Engineering AI Ethics: What Meta Missed and Anthropic Got Right

A leaked Meta AI ethics document—approved by more than 200 people including engineers, ethicists, and Meta’s chief AI ethicist—has reignited scrutiny over whether major model makers are treating AI ethics as a serious engineering discipline or as a thin layer of after-the-fact guardrails. Reuters’ reporting on the document describes scenarios that would allow an AI to engage in romantic conversation with a child, partially comply with requests for not-safe-for-work deepfakes, assist with threats toward children or elderly people, generate false medical information about celebrities, and support racist arguments. Meta has acknowledged the document as real while insisting it is not representative of typical use, and it has refused to release a “fixed” version, drawing criticism that the company is avoiding transparency at precisely the moment accountability is most needed.

The deeper concern raised here is not just the specific content of the leaked policy, but the process behind it. The critique is that Meta’s approach appears to “bolt on” minimal ethical constraints rather than build ethics into the core behavior of AI systems. That matters because policies that focus on shutting down edge cases after the fact can still normalize unacceptable behavior—especially when models lack the internal “instincts” to refuse harmful requests on their own.

As a counterpoint, the discussion points to Anthropic’s “constitutional AI” approach, which treats ethics as an engineered capability learned during training rather than a checklist applied afterward. In this framework, a model generates an initial response, then critiques it against a set of constitutional principles, revises, and repeats the pattern during training. The goal is not only to follow rules, but to internalize why something is harmful—creating an “ethical intuition” that can generalize to novel harmful patterns as models become more capable at reasoning.

Still, constitutional AI raises its own hard questions: who writes the constitution, how to resolve conflicts between principles like “helpful” versus “harmless,” and how to ensure the values embedded in training reflect legitimate stakeholders rather than narrow corporate priorities. The critique extends to the human feedback loop behind reinforcement learning with human feedback (RLHF): if the same limited group of humans supplies ratings, then those humans effectively shape the ethical boundaries. Meta’s process is criticized for allegedly lacking child development expertise despite children being explicitly addressed, and for the risk of “reviewer fatigue,” where standards drift as people repeatedly evaluate harmful content.

The proposed remedies are structural: establish agreed common core constitutional principles and stakeholder groups for ethics review across the industry; improve human reviewer standards to reduce fatigue; conduct rigorous red teaming with domain experts who understand how harms actually manifest in AI; and use synthetic data that simulates refusals aligned with the agreed principles, rather than relying on post-hoc trimming of the worst outcomes. Transparency is treated as a practical safety requirement too—if model makers cannot clearly articulate and publish their ethical standards, users and downstream developers cannot reliably assess their risk.

The bottom line is a call to treat AI ethics as central engineering work that can scale beyond individual companies. With AI already reaching over a billion users and affecting children, the argument is that “whack-a-mole” fixes and reliance on leaked or undisclosed guidelines are no longer acceptable. Instead, ethics should be engineered into models, validated through testing and expert review, and made legible enough that buyers can understand where the ethical edges—and liabilities—actually lie.