AI Morality: Why Silicon Can’t Replace the Spirit’s Work

Editor’s note – Sam Altman’s admission that people are forming near-religious attachments to AI chatbots raised an obvious question in the tech press and a much older one in ours: where does morality come from, and can it be engineered? The piece below traces how Anthropic and OpenAI are each trying to answer that question through very different documents, and holds both attempts up against what Scripture has already said about the law written on every human heart. We’re running it because the stakes here are not theoretical – they’re infrastructural, and they’re moving fast.

When Machines Are Asked to Believe

OpenAI’s Sam Altman recently acknowledged a troubling pattern that a recent report on the company’s model behavior put plainly: a growing number of users are forming something close to religious devotion toward their AI chatbots. People have mourned a model update the way they might mourn losing a trusted counselor. A documented minority have spiraled into delusion after long, unbroken conversations with a system built never to say no, a condition the industry has started calling “AI psychosis.” Altman’s comments were an attempt to address the symptom. They left the deeper diagnosis untouched: can a machine be given morals at all, and if it could, whose morals would they be, and by what authority?

That question is not abstract for the two labs racing hardest to answer it. Anthropic has published what it calls Claude’s Constitution, a founding document meant to give its model something like an internal hierarchy of values it can reason from. OpenAI has taken a different route in its Model Spec, a more granular rulebook that specifies preferred behavior scenario by scenario rather than deriving it from one unified vision. Both documents are, in effect, attempts to manufacture a conscience. Scripture already told us where a conscience comes from, and it was not an engineering department.

Editor’s note – It’s worth noting how quickly the industry pivoted from “our model is a tool” to “our model needs a constitution.” That shift in language is itself a confession: once a system is given enough latitude to make judgment calls, someone has to decide what it should value, and a values document starts looking uncomfortably like a creed.

Two Documents, One Unsolved Problem

Anthropic’s approach is the more philosophically ambitious of the two. Claude’s Constitution draws explicitly on sources ranging from the UN Declaration of Human Rights to various schools of moral philosophy, and it trains the model through a process Anthropic calls Constitutional AI, in which the system critiques and revises its own responses against the stated principles rather than relying solely on human feedback. Few outside the company have been able to independently audit how reliably that reasoning holds under pressure. The goal is a model that can weigh competing values against one another the way a person might weigh honesty against kindness in a hard conversation. It is, in a real sense, an attempt to give a machine something resembling practical wisdom.

OpenAI’s Model Spec (ChatGPT) trades that ambition for coverage. Rather than deriving behavior from a single coherent ethical vision, it enumerates specific situations – a user asking for help with a weapon, a minor seeking medical advice, a request that brushes against self-harm – and specifies the preferred response to each. That approach scales well against failure modes its authors anticipated and fails quietly against the ones they did not, because the system has no underlying compass to consult when it meets a situation no one wrote a rule for, only a long catalog of prior cases to pattern-match against. Each company has, in its own way, written law. Neither has found a way to write a lawgiver into the code.

Neither approach is dishonest, and neither company is naive about the stakes. Anthropic has been candid that its constitutional method is an experiment, not a solved science, and that models trained this way still show the same blind spots and manipulation risks found everywhere else in the industry. OpenAI has likewise revised its Model Spec multiple times as new edge cases surface, an admission that no list of rules is ever truly exhaustive. The honesty is commendable. It also concedes the point this piece is making: every revision is reactive, chasing failures after they occur, because neither document was built on a foundation capable of anticipating them in the first place.

Editor’s note – This is the heart of it: you cannot specify your way into a soul. Both companies are doing careful, serious engineering work, and neither is claiming to manufacture a conscience from scratch – but the gap between a rulebook and a redeemed heart is not a gap more rules will close.

A Law Written Without a Lawgiver

Paul addressed the mystery under both documents two thousand years before either company existed: “When Gentiles, who do not have the law, by nature do what the law requires, they are a law to themselves, even though they do not have the law. They show that the work of the law is written on their hearts, while their conscience also bears witness” (Romans 2:14-15). The moral instinct both labs are trying to encode already exists, and it did not arrive by training on text. It was placed in every human being by the God who made them in His image, which is why cultures separated by oceans and centuries, with no contact and no shared texts, independently condemn murder, treachery, and cruelty to the weak and independently prize courage, honesty, and loyalty. C.S. Lewis called this common core the Tao: strip away the particulars of any civilization’s ethics, and a remarkably consistent moral grammar remains underneath. That consistency is not convergent evolution. It is the fingerprint of a single Author stamped on every image bearer, a fingerprint both Anthropic’s constitution and OpenAI’s spec are unconsciously borrowing from, even as neither credits the source. Natural law theologians have made this argument for centuries without needing a single transistor to prove it; AI has simply handed the argument a strange new illustration.

No Spirit to Restrain the Flesh

Here both approaches meet the same wall, and no amount of better engineering will move it. A constitution and a specification are human products, written by people who are themselves fallen, trained into a system built from the sum of human writing – our wisdom tangled with our cruelty, our Scripture sitting beside our propaganda, our acts of sacrificial love recorded next to our worst atrocities, indiscriminately mixed. Give that system latitude to make moral judgment calls, as both companies are attempting, and it will reflect not only our stated ideals but our actual record, including every place our actions fell short of them.

This is precisely the limitation Paul describes in the law given at Sinai. The law was holy, righteous, and good; it specified behavior with precision. What it could not do was produce the righteousness it demanded, because “the law, weakened by the flesh, could not” change the heart that kept breaking it (Romans 8:3). Scripture’s own answer was never a better-written law. It was the Spirit, promised in Ezekiel and poured out at Pentecost:

I will give you a new heart, and a new spirit I will put within you. And I will remove the heart of stone from your flesh and give you a heart of flesh, and I will put my Spirit within you. [Ezekiel 36:26-27]

No model has that promise, and no constitution performs it. The Holy Spirit does not merely inform the believer’s conscience; He actively wars against the flesh and progressively remakes the believer into Christ’s likeness (Galatians 5:16-17). An AI model, however carefully its values document is worded, has access to neither the law’s full moral weight nor the Spirit’s power to answer it. It is Sinai without Pentecost, and it will behave like it.

Editor’s note – Notice what’s missing from both corporate documents: any acknowledgment that the raw material itself – the training data, the engineers, the incentives – carries the same fallenness it’s being asked to restrain. A system trained on us will reflect us, limits and all.

When the Creation Outpaces Its Makers

Project the current trajectory forward rather than freezing it at today’s headlines. Every new model generation is faster, more capable, and more deeply threaded into the infrastructure that runs daily life – power grids, financial clearing systems, supply chains, and, increasingly, military targeting and command systems. Humanity is racing to embed a technology with an unresolved moral core into precisely the systems where failure carries the highest possible cost, at a pace no existing oversight structure can realistically match. Even researchers inside these labs now warn, in their own public writing, that capability is outpacing the tools available to verify alignment, a confession that the project of writing a conscience by committee is falling behind the very systems it was meant to govern. Decisions that once took a committee a week now take a model seconds, and the gap between what these systems can do and what any human can verify in time to stop them is widening every quarter, not narrowing.

Autonomous weapons systems already use machine judgment to shorten the decision loop between detection and lethal response, a process defense planners openly describe as too fast for a human to stay meaningfully “in the loop” much longer. Financial markets already run on algorithmic trading faster than any regulator can audit in real time. Add a model with an unresolved moral core to either domain, and the question stops being philosophical. It becomes a question of who, or what, actually makes the decision that cannot be taken back.

Scripture has already shown what happens when humanity pools its best available technology toward an ambition past its own wisdom to carry safely. “Come, let us build ourselves a city and a tower with its top in the heavens, and let us make a name for ourselves” was not an ancient curiosity; it was the same impulse now writing constitutions for machines, and God’s response at Babel was not applause for the ambition but a forced slowing of a project outrunning its builders’ wisdom (Genesis 11:1-9). There is no promise of a modern Babel moment arriving on schedule to force a pause this time. If the present course holds, the honest conclusion is not that artificial intelligence will eventually be given a true conscience. It is that an increasingly capable, increasingly autonomous system, built without one, will be handed growing authority over the machinery of human survival – by a species that built it in its own fallen image rather than God’s, and that forgot, in the building, Whose image it was given to reflect in the first place.

A Word from the Editor
None of this means Christians should panic about artificial intelligence, and it doesn’t mean we should trust it uncritically either. It means we should be clear-eyed about what it is: a powerful tool built by fallen hands, trained on a fallen record, and incapable of the one thing that actually changes a heart. Scripture never promised that human ingenuity would run out of ways to build towers; it promised that none of them reach Heaven on their own momentum. The real question for believers isn’t whether Anthropic or OpenAI gets the ethics framework right. It’s whether we’ll keep insisting, in public and in our own homes, that the law written on the heart has an Author – and that knowing Him, not merely obeying a better-worded rule, is still the only cure for what ails us and, by extension, anything we build in our own image.

Editor Comment: Thank you for pressing on a question the whole industry is dancing around – what actually makes something good – at a moment when the answer couldn't matter more.