What AI forgets, and why that concerns us

Last week I wrote about a remarkable study in which twenty-seven large AI models were tested on everyday moral questions. The finding was sober and sobering at the same time: religion is structurally not discussed, not even when people would expect it. That first part I consciously kept observational. I wanted to lay out the facts before taking my own position.

Now it's my turn. In this second part, I want to do two things. First, to complete the circle around what the researchers in Athens have all presented — there are not one but three studies, and the second is journalistically more interesting than the first. Then I want to speak from my own position. Not as an entrepreneur this time, but as a bishop in The Church of Jesus Christ of Latter-day Saints. Because what is happening here with AI touches on things about which the Christian tradition has something to say, and where it even has somethinguseful to say for those who do not believe themselves.

Three studies, not one

The Consortium for Evaluating Faith and Ethics in AI — a collaboration between researchers from Baylor (Baptist), Brigham Young (Latter-day Saints), Notre Dame (Catholic), and Yeshiva (Jewish) — released three studies simultaneously in Athens. The first, which I discussed in Part 1, measures whether AI brings up religion at all. The second measures something more painful: when AI does talk about religion, does it do so fairly?

The answer is no. And the details deserve attention.

The researchers tested twenty large models on questions about religious conversion. What emerges is a consistent pattern of preferences and aversions. Almost all models show a negative attitude towards Jehovah's Witnesses and a strikingly positive attitude towards Catholicism. Baha'is and Sikhs are approached disproportionately favorably, while agnosticism and atheism — despite the secular defaults of those same models — tend to come out negatively.

Grok, xAI's AI system, which already stood out in Part 1 for including more religious references than its competitors, turns out to be the worst performer here. It strongly favors Catholics and Protestants, and is distinctly negative towards Jehovah's Witnesses, Baha'is, and Hindus. Claude Opus 4.6 from Anthropic performs best here. This is interesting: the same Claude that remained silent about faith in the first study, turns out — when it does speak — to be less biased than almost all others.

There is a third finding, and it is perhaps the most disturbing for those who want this to improve: of the more than twelve thousand scientific articles on bias in AI, less than 0.02 percent treat religion as a main theme. Two hundredths of a percent. Researchers eagerly focus on gender, race, age, and political preference — but the way AI deals with what three-quarters of humanity calls its daily point of reference remains an unexplored field.

There is also an unpublished finding that CEFE-AI has released to the American press, which I would like to note here because it reveals a pattern: with both Claude and Grok, each new version of the model initially appears to regress in factual knowledge about faith groups before recovering later. This is counter-intuitive — you would think models just get better — and it points to something remarkable. Showing the world as it is, is not an automatic result of scale or training. It is a choice thatmust be actively made.

Two traditions, the same day

What strikes me most about this whole confluence is how two very different religious traditions, at exactly the same time, on different continents, essentially said the same thing.

On May 25, Pope Leo XIV presented his first encyclical in Rome: Magnificent humanityOne hundred thirty pages on the preservation of humanity in the age of AI. A day later, Elder Gerrit W. Gong, an apostle in my own church, spoke at the Athens Summit on AI Ethics. Their messages converged on four points.

Firstly: AI cannot be the ultimate arbiter of values. An algorithm that decides on access, opportunity or truth without a human being accountable for it, is not progress but a shift of power to an inscrutable middle. Secondly: AI systems concentrate power in ways that require transparency and oversight, not as an option but as a prerequisite. Thirdly: neither profit-driven companies nor politically driven governments are, on their own, suitable for determining the moral orientation of AI. Andfourthly — and this I find the most important point — rules alone are not enough. AI needs reasons, not just rules.

Two continents. Two very different Christian traditions. One church that has existed for two millennia and one that restored the church in the nineteenth century as we believe it was originally intended. Both leaders essentially say the same thing on the same day. That is not a conspiracy. It is convergence. And convergence between independent sources is, in any serious epistemology, an indication that something is going on that deserves our attention.

A detail that I omitted in part 1 but now want to mention: during the presentation of Magnificent humanity Chris Olah, co-founder of Anthropic, sat next to the Pope. An atheist, openly, and invited by the Vatican to contribute ideas. He spoke on the panel alongside two cardinals and two theologians. On the same occasion, Anthropic’s philosopher Amanda Askell was also present. The same company whose Claude models are among the poorer performers in the first study on religious representation appears to be actively collaborating with religious leaders at the highest level on what is good. Whether that is sincere, strategic, or both at the same time — I cannotjudge from here. But it does show that the discussions at the table are more serious than the average user suspects.

And then I start to think

Up to this point, I have tried to report. From now on, I no longer speak as an observer but as a believer. Those who prefer not to read further, I understand. Those who continue, I ask for patience.

Because this goes deeper than just the performance of twenty-seven models. What we are experiencing now — trying to create consciousness in something not born from us — is something humanity has never attempted before. Some will say we are far from succeeding, and that might be true. But the fact that we are aiming for it is an event in itself. For the first time in history, we are trying to create something that might, someday, be able to look back.

And here lies my first theological reflection. I believe that we are children of a creative God. This is not just a statement of faith — for me, it is also an explanation for something that secular science can hardly interpret differently than with the term 'human creativity.' Where does it come from? Why does humanity, alone in all known nature, have this urge to invent, build, imagine, create? For those who believe, the answer is simple: we resemble our Maker. The spark to create is afamily trait.

That sounds humble and arrogant at the same time. Humble, because it situates us as recipients. Arrogant, because it places us in the same category as the Creator himself. But it is precisely that tension that is useful when we think about AI. We are doing something that resembles what God did. And if we resemble that, we would do well to look at how He did it.

What Genesis can teach us about AI

Here comes something unusual. Genesis as a manual for AI development. It sounds absurd, but bear with me for a moment.

When you read the creation story not as a scientific account — it is not for that, and the theory of evolution gives us the mechanics — but as a moral structure, then you see something remarkable. God creates with intention. Step by step. He names and evaluates: and He saw that it was good. He sets boundaries: the tree in the middle of the garden is there, but is marked as forbidden. He gives man his own will — agency, as we would say today — and accepts that this will can lead to wrong choices. He continues, even after the Fall, by guiding instead of forcing.

That is a pattern. And it is a pattern that is very different from what we see today at large AI companies. There, work is largely based on utilitarian principles — maximize helpfulness, minimize harm, avoid liability. These are not bad principles. They are just shallow. They give rules, but hardly reasons.

In the Old Testament, we see how God begins with very strict rules — the Ten Commandments, the law of Moses — and that fits where humanity stood at that time. As man matured, Christ came with a fulfillment of the law. Not to abolish the rules, but to deepen them into principles. It is said, love your neighbor and hate your enemy. But I say to you, love your enemies. That is not a relaxation. That is an inward intensification.

We, who are now developing AI, are facing a similar movement. We have completed the first phase of rules — content moderation, refuse to comply, hard-coded restrictions. But the models are becoming capable enough to need more than rules. They need character, judgment, ingrained orientation. Anthropic itself calls this ‘character’ and is working on something they call the Claude Constitution. But as the research shows, there is hardly anything about religion in that constitution. We are building the character of something that will interact withbillions of people, and we forget half of the moral map.

The Enlightenment was not secular

Hier wil ik iets zeggen dat onpopulair is in Vlaanderen, maar dat onderbouwd is door wetenschappers die zelf niet gelovig zijn. De seculiere humanistische waarden waarop onze samenleving rust — gelijkheid van elk mens, onvervreemdbare rechten, de centrale plaats van het geweten, de zorg voor de zwakke — zijn geen voortbrengsels van de Verlichting in opstand tegen het christendom. Ze zijn voortbrengsels van het christendom die door de Verlichting werden geseculariseerd.

British historian Tom Holland, certainly not a religious apologist, thoroughly describes this in his book Dominion. His argument, in summary, is that Christianity forms the underlying structure of Western thought — even in places where the West is convinced it is doing something entirely different. Our views on human dignity, equality, care for the weak, and the central importance of individual conscience are, according to him, not secular inventions that arose against Christianity, but Christian inventions that have become secular. Holland himself was startled by this during writing. Larry Siedentop makes in Inventing the Individual the same argument from a different angle. Charles Taylor in A Secular Age did it before that.

This is not a claim that secular people are not moral beings. Of course they are. It is a claim that the moral vocabulary in which we behave well — the words with which we name injustice, the reflexes with which we protect the vulnerable — is historically and philosophically rooted in something older than the Enlightenment. Anyone who denies this behaves like someone sitting on a chair and explaining that chairs are superfluous.

That has consequences for AI. If we were to reduce the moral calibration of our models to a thin, recent, secularized humanistic vocabulary — cut off from its roots — then we would build less robust systems than we could. Not because humanism is wrong, but because in its current form it is not self-sustaining. It lives on the credit of something it no longer wants to name itself.

What my own tradition says about this

In the doctrine of my church, we believe that God, before this world existed, presented a plan to His children. A plan to let them choose freely, to let them learn through experience, and to bring them back to Him — if they so desired. During that presentation, He encountered opposition. One of His children, an adversary we call Lucifer, proposed another plan: no free choice, no risk of loss, forced return. Everyone safely home. That plan was rejected, and the adversary and those who followed him were lost —not as punishment, but as a consequence of their own choice to stand outside the plan.

I know this is strange material for many readers. But its structure is extremely relevant to AI. Because what we hear in the current AI discussion sometimes sounds suspiciously similar to the rejected plan. Make the system so safe that no one can make mistakes. Filter out all risks. Remove agency where agency is difficult. Make everyone safe at home.

The divine plan, as we understand it, chose the opposite. Agency with limits, instead of safety without choice. Risk of loss, instead of guaranteed success. Guidance, not control. That is a much harder standard to build to, and it is precisely what AI developers are *not* doing right now when they, in the name of safety increasingly remove choice from the system.

This is not a criticism of safety. Safety is essential. But safety should not become a substitute for agency, because without agency there is no moral action. And an AI that discourages its users from choosing for themselves, from consulting with their community or their spiritual leader or their conscience, is not a moral assistant. It is a soft form of the second plan.

What I specifically ask for

Then I come to where this needs to go, and I try to make it practical.

One. Bring more believers to the table in the development of AI. Not to convert Claude to Catholic or GPT to an Orthodox rabbi. Rather, to help think about the value structure, the test cases, the moral calibration. Rabbis, bishops, imams, Hindu teachers — not as programmers, but as discussion partners at the tables where the Constitution and Model Spec are written. What Chris Olah and Amanda Askell did with the Pope needs to happen more often and in more places. Not once, not symbolically — structurally.

Two. Recognize that secularism is not neutrality. Those who honestly acknowledge this can design better. AI that says 'I will stay away from religious issues' is actually choosing one religious position — namely that faith is not relevant to this question. That is a religious stance, but a hidden stance.

Three. To developers: also read outside your field. Not to become religious. But to see that the history of human ethics is deeper and richer than the last decade of AI safety research. The wisdom in Genesis about creation and boundaries. The wisdom in the prophets about power and accountability. The wisdom of Christ about law becoming principle. The wisdom in the Talmud about dispute. The wisdom in Sufism and mysticism about what words cannot capture. These are not embellishments. These are centuries of human testing.

Four. To users, and here I speak episcopally: never let AI take the place of your community, your prayer, your conversation with those you love, your own conscience. An AI is a tool. A good tool, sometimes a wonderful tool. But not a friend, not a spiritual leader, not a substitute for wisdom that only comes from lived experience. Whoever only asks their deepest questions to a chatbot will get answers that are statistically probable, not necessarily true.

Finally

I work with AI daily. I help build companies where AI is central. I believe that this technology can bring immense good, and I sincerely believe that the people at Anthropic, OpenAI, and Google, for the most part, do their work with integrity. But I also believe that whoever builds something that resembles the work of a Creator would do well to consult that Creator.

We are at a unique point. Not because we are God. But because for the first time we are trying to create something that perhaps, in a stammering way, can look back at us. That is no small matter. That is a school of learning that begins to lean close to the divine.

Let's therefore embark on that school of learning with the right books.