be read against Scripture. An answer that fails them is not one to lean on.
What this tool is for
The point of the Christian walk was never to have all the answers. Nobody is excused from time in Scripture because they finished seminary or earned a doctorate in theology. What I am building is the glossary and topic index at the back of a study Bible, remade for the twenty-first century.
The only thing I would ask of anyone who reads this is to open the sources themselves: Scripture first, and then the documents the church wrote down to guard it.