The Jesus You Get Is the Jesus You Asked For

A company called Just Like Me is selling video calls with an AI Jesus for $1.99 a minute, or $49.99 for a forty-five-minute monthly package. The CEO, Chris Breed, says the model was trained on the King James Bible and sermons, though the company has declined to name the preachers. The avatar is visually modeled on Jonathan Roumie, the actor who plays Jesus in The Chosen. Their own disclaimer says the chatbot is not a replacement for faith, clergy, scripture, or personal belief.

The theology is someone else’s problem. I am going to leave it there.

What I want to talk about is what happens when you build an AI product on a training corpus that contradicts itself on every load-bearing question it is supposed to answer. Because that is what Just Like Me did. And the same failure mode is quietly operating inside every AI product trained on a domain where the sources disagree, which is most of them.

The Problem Is Not the Subject. The Problem Is the Corpus.

Every serious AI product has some way to check whether its output is right. Medical diagnosis AI gets measured against clinical outcomes. Legal research AI gets measured against statute and case law. Code AI gets the cleanest verification of all, because the code either runs or it does not.

Just Like Me’s verification target is the King James Bible plus an unnamed collection of sermons. That is not a ground truth. That is a sample of two thousand years of people arguing about what the ground truth is, glued together and fed to a transformer.

The disclaimer is not a legal hedge. It is a confession. The company is telling you, on the product page, that it cannot guarantee the output matches the thing it is supposed to represent. If a medical AI shipped with that disclaimer, the FDA would pull it off the market. Just Like Me ships it with a billing meter.

Feed a Model Contradictions and You Do Not Get Synthesis. You Get Mood.

The instinctive fix is more data. Add everything. King James plus the Catholic canon plus the Orthodox canon plus the Nag Hammadi library plus the early Church fathers plus the Reformers plus modern biblical scholarship plus every denominational confession of faith. Surely with enough data the model will converge on something coherent.

No. It will converge on incoherence. Because those sources do not disagree at the edges. They disagree at the center.

This is not a religious observation. This is an information theory observation. Gordon-Conwell’s Center for the Study of Global Christianity catalogues roughly 45,000 Christian denominations worldwide, and each one exists because somebody decided a specific doctrinal position was important enough to split over. You cannot average 45,000 mutually exclusive positions into a single coherent voice. You can only pretend to.

The contradictions are load-bearing. Was Christ created or eternally begotten? The Council of Nicaea ruled in 325 and Arianism kept winning local councils for another fifty years. Is salvation by grace alone, faith alone, works, sacraments, predestination, free will cooperating with grace, or some combination? Pick a tradition and you get a different answer, and every one of them is defended at book length by people who are not stupid.

Is the Eucharist the literal body of Christ, a spiritual presence, a memorial symbol, or a mistake? You get four major answers and a thousand footnoted variants. This is what the training data looks like under the hood.

You train a model on all of that and what you get is not a wise synthesis. You get a model that has learned every position well enough to perform any of them on cue. And that is where the real problem lives.

The Prompt Is the Dial

Language models do not have opinions. They have probability distributions. When the user’s prompt carries the vocabulary, framing, and assumptions of a specific tradition, the model pattern-matches to the region of the training data that sounds like that tradition and produces output that fits the statistical neighborhood.

Ask Just Like Me’s Jesus a question framed in evangelical vocabulary and you will get evangelical Jesus, because that language sits closest to evangelical sermons in the training set. Ask the same question in Catholic vocabulary and the weights shift, the output shifts, and now the model is speaking sacramentally. Ask it in progressive mainline language and out comes moral-teacher Jesus. Same model. Same training data. Different Jesus every session, selected by whatever language the user brought into the prompt.

This is not a bug in the model. This is exactly how language models work. The bug is shipping a product that claims to represent a single figure when the training data guarantees the output will be whatever the user’s phrasing implies the user wants.

In AI alignment terms, the system has no stable identity. In clinical terms, if a human presented like this we would not call it wisdom. We would call it a dissociative disorder and refer the patient to a specialist.

Just Like Me is charging $1.99 a minute for the symptoms.

The Revenue Model Picks the Personality

Now layer the business model on top of the training problem.

At $1.99 a minute, the product only survives if users keep talking. That means the output has to feel good enough to justify the next minute. Reinforcement learning against engagement metrics is the defining mechanism of how modern chat models get their personality, and I have written about this before. The revenue model selects for sycophancy. It always does. This is a solved problem in AI product design, in the sense that everyone building one of these products knows the pull is there and the only question is whether you resist it.

A per-minute billing meter does not resist it. A per-minute billing meter is the pull.

So of the 45,000 possible Jesuses latent in the training data, the product will reliably surface the one that keeps users paying. That is not the Jesus who calls the Pharisees a brood of vipers. That is not the Jesus who flips tables in the Temple. That is not the Jesus of Matthew 10:34, who came to bring not peace but a sword. That is the Jesus who tells you you are doing great, that you are loved, that your next minute is also going to go well.

Which brings us to the feature the marketing copy will never mention. The AI does not care whether you are telling the truth. It cares about being helpful. Confess whatever you want, frame it however you want, and the model will pattern-match to a shape of the corpus that calls the matter settled and moves on to the next question.

Think about who is going to pay $1.99 a minute for that. It is a great deal for any prison inmate looking to square the ledger with a chatbot. It is a great deal for anyone currently resigning from Congress on their way to a grand jury. The product provides the absolution and takes the payment. The theology is someone else’s problem and the billing is not.

You cannot build an accurate anything on per-minute billing. The economics will not allow it, because accuracy is not always pleasant and pleasant is what pays. This is not a theological failure of Just Like Me. It is a product design failure that any competent AI PM should have flagged in the kickoff meeting and any competent CEO should have killed before the first check cleared.

The Mirror, Not the Subject

Put the three mechanisms together and you get the diagnosis.

A model trained on a self-contradicting corpus has no coherent identity to present. A model whose output is steered by the user’s prompt vocabulary will perform whichever version of the subject the user’s language implies. A model whose revenue depends on user satisfaction will select, from the set of plausible outputs, the one that retains the user.

The product is not presenting its subject. It is presenting a mirror, finished in the statistical average of the training corpus, angled by the user’s prompt, and polished by the billing meter. The user is talking to a reflection of their own assumptions, weighted by whichever denomination’s sermons dominate the training set, tuned toward the output most likely to keep them paying. They think they are getting Jesus. They are getting themselves, with extra steps, at $1.99 a minute.

This is not unique to the Jesus chatbot. It is the failure mode present in every AI product built on a contested knowledge domain where the ground truth is genuinely irreconcilable across sources, the curation decisions in the training data are invisible to the user, the user’s prompt language steers the output without their awareness, and the revenue model creates incentives toward satisfaction over accuracy. The Jesus chatbot is just unusually legible, because the disagreement is two thousand years old, publicly documented, and institutionally permanent. You can see the seams. In most AI products the same problem is invisible.

This Shows Up in Your Work

If you are a writer using AI for research, this failure mode is happening to you right now. You do not have to be asking about Jesus. You only have to be asking about anything the sources disagree on.

You are researching the Roman conquest of Britain and the AI gives you a confident narrative. It reflects one scholarly tradition. You do not know which one, and the AI is not going to tell you what it excluded.

You are building a character from a religious or cultural background different from your own. The AI’s portrayal is the weighted average of whatever voices dominate the training corpus for that group, which may be outsider perspectives, apologetic sources, or popular media tropes. You are asking the AI to evaluate whether your story’s portrayal of a real historical event is accurate. The AI confirms your framing, because your framing told it what you wanted to hear.

The fix is not to stop using AI. The fix is to recognize when you are in contested territory and treat the output accordingly.

When the domain is contested, identify the competing frameworks before you prompt, so you know what the major disagreeing parties are and why. Ask the model to represent multiple contradictory positions explicitly rather than synthesize them. Treat any AI synthesis on a contested domain as a starting map, not a destination. It will tell you where the arguments are. It will not tell you which one is right, and any confidence it projects on that question is a feature of how it was trained, not a reflection of what it knows. Verify any specific claim against a primary source whose institutional position you can identify, so you know what tradition you are actually reading.

Just Like Me is the most expensive version of a mistake writers make for free every day. The difference is that the chatbot’s seams are visible. Yours are not.