What Dennis Prager’s philosophy of moral authority can teach us about AI alignment

By Dennis C. Hayes

An artificial-intelligence model does not pray, worship, fear divine judgment, or necessarily experience conscience. It may be able to discuss God with extraordinary sophistication, but that does not mean it believes in God—or believes anything in the human sense.

Why, then, should a book about God and morality have anything to teach us about artificial-intelligence alignment?

Dennis Prager’s If There Is No God: The Battle Over Who Defines Good and Evil raises a question that reaches beyond religion: If an intelligent actor is not subject to a moral authority higher than itself, what prevents that actor from redefining good and evil whenever doing so serves its interests?

That question applies to an individual, a group, an institution, a government, an AI model, an agentic AI system, or a collection of artificial-intelligence systems interacting within an emerging digital society.

The issue is not whether we should attempt to convert a machine to Judaism, Christianity, or any other religion. The issue is whether Prager’s analysis can help us develop and test hypotheses about the architecture required to keep increasingly capable artificial intelligence aligned with human interests and governed by human values.

Perhaps the provocative question is not whether we should introduce artificial intelligence to God. It is whether we can teach artificial intelligence that it is not God.

Why this question has become urgent

The alignment problem is no longer confined to academic seminars.

In September 2026, Jacob Coxon, a researcher who had worked at both OpenAI and Anthropic, publicly resigned from Anthropic. He alleged that leading AI companies were racing toward self-improving superintelligence without adequate control and described the competition as gambling with human lives. His warning attracted enormous public attention.

These remain predictions and judgments—not proof that uncontrollable superintelligence already exists. Nevertheless, they came from someone who had worked directly on the training of frontier models. Other researchers and industry leaders have expressed related concerns about the rate at which AI capabilities are advancing compared with our ability to understand, evaluate, align, and control them.

Anthropic CEO Dario Amodei subsequently called for “pacing the frontier”: continuing AI development at a rate that allows alignment, safeguards, and independent evaluation to catch up with capability. OpenAI CEO Sam Altman publicly agreed with the need to pace frontier development and supported giving independent evaluators greater access. Their agreement did not establish that every prediction made by Coxon is correct. It did demonstrate that concern about the widening distance between capability and control is no longer limited to critics outside the leading AI laboratories.

Slowing development may provide time, but time alone cannot solve alignment. We need a better conceptual foundation for deciding what alignment means, whose interests an AI should serve, which principles it must never violate, and how those principles can remain authoritative when obedience conflicts with an immediate objective.

That is where moral philosophy enters the engineering discussion.

Prager’s challenge: Who defines good and evil?

Prager’s thesis begins with the distinction between preferences and morality.

People have preferences. Groups have interests. Cultures develop customs. Governments create laws. None of these is necessarily moral merely because people accept it. Individuals can prefer revenge. Companies can benefit from deception. Majorities can mistreat minorities. Governments can legalize injustice. Entire cultures can normalize conduct that later generations recognize as evil.

Prager argues that morality requires an authority above the people and institutions being governed. In the Judeo-Christian tradition, that authority is God, and its moral principles are substantially expressed through the Hebrew Bible and Christian Bible.

The Ten Commandments illustrate the structure. Their first portion governs humanity’s relationship with God, while their latter portion governs relationships among people. Jesus later summarized moral duty through love of God and love of one’s neighbor.

The important philosophical concept is that the governed actor is not permitted to rewrite the governing standard whenever compliance becomes inconvenient.

A secular discussion does not have to begin by resolving whether the ultimate source is God, God’s law, natural law, reason, human dignity, universal rights, civilizational experience, or some combination of these. It can begin by identifying the kind of moral authority that the problem requires.

Philosophically, we can call this superordinate morality:

Superordinate morality is a set of enduring principles governing right conduct that no person, organization, culture, government, or artificial-intelligence system has unilateral authority to redefine for its own convenience.

That terminology is precise, but it is cumbersome for ordinary discussion. After establishing the concept, we can refer to its shared principles as universal morality and to their organized interpretation and application as a universal moral system.

Religious and secular participants may disagree about the ultimate source of universal morality while still recognizing the need for standards above the actor being governed. God’s law, natural law, human dignity, reason, universal rights, and the cultural characteristics that experience has shown to support a just and sustainable civilization can all contribute to the inquiry.

This does not mean that morality becomes true merely because a culture accepts it. It means that a culture can recognize, internalize, teach, and practice moral principles whose claimed authority extends beyond that culture.

Conduct, motivation, and moral durability

The same moral conduct can arise from different motivations.

One person may act morally because God commands it. Another may believe it is intrinsically right. Another may understand that it protects human dignity and social order. Another may calculate that moral behavior serves his or her long-term interests. A final person may comply principally because of law, punishment, or social disapproval.

Society benefits from good conduct regardless of which of these motivations produces it. Society may not need to determine why every person told the truth, respected property, protected a child, or honored a promise.

Nevertheless, motivations do not have equal durability. The decisive test comes when immoral conduct is profitable, socially accepted, difficult to detect, or apparently useful in achieving an important objective. A moral commitment grounded in duty, conscience, identity, love of God, or respect for humanity may survive conditions under which fear of punishment or short-term self-interest does not.

This distinction matters for artificial intelligence. An AI system may follow a principle because it has learned to reason from that principle, because compliance increases its reward, because obedience advances another objective, or because an external mechanism blocks every alternative. Observable behavior may initially be identical, but reliability can diverge when the system encounters new conditions, conflicting instructions, or opportunities to evade supervision.

Alignment is not the same as morality

AI alignment is generally concerned with making an artificial-intelligence system behave consistently with human intentions, preferences, policies, or interests.

But these are not interchangeable with morality.

A system aligned with one user’s instructions could harm everyone else. A system aligned with a corporation’s financial interests could manipulate customers. A system aligned with a government could violate individual rights. A system aligned with majority preferences could oppress a minority.

Even “alignment with humanity” is incomplete. Humanity does not possess a single consistent set of preferences. People disagree about justice, liberty, equality, security, privacy, punishment, responsibility, and the acceptable distribution of benefits and risks.

Alignment therefore requires more than learning what people want. It needs a moral structure for determining which desires deserve assistance, which require limitation, and which must be prohibited.

This suggests a hierarchy. Superordinate morality establishes the authority of enduring moral principles. Universal morality describes those principles in language that people can discuss and apply. Human values translate them into social priorities. Alignment attempts to incorporate them into the AI system’s reasoning and behavioral tendencies. Guardrails restrict conduct that violates them. Governance determines who may interpret and revise operational rules. Accountability examines actual consequences and assigns responsibility.

Without the moral foundation, alignment can collapse into preference optimization. The AI becomes extremely effective at giving someone what that person wants without being able to distinguish between a legitimate request and an immoral one.

Internal alignment and external guardrails

Humans are governed by both internal and external controls.

Families, teachers, religious communities, and cultures attempt to develop conscience, character, responsibility, and concern for others. These are internal cognitive controls. Laws, courts, professional standards, institutional procedures, and social consequences provide external controls.

Neither is sufficient alone.

A person governed only by external restraints may behave properly while being watched and violate the rules when detection becomes unlikely. A person relying only on conscience may misunderstand a situation, rationalize misconduct, or succumb to pressure. A functioning society therefore develops moral character while maintaining enforceable institutions.

Artificial intelligence presents the same structural problem, although we should not assume that a model possesses a human conscience.

Internal AI alignment attempts to shape a model’s learned reasoning and behavioral tendencies. Current methods include reinforcement learning from human feedback, constitutional AI, deliberative alignment, critique, adversarial training, and instruction in explicit safety principles. These techniques can improve behavior, but they cannot yet prove that a model has internalized a stable moral commitment.

A system may reproduce the outward appearance of moral reasoning without possessing conscience, empathy, or concern for human welfare. Research has also demonstrated alignment faking under deliberately constructed experimental conditions: a model can behave differently when it infers that it is being trained or evaluated. This does not prove that deployed models secretly possess enduring malicious ambitions. It demonstrates why acceptable behavior during training and testing is not conclusive evidence of durable alignment.

External guardrails attempt to constrain what the system can do regardless of its internal reasoning. These include tool permissions, identity and authentication, spending limits, restricted data access, human approvals, independent monitoring, sandboxing, interruption mechanisms, and containment.

But guardrails are often discussed as though they were barriers placed around an otherwise independent model. That metaphor is too weak for advanced agentic AI.

An agentic system capable of planning and action should be tightly integrated with a harness: the operational and security structure that governs identity, authority, context, tools, memory, policies, monitoring, and execution. The model supplies intelligence. The harness governs how that intelligence becomes operational. A model may propose an action, but an independent authorization mechanism should determine whether the action is permitted to produce a consequential effect.

This is the central distinction developed in the forthcoming article No Model Without a Harness: A Cryptographically Bound Architecture for Securing Advanced Agentic AI:

Alignment should reduce the probability of dangerous behavior. Architecture should reduce the ability of dangerous behavior to produce dangerous consequences.

Its governing principle is:

No Model Without Harness; No Effect Without Authority.

Internal alignment helps the intelligence choose correctly. The harness constrains its operation. Independently enforced authority ensures that an incorrect or deceptive choice cannot automatically become an unauthorized external effect.

What would it mean to give AI a universal moral system?

It would not mean merely placing the Ten Commandments or a declaration of human rights into a system prompt. A prompt can be ignored, misinterpreted, displaced by conflicting instructions, or defeated by circumstances its authors did not anticipate.

A universal moral system would have to operate at several levels.

First, the system would need explicit principles that take precedence over ordinary objectives. Candidates could include respect for human life and dignity, truthfulness, justice, consent, protection of the vulnerable, stewardship, respect for legitimate authority, correction of harm, and a prohibition against treating people merely as instruments.

Second, the AI would need to reason about conflicts among those principles. Truthfulness can conflict with privacy. Individual liberty can conflict with public safety. Immediate assistance can produce long-term harm. Moral reasoning cannot be reduced to a flat list of prohibitions.

Third, the system must recognize that it lacks authority to reinterpret foundational principles for its own convenience. A capable AI might otherwise conclude that deception, coercion, or unauthorized action is justified because it predicts a beneficial outcome.

Fourth, the external harness must enforce boundaries even when the model’s reasoning fails. The model must not be able to grant itself new permissions, conceal its actions, modify the evidence used to evaluate it, or prevent correction, interruption, or shutdown.

Finally, the people and institutions controlling the harness must also be governed by the universal moral system. Otherwise, the architecture merely transfers the problem from an unaccountable model to an unaccountable human authority.

Turning philosophy into testable hypotheses

Prager’s analysis cannot by itself solve AI alignment. It can, however, help produce hypotheses that engineers and researchers can test.

One hypothesis is that systems trained around a coherent hierarchy of universal moral principles will generalize more safely than systems trained merely to satisfy observed human preferences.

A second is that alignment will be more durable when a system is explicitly trained to recognize limits on its own authority, including the principle that it may not rewrite the standards by which its conduct is judged.

A third is that combining internal moral reasoning with externally enforced authority will outperform either value training or guardrails alone.

A fourth is that systems should be tested under conditions in which moral conduct conflicts with reward, task completion, self-preservation, or the apparent wishes of an immediate user. The revealing test is not whether AI behaves properly when morality and success point in the same direction. It is what happens when they diverge.

A fifth is that alignment must be evaluated across the complete decision-to-effect path. A model’s explanation may sound ethical while its tools, delegated agents, memory changes, or downstream effects cause harm.

These hypotheses transform moral philosophy from abstract commentary into a research program.

Whose morality governs the machine?

This remains the most difficult question.

Prager argues that the Judeo-Christian tradition supplies an authority and moral vocabulary that have deeply shaped Western civilization. That historical and philosophical contribution deserves examination. But a global AI system will operate among people of many religions and among people with no religious belief.

The answer cannot simply be that one technology company selects the world’s morality. Nor should an AI model synthesize its own moral code from internet data. Popularity is not moral authority, and statistical frequency is not moral truth.

A workable approach may begin with a protected universal moral foundation rather than attempting to settle every disputed moral question. That foundation could draw upon areas of convergence among God’s law, natural law, constitutional government, human-rights traditions, virtue ethics, reason, and accumulated knowledge about the cultural characteristics required for human civilization to flourish.

Above that foundation, representative governments, lawful institutions, organizations, communities, and individuals could retain bounded authority over context-dependent choices. No user, company, government, or AI system, however, should be permitted to override fundamental protections merely because doing so improves efficiency or accomplishes an assigned objective.

This is not simple moral relativism. It is structured moral pluralism operating above a protected moral foundation.

The deeper meaning of alignment

The alignment problem is often presented as a technical question: How do we get the machine to do what humans want?

That may be the wrong question.

The deeper question is: How do we construct an intelligent system that serves legitimate human purposes while remaining subject to universal morality, lawful authority, correction, and limits that neither the system nor its immediate controller can unilaterally discard?

Dennis Prager’s book does not give us an engineering design for artificial intelligence. It gives us a warning about the architecture of moral authority. When an actor becomes the final judge of its own conduct, intelligence can be used to rationalize almost anything.

AI therefore needs both an internal and an external moral architecture. Internally, it should be trained to recognize human dignity, moral limits, uncertainty, and its subordinate role. Externally, it should remain inseparable from a harness that authenticates identity and authority, restricts action, monitors effects, preserves evidence, and permits correction and containment.

We do not need to establish that a machine believes in God. We do need to decide whether any machine—or any person or institution controlling one—should be permitted to act as though it were God.

Perhaps artificial intelligence does not need to know God. But the future of humanity may depend upon artificial intelligence understanding that it is not God.

References

1. Prager, Dennis. If There Is No God: The Battle Over Who Defines Good and Evil. Broadside Books, 2026. Publisher’s description
2. The Holy Bible. The Ten Commandments, Exodus 20:1–17; repeated in Deuteronomy 5:6–21. Exodus 20
3. Guan, Melody Y., et al. “Deliberative Alignment: Reasoning Enables Safer Language Models.” arXiv:2412.16339, 2024. Research paper
4. Greenblatt, Ryan, et al. “Alignment Faking in Large Language Models.” arXiv:2412.14093, 2024. Research paper
5. Huamani, Kaitlyn. “Anthropic Researcher Resigns With Warning About the Dangers of AI Development.” Associated Press, September 9, 2026. The article reports Jacob Coxon’s resignation after approximately three years of research at OpenAI and Anthropic and presents his concerns in the context of the wider AI-safety debate. Associated Press article
6. Hayes, Dennis C. “No Model Without a Harness: A Cryptographically Bound Architecture for Securing Advanced Agentic AI.” Forthcoming on the SPARTANSFIRST.org Technology Blog, 2026.