When people talk about "AI safety," they often mean reliability, jailbreak resistance, or preventing obviously harmful outputs. But at the frontier, safety is unavoidably moral: it's about what an AI should do when values conflict—truth versus kindness, user autonomy versus harm prevention, and individual freedom versus societal risk. Anthropic and OpenAI both take this moral dimension seriously, but they operationalize it through noticeably different "moral architectures": how values get written down, how they're trained into models, and how risk gates constrain deployment.
Two moral artifacts: "constitutions" and "specs"
Anthropic increasingly treats alignment as a kind of character formation. Its public "Claude's Constitution" is positioned as the "final authority" on Claude's values and behavior, explicitly framed in human moral language—virtue, wisdom, moral uncertainty—and written primarily for the model as the intended reader. It also makes a clear ordering of priorities: broadly safe, then broadly ethical, then compliant with Anthropic's guidelines, then genuinely helpful. In other words: don't undermine human oversight and avoid dangerous behavior even when that constrains helpfulness.
That moral stance is tied to Anthropic's technical lineage of Constitutional AI—training methods where the model learns to critique and revise its own outputs according to a set of principles, reducing reliance on direct human preference labeling for every edge case. The aspiration is that a model internalizes general moral reasoning (and not just a list of "don'ts"), so it can generalize better to novel situations.
OpenAI, by contrast, presents alignment more like a governance document for behavior: the Model Spec. It is explicitly built around a chain of command and levels of authority (Root/System/Developer/User/Guideline), designed to make trade-offs legible and enforceable. Morality in this framing is less "becoming virtuous" and more "obeying a structured hierarchy of constraints," with the goal of predictable outcomes across a huge range of deployments.
Importantly, OpenAI's latest Model Spec makes its moral commitments unusually explicit in "red-line principles": it prioritizes human safety and human rights, prohibits use cases involving severe harms (e.g., mass violence, WMD-related assistance, child abuse, persecution or mass surveillance), and emphasizes human control—rejecting uses that undermine autonomy or civic participation. It also frames privacy as a core commitment.
Rules versus judgment: different "moral philosophies" in practice
Anthropic's constitution argues openly about a classic tension: rules are predictable and testable, but can fail in unforeseen contexts; judgment is flexible and context-aware, but can be harder to audit. Anthropic says it generally prefers cultivating "good values and judgment" over rigid procedures—while still keeping some hard constraints for high-stakes domains (for example, strong limits around bioweapons-related uplift).
OpenAI's Model Spec expresses a different balance: it tries to preserve user freedom and intellectual exploration within firm boundaries. The spec explicitly aims to "maximize helpfulness and freedom" while "minimizing harm," and it clarifies that only a limited set of high-authority rules should be non-overridable—reflecting a "minimal necessary constraint" philosophy for a foundational technology.
Neither approach eliminates moral disagreement. Anthropic makes values more "human-readable" and invites critique of the constitution itself; OpenAI makes values more "system-operational" and invites critique of how the chain of command resolves conflicts. The trade-off is between moral transparency (values described in philosophical terms) and behavioral determinism (values expressed as enforceable instruction hierarchies).
Scaling safety: Responsible Scaling Policy vs Preparedness Framework
At frontier capability levels, "morality" also includes decisions about when not to ship and how to avoid enabling catastrophic misuse. Here, each company has a distinct safety gating framework.
Anthropic: Responsible Scaling Policy (RSP)
Anthropic's RSP v3.0 explicitly reframes safety as a collective action problem: if one company slows down unilaterally, another may push ahead with weaker protections, potentially increasing ecosystem risk. As a result, Anthropic now separates (1) what it expects to do as a company from (2) more ambitious industry-wide recommendations—stating it "cannot commit to following them unilaterally."
To compensate, RSP v3.0 introduces more public-facing accountability mechanisms: Frontier Safety Roadmaps (published in redacted form) and recurring Risk Reports that discuss threat models, mitigations, and whether risks are justified by benefits. The policy states Risk Reports will be published every 3–6 months, and it anticipates some external review.
OpenAI: Preparedness Framework
OpenAI's Preparedness Framework (v2) focuses on identifying and tracking frontier capabilities most associated with "severe harm," explicitly naming tracked categories like biological/chemical, cybersecurity, and AI self-improvement capabilities. It defines "severe harm" at a high bar (e.g., mass casualty scale or massive economic damage), and it ties deployment to safeguards that "sufficiently minimize" risk.
Operationally, the framework describes governance: a cross-functional Safety Advisory Group (SAG) makes recommendations about required safeguards; leadership can approve or reject; and the Board's Safety and Security Committee provides oversight of those decisions.
What these differences add up to
Both labs increasingly converge on a shared premise: powerful AI requires explicit normative commitments plus measurable, gated risk management. The differences are in emphasis:
- •Anthropic's center of gravity is moral internalization (constitution-shaped behavior and "judgment"), paired with a scaling policy that increasingly treats safety as an ecosystem-wide coordination challenge.
- •OpenAI's center of gravity is rule-governed behavior and institutional process (Model Spec hierarchy + preparedness gating), paired with board-level oversight structures intended to hold releases to documented safety thresholds.
A fair critique of both approaches is that "morality" is still being specified by a relatively small set of actors inside private organizations—then translated into global-scale infrastructure. A fair defense is that publishing constitutions/specs and safety frameworks is a real step toward scrutiny, standard-setting, and external pressure—especially compared with opaque alignment-by-default. In practice, building AI safely likely requires borrowing from both philosophies: explicit moral commitments people can debate, and auditable thresholds institutions can enforce.