Anthropic Consults 15+ Religious and Cultural Groups on AI Ethics, but Key Results Remain Unverified

AI developers are holding talks with religious leaders, philosophers, and ethicists in an effort to improve the ethical behavior of their models. According to an Associated Press (AP) report dated October 25, 2026, technology companies including Anthropic are trying to bring the knowledge accumulated over the long histories of religion and philosophy into AI development. Chris Olah, a co-founder of Anthropic, said at a meeting held at the Vatican that questions about AI ethics cannot be answered by a single laboratory. His position is that perspectives from a range of cultures and schools of thought must be incorporated alongside technical expertise.

However, the fact that dialogue with religious communities is under way must be distinguished from AI actually having acquired ethical judgment. Seeking advice from outside experts is meaningful, but it does not by itself mean that AI safety has improved. What engineering can verify is observable behavior: what responses a model generates and what actions it chooses under particular conditions. That is separate from the philosophical question of whether machines can have a human-like conscience, faith, consciousness, or moral agency.

For example, even if an AI generates text expressing ethical concern, or responds in a way that appears to reflect on its own judgments, this does not prove the existence of inner conviction or remorse. What matters is what technology produces such behavior and how much it contributes to real safety gains. Comparing Anthropic's official documents, its Constitutional AI research, and the Vatican's public documents shows that a gap remains between the ethical AI that companies aim for and the technical results demonstrated so far.

Anthropic Has Talked With 15+ Religious and Cultural Groups: How Far Has This Reached the Model?

In an official announcement on May 19, 2026, Anthropic said it had held early-stage consultations with people from more than 15 religious and cultural groups, including religious scholars, clergy, philosophers, and ethicists. The participants span major religious traditions as well as philosophy, secular humanism, and a range of political viewpoints. The company says that in thinking about AI safety and social impact, the insights of people who have long reflected on human values and ethics are important, not just those of engineers.

The purpose of the dialogue is not to make Claude follow a particular religion or doctrine. Anthropic explains that it aims for a model that can understand diverse perspectives, whether religious, secular, or political, with equal depth. The goal is for AI, when facing complex situations, to make judgments that take different ways of thinking into account without leaning excessively toward any one set of values. However, this is a development policy the company has set out; it has not been demonstrated that Claude can behave neutrally toward all values.

It is also unclear how far these consultations have been reflected in Claude's training. Anthropic explains that it gathered outside input when drafting its behavioral guidelines, which it calls a constitution, and that the dialogue gave it ideas for new experiments. But details of which participants' remarks or proposals were used in which training stage of which model have not been made public. The fact that outside dialogue took place and the claim that it changed a model's parameters or actual behavior need to be verified separately.

{
  "type": "bar",
  "title": "公開資料から確認できる対話活動と技術検証の状況",
  "unit": "件数・集団数",
  "categories": ["宗教・文化集団（15超）", "言及された他社（4社）", "公開された定量ベンチマーク（0件）"],
  "series": [
    { "name": "公開情報に基づく数値", "values": [15, 4, 0] }
  ],
  "caption": "15は「15超」という公表値の基準点を表す。項目ごとに数えている対象が異なるため、数量の直接比較には適さない。",
  "source": "Anthropic公式発表・AP通信報道"
}

Contact with religious communities is not limited to Anthropic. According to AP, representatives of Meta, Google, and Amazon have also met with Vatican officials, and Microsoft has actively engaged in dialogue with religious leaders. However, each company's efforts are independent activities, not an industry-wide program based on common standards or methods. Sufficient information has also not been disclosed on how far the content of these exchanges has been reflected in each company's commercial models.

The Priorities in Claude's "Constitution" and How They Are Taught to the AI

To understand the mechanism by which religious and ethical insights could be reflected in AI behavior, one needs to grasp the idea of "Constitutional AI," which Anthropic uses. The company's Claude's Constitution is a document that describes in natural language the values and principles of conduct expected of the model, and it plays an important role in shaping Claude's training policy. Rather than only specifying in detail how to respond to individual questions, it sets out basic thinking on what to prioritize when a situation is hard to judge.

The constitution Anthropic published in 2026 sets out four principles Claude should follow, along with their order of priority:

1. Broadly safe: Not undermining human oversight and correction of AI.
2. Broadly ethical: Behaving honestly and avoiding danger and harm through sound value judgments.
3. Compliant with Anthropic's guidelines: Following the operational guidelines the company sets for particular situations.
4. Genuinely helpful: Being substantively useful to users.

When these conflict, Claude is in principle designed to give priority to the higher item. In other words, maintaining human oversight and avoiding ethical problems are weighted more heavily than meeting a user's request. However, this order of priority was designed by Anthropic; it is not an ethical system the AI derived independently. And while the company emphasizes flexible value judgment, it also sets strict prohibitions on actions that could lead to serious harm, such as materially assisting a biological weapons attack.

How are such natural-language principles built into actual machine learning? The underlying mechanism is described in the paper "Constitutional AI: Harmlessness from AI Feedback," published by Anthropic researchers in December 2022. The method proposed there has two broad stages: supervised learning that uses self-critique and self-revision, and reinforcement learning using AI feedback (RLAIF).

In the first stage, supervised learning, the model is made to generate answers to harmful questions and then to critique and revise those answers against the constitution's principles. For example, if an answer encouraged dangerous behavior, the model itself is made to point out the problem and rewrite the answer in line with the principles. The revised answers are used as training data to fine-tune the model, teaching it a tendency toward desirable responses.

In the reinforcement learning stage that follows, a separate AI model compares and evaluates multiple candidate answers generated by the model, based on the constitution's principles. A preference model is trained from those evaluations, and the reward it outputs is used to optimize the original model. This uses AI evaluations as the learning signal instead of having humans rate every answer directly. Note that "AI evaluation" here does not mean that no human involvement is needed. The constitutional principles that serve as evaluation criteria are themselves designed by humans.

In the 2022 experiments, the model used to generate initial answers was an assistant that prioritized helpfulness and had not been specially tuned for harmlessness. The point was to deliberately generate answers that might be problematic and then revise them according to principles, in order to examine the learning effect. The study evaluated practical metrics such as whether dangerous responses could be suppressed while avoiding unnecessary refusals.

However, this paper describes the technology as of 2022, and it is not certain that exactly the same training process is used in today's commercial Claude models. What was evaluated was also observable model behavior such as harmlessness and helpfulness. It was not research testing religious learning, human-like moral growth, or the existence of consciousness.

"Ethical Reminders" Reduced Problem Behavior, but the Cause Remains Unclear

In its May 2026 announcement, Anthropic also reported a new behavior-control experiment inspired by dialogue on religion and ethics. Focusing on how people facing difficult decisions recall the words of a trusted advisor or their own beliefs, the company tried a tool that lets Claude check its ethical principles while working. Specifically, when the model calls the tool as needed, it receives a short reminder summarizing its ethical principles of conduct.

According to the company, Claude sometimes called this tool on its own initiative, for example just before an action that could have serious consequences. It was also observed pointing out possible conflicts of interest on its own part. In experiments where the tool was built into the decision-making process, the company reports that the rate of behavior deviating from intended principles dropped substantially across multiple internal alignment evaluations.

This is an interesting result as a way to improve AI safety, but it calls for cautious interpretation. The calling of a tool to check ethical principles can be explained as part of the model's reasoning process and API handling. It gives no grounds for interpreting that the AI exercised a human-like conscience and held back from an action after inner conflict.

More importantly, not enough information to verify the experimental results has been made public. Anthropic's announcement does not include specific figures showing how much problem behavior decreased, statistical confidence intervals, the number of tasks evaluated, detailed versions of the models used, or benchmark specifications. It is therefore difficult for outsiders to judge how large the reported improvement is or whether it can be reproduced under other conditions.

{
  "type": "bar",
  "title": "倫理リマインダー実験に関する情報公開の状況",
  "unit": "開示の有無",
  "categories": ["定性的な挙動の説明", "統計的効果量", "誤差範囲・信頼区間", "査読済み論文"],
  "series": [
    { "name": "公開状況（1=あり、0=なし）", "values": [1, 0, 0, 0] }
  ],
  "caption": "2026年5月19日のAnthropic公式発表における開示状況。数値は各項目の公開有無を表し、実験の性能を示すものではない。",
  "source": "Anthropic公式発表"
}

Another important problem is that it is not known what produced the improvement. Anthropic itself acknowledges that it has not yet separated whether the effect comes from the ethical content in the reminder or from the model getting a chance to pause and reconsider. In other words, it cannot currently be determined whether rereading ethical principles was effective, or whether simply adding an extra reasoning step was.

In large language models, additional reasoning steps or computation can sometimes improve the accuracy of answers and the stability of judgments. Controlled experiments are therefore needed, comparing against reminders with no ethical content or against conditions that add only the same amount of reasoning time. Without matching these conditions, one cannot conclude that religious or ethical insight itself contributed to the safety gains.

The report is a preliminary result based on Anthropic's internal evaluations, not a peer-reviewed paper or an independent preprint with detailed experimental data. The company has said it plans to share further results, but until additional verification is published, qualitative reports of improvement need to be distinguished from scientifically established effects.

The Difficulty of Reflecting Diverse Religious Values in AI, and Remaining Questions of Verification

Efforts to bring religious and ethical insight into AI face not only technical problems but also the hard question of how to reconcile differing values. Standards of good and evil, and the values people prioritize, do not necessarily coincide across religions and cultures. If a company tries to incorporate multiple positions without adequately understanding each doctrine and tradition, and extracts only the shared elements, the original meaning may be lost.

Paolo Carozza, a professor at the University of Notre Dame who spoke to AP, pointed out that in the process of trying to merge various religious traditions into one, each one's distinctive features may be diluted and converge on an inoffensive middle-ground view. He also mentioned the danger that Western views of religion could be implicitly privileged. These concerns are not reports that flaws have actually been confirmed in a particular AI model, but critical views on how religion and values are handled.

The problem is that religion cannot always be treated as a simple set of ethical rules. Religious traditions are not made up of doctrine alone; they are deeply tied to ritual, culture, communal relationships, and historical practice. Simply combining principles extracted from different traditions does not guarantee that the result is fair to each culture. The company's discretion inevitably enters into the final decisions about whose values to adopt and how to rank them.

In a related development, the Catholic Church on May 15, 2026 published Pope Leo XIV's encyclical "Magnifica Humanitas." The document argues that human dignity and the benefit of society as a whole should be at the center of technological development, including AI. It also notes that developers themselves do not fully understand the inner workings of AI, and stresses the need for ethical and social judgment alongside deeper scientific research.

The encyclical also expresses concern about the concentration of technological power in a small number of companies and organizations. If important decisions about AI design and operation are made by a limited group of people, their values could affect society as a whole. To prevent this, it calls for clearer accountability, independent oversight, and appropriate legal regulation. However, the encyclical is a document discussing, from a religious and social standpoint, what desirable technological development looks like; it is not academic research that experimentally clarifies the capabilities and limits of AI.

Positions also need to be distinguished on whether AI has a conscience or moral agency. In Claude's constitution, Anthropic explains that great uncertainty remains at present about whether AI could have consciousness or moral status. The Vatican's encyclical, for its part, takes the position of clearly distinguishing human experience and conscience from AI's information processing. Either way, the fact that an AI generated text expressing emotion or ethical conflict cannot scientifically prove the existence of subjective consciousness or suffering.

To assess whether dialogue with religious communities can improve AI safety, reproducible experimental results are needed, not just corporate philosophy and qualitative reports. For the ethical-reminder experiment in particular, this means clarifying the model version, evaluation tasks, improvement rates, and statistical margins of error, and comparing against a control group given the same reasoning time. Transparency is also important so that it can be traced how insights from outside experts were reflected in constitutional revisions, safety evaluations, and model training.

The efforts of Anthropic and others are notable as a move to consider AI safety in light of society's diverse values rather than leaving it to engineers alone. But incorporating religious and philosophical insight is not the same as actually improving AI behavior as a result. The focus going forward is whether the principles drawn from dialogue can be shown, in a form third parties can verify, to be implemented as concrete technology and to improve safety under specific conditions.