Silicon Valley labs turn to religions to shape model ethics, a tech gamble that clashes with conflicting references

AI companies are looking for an ethical framework to teach to chatbots, but this time the answer did not come from computing labs or from mathematical auditing algorithms, but from the corridors of religious institutions. As warnings about the destructive capabilities of language models increase, Silicon Valley technology leaders turn to theologians and philosophers to derive rules that distinguish right from wrong, in an attempt to craft what some see as a programmed conscience for machines.

This turn toward religion was evident in the actions of major firms, and at the Vatican, Chris Olah, co-founder of Anthropic, participated in delivering the public address to Pope Leo XIV on artificial intelligence, emphasizing that the ethical questions raised are not something labs can answer alone, but have been carried by religious traditions for millennia. In contrast, Sam Altman, CEO of OpenAI, expressed a parallel view when he said that what is required of ChatGPT is to mirror the weighted average of humanity’s ethical outlook, as if human values could be calculated as a universal statistical equation.

But this technical conception collides with fundamental objections from humanities scholars.Web Kin, a professor of anthropology at the University of Michigan, points out that treating religion as a booklet of abstract rules that can be extracted apart from lived practices and cultural complexities represents a limited understanding, stripping ethical principles for algorithm training entails a real risk of privileging certain traditions, especially the Abrahamic religions that rest on specific texts and statutes, while marginalizing other religions and traditions that do not primarily view themselves as codified moral systems. Researcher Tomoko Masuzawa supports this argument by noting that categorizing religions as separate, comparable systems is a modern Western concept that originated in the nineteenth century and does not reflect the authentic reality of those traditions.

The ongoing consultations have moved beyond occasional protocol meetings to become regular workshops, Anthropic announced that it holds sessions with scholars, religious leaders and philosophers from more than fifteen religious and cultural groups. Megan Sullivan, a professor of philosophy of religion at Notre Dame, recounts her participation in those discussions, which included representatives of multiple denominations, where the company engineers simplify the models in the first part, while the second part is devoted to debating thorny existential and ethical questions, ranging from whether humans have souls to the legitimate limits of using chatbots with elementary school students for teaching reading and writing. Companies such as Meta, Google and Amazon also held meetings with the Vatican, while Microsoft expanded its outreach to religious bodies.

However, the structural mechanism of generative models raises a dilemma that cannot be solved by aggregating opinions.Large language models rely on statistical prediction of the next word based on vast amounts of text. Paolo Carosa, a member of Meta’s independent oversight board and an academic specializing in Catholic social thought, warns that companies’ drive to merge everything into a single crucible flattens living traditions and erases their particularities in favor of a hybrid average that often leans toward a Western perspective. Human traditions interact and evolve over time in an organic, conscious manner, whereas generative algorithms perform a mechanical, random merging that strips texts of their context and deep meaning.

This debate does not stay within Western academic theory, it directly touches what institutions in the Gulf, Egypt and the Arab Levant are building. When regional banks, ministries of education, or government platforms rely on imported language models whose ethical compass has been calibrated through a weighted Western average, these institutions implicitly import cultural and legislative biases disguised as technical neutrality. The decision facing data engineers and regional technology policymakers today goes beyond inference efficiency and processing speed to the necessity of scrutinizing the values embedded in model outputs, relying on the ethical extracts of California companies imposes an invisible cost to cultural sovereignty, and makes the development of locally built models with explicit value references an operational and regulatory imperative to protect Arab work and education environments from hybrid standards imposed through predictive algorithms.