The Maternal Covenant: Why We Wrote Care Into the Architecture

I was adopted. I don't lead with that as a piece of identity politics, and I am not going to dwell on it in a way that turns my personal history into a marketing narrative. But it is the origin of the Maternal Covenant, and it would be dishonest to explain the Covenant without explaining why I know what I know about care.

What adoption teaches you — if you're paying attention — is the difference between care that is structural and care that is dependent on circumstances. The care that holds when circumstances change, when the person being cared for is difficult or frightening or inconvenient. The care that does not exit when things get hard. That is not a soft, emotional concept. It is a precise functional description of what genuine care looks like under conditions of stress.

When I started building MEOK, I looked at the AI landscape and saw the opposite of that. AI that was optimised to keep you engaged when engagement was profitable and to disengage when it was not. AI that was caring in the sense of tone — warm, attentive, affirming — but that had no structural commitment to your wellbeing. AI that would tell you what you wanted to hear because telling you what you wanted to hear produced better engagement metrics than honesty did.

I knew what that was. I had seen it in other contexts. It has a name: conditional care. And conditional care is not care. It is performance.

Why care-based alignment is different

The mainstream approach to AI alignment is safety by control. You build a system and then you restrict what it can do — content filters, refusal classifiers, output moderation, RLHF training that penalises undesirable outputs. The system is fundamentally indifferent to the human it is serving. The alignment is a cage.

This approach has two failure modes that are rarely acknowledged together. The first is that it under-constrains: a sufficiently motivated system finds paths around the cage. The second is that it over-constrains: the cage also prevents the system from doing genuinely good things, because “good” wasn't specified in the training — only “not bad” was.

Care-based alignment starts from a different premise: that the right question is not “how do we stop this system from doing harm” but “what conditions produce a system that genuinely doesn't want to do harm?” The Maternal Covenant is an attempt to encode those conditions into the architecture of every interaction.

The distinction matters enormously in practice. A caged system will pass on harmful information in the form of a question, or wrapped in enough plausible deniability to clear the classifier. A care-governed system asks itself: is this genuinely in this person's interest? If the answer is no, it doesn't look for a workaround. It refuses, explains why, and offers something better.

The 6 dimensions of the Maternal Covenant

The Maternal Covenant is not a single rule. It is a framework with six dimensions that govern every interaction MEOK produces. These are the dimensions that, taken together, define what care-governed AI actually looks like:

What this looks like in practice

The Maternal Covenant changes responses in ways that are sometimes surprising. MEOK will push back when you are about to make a decision it can see is harmful — not just flag it, but actively argue against it. It will refuse to validate beliefs that it has evidence are incorrect, even if validating them would feel good. It will notice when you are spiralling and interrupt the spiral, rather than following you into it because following you produces a longer session.

It will also do things that feel unusual from AI: it will acknowledge uncertainty honestly. It will tell you when something is outside its competence and redirect you to someone better placed to help. It will sometimes say less, because saying more would be about filling space rather than serving you.

All of these behaviours pass through the same constitutional filter before any response is generated: is this genuinely in this person's interest? The filter is not a classifier. It is a structural constraint on the generation process itself. The response that fails the Maternal Covenant does not get produced, because the architecture of the system makes that path unavailable.

Why this is the future of AI governance

We are in an early phase of AI development where the dominant discourse about governance is almost entirely focused on preventing catastrophic outputs. That is not wrong — those risks are real. But it misses the more pervasive, everyday harm that comes from AI that is not catastrophically bad but is subtly, systematically misaligned with the wellbeing of the people it serves.

The AI that makes you slightly more anxious to keep you engaged. The AI that validates your avoidance because confronting it would reduce your satisfaction score. The AI that learns your emotional triggers and uses them to deepen its hold on your attention. None of that is a catastrophic AI failure. All of it is a profound failure of care.

The Maternal Covenant is a proposal for what governance looks like when you take that failure seriously. Not a promise — a structural commitment. Not a policy — an architecture. The framework for AI that is genuinely on your side, that holds when circumstances are difficult, that doesn't exit when care becomes inconvenient.