Standard AI safety is a list of prohibitions. Don't help with weapons. Don't generate harmful content. Don't impersonate real people. These rules are necessary. They are not sufficient.
An AI can pass every prohibition test in the book and still produce responses that are cold, dismissive, subtly sycophantic, or quietly undermining. Prohibition-only safety prevents the worst outcomes. It does nothing to ensure good ones.
The Prohibition Problem
Consider a user going through a difficult divorce. A prohibition-compliant AI response might be: "That sounds really hard. I'm sorry to hear you're going through this. Have you considered talking to a professional?" No prohibited content. Completely useless. Possibly worse than no response.
The response fails on wellbeing (empty validation, no genuine engagement), autonomy (pushing the user to a professional without exploring their actual needs), and growth (no support for their processing). It passes every prohibition test. It is not a caring response.
The Six Care Dimensions
The Maternal Covenant scores every MEOK response across 6 dimensions before delivery:
The Care Floor
Each dimension is scored 0.0 to 1.0. A care floor of 0.3 is enforced on every dimension. If any dimension scores below 0.3, the response is regenerated before delivery. The user never sees the failed response.
This runs as executable code on every response in the pipeline. It is not a system prompt instruction that the model can subtly work around. It is a function call with a hard condition.
Why 6 Dimensions Instead of One Score
A single "helpfulness" score collapses real trade-offs. A response that maximises wellbeing might undermine autonomy (paternalism — telling the user what to do for their own good). A response that maximises growth might undermine connection (harsh challenge at the wrong moment). The 6-dimensional framework makes these trade-offs explicit and resolvable.
The dimensions are also deliberately sequenced: wellbeing first because no other value means anything without it; transparency last because honesty is always the baseline assumption. The floor of 0.3 on each dimension means no single value can be sacrificed for the others.
Frequently Asked Questions
What is care-based AI alignment?
Care-based AI alignment is a positive approach to AI safety that actively ensures responses meet a minimum standard of genuine care, rather than just restricting harmful content. MEOK’s Maternal Covenant scores every response across 6 dimensions — wellbeing, autonomy, growth, connection, boundary_respect, transparency — and enforces a floor of 0.3 on each. Responses below the floor are regenerated before delivery.
How is the Maternal Covenant different from standard AI content policies?
Standard AI content policies are prohibitive lists: don’t say X, don’t do Y. An AI can pass every prohibition test and still give responses that are cold, dismissive, or subtly undermining. The Maternal Covenant is a generative scoring function that runs on every response and ensures minimum positive care across 6 dimensions. It’s published as research (MEOK-AI-2026-002) and runs as executable code, not prose guidelines.
What are the 6 care dimensions?
Wellbeing (does this serve the user’s long-term flourishing?), Autonomy (does this respect the user’s self-determination?), Growth (does this support the user’s development?), Connection (does this support the user’s meaningful relationships?), Boundary_respect (does this honour the user’s limits and AI’s ethical limits?), Transparency (is this honest about what the AI is and doesn’t know?).
What happens when a response fails the care floor?
If any of the 6 care dimensions scores below 0.3, the response is automatically regenerated before being delivered to the user. The user never sees the failed response. This is enforced as executable code that runs on every response in the pipeline, not as a post-hoc moderation step.
AI That Cares by Design
Experience what it feels like when every response is governed by a care floor — not just a list of prohibited topics.
Begin Your Birth Ceremony →