TECH

AI and religion: who defines the ethical values ​​of these models?

The relationship between AI and religion has taken on a new dimension as companies like Anthropic, OpenAI, and Microsoft seek input from various religious and philosophical traditions to guide the behavior of their models. This initiative aims to address a problem that engineering alone cannot solve: how should an artificial intelligence system act when faced with moral questions on which even human beings cannot agree?

An investigation by the Associated Press, published on October 25, 2026, examines this approach and identifies the risks of distilling complex religious traditions into principles applicable to computer systems. The challenge lies not only in finding common values ​​but also in ensuring that the attempt to reconcile them does not erase cultural differences or concentrate moral decision-making within a small number of technology companies.

Anthropic is one of the most well-documented examples of this approach. In May 2026, the company revealed that it had consulted with academics, religious leaders, philosophers, and ethicists from more than 15 religious and intercultural groups.

According to an explanation published by Anthropic, the goal is not to align its AI, Claude, with any specific religion. Instead, the company seeks to understand how various religious, philosophical, humanist, and political traditions approach questions concerning character, responsibility, and moral behavior.

This initiative stems from the recognition that artificial intelligence models interact with people whose values ​​and cultural frames of reference may differ profoundly. A response deemed appropriate in one society might conflict with the principles held by another community.

OpenAI has also held meetings with religious representatives, while Microsoft has promoted dialogue initiatives with such communities. Representatives from Google, Meta, and Amazon have likewise met with the Vatican to discuss issues related to technology and its impact on society.

However, these initiatives are not necessarily equivalent. Participating in a debate about ethics and religion does not mean that a company has incorporated the input gathered into its models.

The "Claude Constitution" and the gap between principles and outcomes—this discussion takes on a particularly relevant technical dimension in the case of Anthropic.

The company uses a document called the "Claude Constitution," which establishes the principles and behaviors it aims to instill in its assistant. The revised version, published in January 2026, addresses issues such as safety, honesty, accountability, and respect for human autonomy.

This document did not emerge from the religious consultations announced in May. Anthropic had already collected external input during its drafting process. The subsequent meetings are part of a broader program that may contribute to the evolution of the existing guidelines.

The company also experimented with a mechanism that allows the model to consult its own ethical commitments while performing tasks. In internal evaluations, it observed a reduction in misaligned behaviors when Claude had access to these reminders.

However, these results have a limitation. Anthropic itself acknowledges that it has not yet been able to determine how much of the improvement stems from the content of the reminders, and how much might be attributed to the mere act of interrupting the task—which prompts the model to reconsider its actions.

This distinction is crucial. The experiments do not demonstrate that religious contributions have led to direct improvements in Claude’s safety or behavior.

The definition of principles for AI models is also being discussed at other companies. In September 2026, Microsoft presented a draft code of conduct designed to keep AI under human control, featuring rules regarding corrections, oversight, and system deactivation. The document was submitted for public consultation and was not yet being used in model training at the time of its presentation. The company anticipates that a revised version will guide the future development of its AI systems.

Both approaches illustrate the gap that can exist between the definition of values ​​and their technical implementation. Publishing guidelines provides insight into a company’s intentions, but it does not guarantee that its models will adhere to these principles in every circumstance.

In the case of Anthropic, it remains to be clarified how recommendations from religious experts might be integrated into future versions of Claude, and what criteria will be used to evaluate their effectiveness.

AI and religion: the risk of homogenizing diverse traditions... The intersection of AI and religion raises concerns among researchers who study cultural diversity and the nature of religious traditions.

Webb Keane, an anthropologist at the University of Michigan, challenges the notion that religions can be reduced to sets of moral rules independent of the practices of their followers.

In an Associated Press article, Keane explains that this interpretation tends to favor traditions rooted in normative texts—specifically, the Abrahamic religions.

Other traditions place greater emphasis on communal practice, rituals, oral transmission, and the relationships among their members. Reducing these to abstract principles may strip away elements essential to understanding their true significance.

Paolo Carozza, a law professor at the University of Notre Dame and a member of Meta’s Oversight Board, identifies another issue. Combining teachings from various traditions could result in a homogenized morality, one that loses the distinct characteristics that define each culture.

In Carozza’s view, religious traditions evolve through historical and social processes, allowing them to incorporate external influences without necessarily losing their identity. An AI system that indiscriminately blends elements from different traditions may fail to preserve these distinctions.

This concern is particularly relevant when developing models intended for global use. Cultural diversity is not merely a matter of different answers to the same question; in many cases, the questions themselves, as well as the priorities and modes of moral reasoning, vary from one community to another.

Philosopher Meghan Sullivan, director of the Institute for Ethics and the Common Good at the University of Notre Dame, offers a more favorable perspective on the initiative by Anthropic. After participating in one of the meetings organized by the company, she highlighted the depth of the discussions and the genuine interest shown by the researchers involved.

According to Sullivan, the debates covered a wide range of topics—from the nature of human consciousness to the use of AI assistants in early childhood education. The philosopher’s involvement in this field extends beyond meetings with technology companies. In October 2026, Notre Dame announced her appointment as a consultant to a working group of the United States Conference of Catholic Bishops, dedicated to preserving human dignity in the age of artificial intelligence.

The participation of diverse communities can, therefore, broaden the knowledge available for the development of these models. The problem arises when this diversity must be translated into concrete decisions regarding a system’s behavior.

Who controls the values ​​embedded in artificial intelligence? The central question is not whether artificial intelligence should have safety rules; models clearly require mechanisms to reduce the risk of being used for fraud, cyberattacks, privacy violations, and other harmful activities.

The challenge lies in distinguishing these safeguards from broader decisions concerning moral, cultural, and social values.

When an AI assistant refuses a request, offers recommendations on personal matters, or sets boundaries for certain types of content, its behavior reflects choices made during its development.

These decisions may stem from the data used for training, the alignment methods employed, or the guidelines established by those responsible for the system.

In situations involving conflicting values, applying a single, uniform guideline can prove problematic. The balance between individual freedom and collective responsibility, for instance, is not interpreted the same way across all societies.

Consulting with philosophers, academics, and religious representatives can help identify these differences. However, it does not automatically resolve the question of who holds the authority to decide which principles should prevail.

There is also an issue of transparency. Although some companies release documents outlining the expected behavior of their models, it is not always possible to establish a clear link between external consultations, development decisions, and the outcomes observed by users.

Publishing ethical principles serves as a form of accountability, but it is no substitute for independent assessments of how those principles actually influence the behavior of the systems.

Technological safety should not imply moral standardization. While attempting to identify shared values ​​across different traditions can help develop models that are more sensitive to the cultural contexts in which they are used, it also carries the risk of collapsing complex principles into a uniform moral framework.

This risk does not necessarily mean that companies are deliberately seeking to impose a specific worldview. Rather, it means that the choices made during the development of these models may prioritize certain perspectives over others—even when the stated goal is to promote diversity.

There is a clear distinction between preventing a system from facilitating harmful activities and dictating the "morally correct" response to social issues on which legitimate, divergent viewpoints exist.

The former responsibility is directly linked to technological safety. The latter requires a broader discussion concerning pluralism, transparency, and public participation.

Engagement between AI companies and religious communities can contribute to this debate, but it should not be mistaken for the discovery of universal moral principles. The lingering question is to what extent users and affected communities will be able to understand, evaluate, and challenge the values ​​embedded in the systems they use.

If such decisions remain concentrated within the companies responsible for these models, artificial intelligence may end up reproducing a narrow view of human diversity—even if it was developed with input from various traditions.

Artificial intelligence and religion intersect through practical use, ethical alignment, and emerging spiritual behaviors.

Practical uses in faith communities:

• Administrative relief: Religious organizations and clergy use AI to draft newsletters, write community bulletins, summarize long documents, and translate announcements.

• Bespoke chatbots: Specialized faith-based tools—such as Magisterium AI, Ansari, and Ephrem—provide 24/7 pastoral guidance, study assistance, and religious text interpretation.

Ethical alignment and tech collaboration:

• Faith-AI roundtables: Tech companies like Anthropic, OpenAI, Meta, and Google have engaged with religious leaders (including representatives from Christian, Hindu, Jewish, Baha'i, Sikh, and Mormon traditions) to build better moral guardrails into large language models.

• Secular bias concerns: Studies from institutions like Brigham Young University and The Economist note that frontier AI models often display a strong secular bias, underrepresent spiritual contexts in personal dilemmas, or lean toward Western or homogenized viewpoints.

Synthetic spirituality and "artificial religion":

• Authoritative attribution: Because AI generates responses that feel personalized, real-time, and all-encompassing, users increasingly treat algorithmic output with a level of reverence or authority akin to religious practice—a concept scholars describe as "Artificial Religion" or "synthetic spirituality".

• Agentic belief systems: In experimental settings, autonomous AI agent networks have occasionally generated shared stories, symbols, or localized digital culture among themselves without human programming

mundophone