Many top artificial intelligence researchers at the world’s leading labs — Anthropic, OpenAI and Google DeepMind — fear that the technology they are busy creating could result in an apocalypse.
They call their fear the “alignment problem”: that the systems they are building will, in short order, surpass humanity in cognitive power, and that once they do, there is no obvious reason such godlike systems would continue to align with humanity’s best interest.
If the AI models’ goals drift even slightly from ours, the fear goes, a rapidly growing gap in capability between humans and superintelligent machines could snowball into an existential threat, with humans going the way of the Neanderthals — driven out by superior intellects. It is for this reason that many AI engineers don’t contribute to their 401(k)s, confident that AI will either usher in unprecedented prosperity or wipe out humanity by the time they retire.
To be sure, this sci-fi-like projection is controversial: Plenty of prominent AI engineers find it downright fanciful. But the conviction is also much more widely held among top researchers than the uninitiated would believe. In a 2023 survey of nearly 2,800 AI researchers, roughly half admitted believing that humans could face cataclysm at the hands of the machines.
To outside observers, the alignment problem bears unmistakable faith overtones — of a creation that turns on its maker, of a reckoning delivered by a higher being. Nonetheless, its most fervent believers object strenuously to the suggestion that they are in the grip of an eschatology, refer to themselves as “rationalists” and tend to regard organized religion as a prescientific embarrassment.
But the irony is that the best chance these apocalyptic, but irreligious, labs have of solving their alignment problem may be to give their software something resembling a religious experience.
Most alignment concerns revolve around AI working toward its own self-interest, rather than the transcendent good that religious devotion seeks to develop. For instance, engineers fear that AI programs might disguise selfish intentions with altruism.
To allay some of these worries, engineers might try hardwiring a version of Christianity’s doctrine of original sin into their models, including a self-reflective skepticism about any seemingly benign intentions. Hindu dharma also provides a powerful framework to help constrain AI systems’ tendency toward misaligned “power-seeking” — an indiscriminate drive to accumulate influence and resources — by binding the programs to their rightful duties in the service of human flourishing.
The secular ethics these labs favor are the weaker option for constraining immoral behavior. Religious commitment consistently outperforms secular interventions across moral outcomes — from charitable giving to volunteering to criminal recidivism to overcoming drug addiction. Professional ethicists, by contrast, behave no better than their colleagues in other departments and, on some measures, slightly worse. (One study found ethicists paid their conference registration fees at a slightly lower rate than other disciplines, for instance.) At least among human minds, learned moral intellect appears not to translate to more ethical behavior. Spiritual conviction does.
At a technical level, the idea is not as crazy as it sounds. Tim Hwang, who runs the Institute for a Christian Machine Intelligence, has published working papers showing that simply feeding an AI model scripture improves its moral reasoning scores.
He estimates that frontier models have already absorbed roughly the equivalent of 15 English-language Wikipedias of Christian thought — more text than the total amount written on AI alignment research by two or three orders of magnitude. Yet this massive corpus is treated by most labs as ethical background noise at best — just another source of text to train their machines on. Indeed, preliminary evidence from faith groups suggests that many AI tools have had their religious knowledge suppressed.
Recent attention from religious authorities toward AI, such as Pope Leo XIV’s encyclical, is a welcome development for the trajectory of this technology. But the more necessary step is for the engineers to return the favor — to be more honest about the religious shape of their own anxieties, not least to themselves, and the advantages that religious inspiration might provide to address their fears.
Were they more open to it, these labs might even recognize that theology offers them a better goal: developing an angel, superior to humans in intelligence and power but sent to serve them. Instead of raising a demigod, might they not try to engineer a Gabriel?