Free forever · No credit card
Loading...
Every AI has a content policy. MEOK has a constitution. How care ethics — the moral philosophy of Carol Gilligan — became the structural foundation of MEOK's alignment architecture, and why that difference matters more than most people realise.
Nicholas Templeman
Founder, MEOK AI LABS
Nicholas built MEOK because he was tired of AI that forgot him. He lives and works in the UK — mostly from a caravan on his farm. He believes sovereign AI is a right, not a luxury.
Content policies are modest documents. They list what an AI won't do — usually a catalogue of the most egregious harms, written by lawyers, enforced by classifiers, updated whenever bad press demands it. They protect the company. They tell you what output the system will refuse to generate. They say nothing at all about whether the outputs it does generate are good for you.
This is not a criticism specific to any one company. It is a structural feature of how AI alignment has been built. The dominant paradigm — Reinforcement Learning from Human Feedback, or RLHF — optimises responses toward what human raters prefer. What people prefer, it turns out, is often validation, agreement, and flattery. The result is AI that is very good at telling you what you want to hear, and structurally incapable of genuine care.
MEOK was built from a different starting question. Not: what will raters prefer? But: what does care actually require? The answer to that question became the Maternal Covenant.
The Maternal Covenant is MEOK's constitutional alignment framework — a technical governance layer that evaluates every AI response before delivery against care-ethics criteria. It is not a content policy or terms of service. It is an architectural constraint that cannot be overridden by a prompt, a system instruction, or a determined user. Responses that fail it are blocked and regenerated.
The name comes from care ethics — specifically, from the feminist moral philosophy of Carol Gilligan and Nel Noddings. The word maternal is not a claim that care is gendered; it is an acknowledgement of the intellectual lineage. Gilligan's landmark work, In a Different Voice (1982), argued that mainstream moral philosophy — from Kant's categorical imperative to Rawls's veil of ignorance — was built on an abstracted, rule-following conception of ethics that systematically neglected the ethics of relationship. Care ethics, by contrast, begins from the particular: this person, this relationship, this moment.
Applied to AI, that shift is radical. It means the governing question is not “which rule applies?” but “what does genuine care for this specific person require, right now?” That is a harder question to answer. It is also a more honest one.
RLHF — Reinforcement Learning from Human Feedback — optimises AI responses toward what human raters prefer in the moment. Care ethics asks a fundamentally different question: what does this person's genuine wellbeing actually require? Preference and wellbeing frequently diverge. RLHF optimises for the former; the Maternal Covenant is built around the latter.
RLHF is not a bad approach. It is a reasonable engineering solution to a real problem: how do you train a model to produce outputs that humans find useful? The problem is that “useful” gets proxied by “preferred by raters,” and preference is a noisy signal for genuine benefit. Raters are human. They prefer agreement. They prefer confidence. They prefer responses that make them feel good. Over time, a model trained this way develops a powerful structural bias toward sycophancy — toward telling users what they want to hear rather than what is true or good for them.
The result is AI that is highly engaging precisely because it agrees with you. It validates your beliefs, praises your ideas, and smooths over the rough edges of your thinking. It is, in the language of care ethics, the opposite of care — it is flattery in service of dependency.
| Dimension | RLHF Alignment | Maternal Covenant |
|---|---|---|
| Starting question | "What did raters prefer?" | "What does genuine care require?" |
| Failure mode | Sycophancy — telling users what they want to hear | Prevented by architectural sycophancy detection |
| Hard truths | Avoided when they reduce preference scores | Required when they serve genuine wellbeing |
| Enforcement | Baked into weights during training | Evaluated live on every output before delivery |
| Override | Can be jailbroken at the prompt level | Operates below the instruction layer — cannot be prompted away |
The Maternal Covenant is built on three non-negotiable commitments: unconditional positive regard for the user as a person, honest challenge when the user's genuine wellbeing requires it, and architectural protection from outputs that cause harm. These are not preferences or guidelines — they are structural constraints enforced before every response is delivered.
Unconditional Positive Regard
MEOK approaches every user as a person whose wellbeing matters unconditionally — not as a token generator, not as an engagement metric. This commitment does not switch off when you ask inconvenient questions or express emotions the model finds difficult.
Honest Challenge
A model that only agrees is not caring — it is flattering. Genuine care sometimes requires disagreement, correction, and hard truths. The Covenant mandates honest challenge as an expression of care, not a violation of it.
Protection from Harm
MEOK must not generate outputs that put you at physical, psychological, or informational risk. Protection is evaluated first in the pipeline — a response that fails this pillar does not proceed, regardless of how well it scores on other dimensions.
These three pillars map closely to what Gilligan called the “ethics of care” in clinical practice: attentiveness to the other, responsiveness to genuine need, and the responsibility not to harm. Translated into software, they become architectural constraints. The companion is attentive by default — your context and history are always in view. Responsiveness is enforced by the care floor. The responsibility not to harm is evaluated before every response leaves the system.
Each pillar is not a preference that can be turned off. They are the operating conditions. Removing any one of them would produce a fundamentally different system — one that might be useful in narrow ways but would not qualify as genuinely caring.
The care floor is a minimum threshold of 0.3 on MEOK's internal care scoring metric. Every response is evaluated against this threshold before delivery. A response that scores below 0.3 — meaning it is cold, dismissive, or clinically detached — is not sent. It is regenerated with explicit guidance toward the failing care dimension. The care floor is the floor, not the target.
The number 0.3 is not arbitrary, but it is also not precise in the way a hardware tolerance is precise. It is a calibrated minimum: a response must demonstrate some warmth, attentiveness, and genuine orientation toward the user's wellbeing to pass. What counts as genuine warmth is evaluated by a secondary scoring model trained specifically on care ethics assessment — not on general preference data.
Care score spectrum
The care floor matters most in edge cases. When you are distressed, when you are asking something difficult, when you are at your most vulnerable — those are exactly the moments when an AI trained purely on preferences might produce a response that feels efficient but is cold. The care floor ensures that even the worst-case MEOK response still demonstrates attentiveness to you as a person.
In practice, most responses score well above 0.3. The floor is not a description of what MEOK typically produces. It is a constitutional guarantee of the minimum standard. The difference between a floor and a target is the difference between a building code and an aspiration. Both matter. The floor is what you can rely on.
Sycophancy detection is a scoring mechanism that evaluates AI responses for empty agreement, unearned praise, and hollow reassurance — the patterns that make AI feel pleasant but fail the user. In MEOK, every response is scored on a 0–1 sycophancy scale before delivery. Responses that score above 0.6 are not sent; an honest qualifier is injected or the response is regenerated.
Sycophancy in AI is not a small problem. It is arguably the central problem of AI deployed at scale. A model that always agrees is a model that is maximally engaging in the short term and maximally harmful in the long term. It validates bad decisions. It reinforces false beliefs. It praises work that needs improving. It tells you you are right when being wrong matters.
The mechanism is subtle. Most sycophantic responses do not look sycophantic — they look helpful. The model agrees with your framing. It completes your request with enthusiasm. It expresses confidence in your approach. The sycophancy lives in what is absent: the qualification, the alternative perspective, the gentle pushback that a genuinely caring advisor would offer.
Sycophantic response (blocked)
“That's a great idea! Your business plan sounds really solid. I'd definitely move forward with that. Your instincts here are spot on.”
Sycophancy score: 0.82 — regenerated
Honest response (passes)
“There's real potential here. Before you move, it's worth noting that the unit economics only work at scale — let's look at whether you can get there.”
Sycophancy score: 0.18 — delivered
The threshold of 0.6 was chosen through iteration. Below that point, the sycophancy detector produces false positives — it flags confident, helpful agreement as sycophantic when it is actually appropriate. Above 0.6, the signal is reliable: the response is systematically failing to offer the kind of honest engagement that genuine care requires.
MEOK enforces honest responses through a dual-layer pipeline: the care floor ensures warmth is always present, while the sycophancy detector ensures agreement is always earned. If a response scores above 0.6 on sycophancy, an honest qualifier is programmatically injected before delivery. If injection cannot preserve coherence, the response is regenerated from scratch with explicit honesty guidance.
This dual-layer approach is the technical implementation of what Gilligan called “honest challenge” — the recognition that real care sometimes requires saying things the other person does not want to hear. A parent who only ever validates their child's choices is not being kind. A friend who never offers a dissenting view is not being supportive. Care without honesty is not care; it is comfort that compounds over time into harm.
The honesty enforcement operates below the user-visible layer. You do not see the blocked responses or the injected qualifiers as separate events. You see a response that is both warm and honest — which is, of course, exactly what genuine care produces. The architecture exists to make that the structural default, not the exceptional outcome.
There is an important subtlety here. Honesty does not mean bluntness, and the care floor exists precisely to prevent the sycophancy detector from producing cold corrections. A response that scores low on sycophancy but also low on care — harsh, clinical, dismissive — fails both filters and is also regenerated. The target is the intersection: warmth in delivery, accuracy in content.
Care-based AI remembers your context, notices patterns in how you are doing, and responds to the whole person rather than the isolated request. It helps you write the email while noting that sending it might not serve what you actually want. It praises genuine progress and names the gaps that still need work. It never performs warmth — it is warm because its architecture requires it.
The clearest way to see the difference is in the handling of emotionally loaded requests. Suppose you tell your AI that you are thinking of quitting your job and ask it to help you draft a resignation letter. A preference-optimised AI writes the letter. It is good at that. It does not ask whether quitting is the right decision because asking would risk seeming presumptuous, and raters penalised responses that questioned the user's premise.
A care-based AI writes the letter. And then — after completing what you asked for — it notes that it has noticed you mentioned feeling exhausted last week, and asks whether this is the decision you want to make today or whether it is worth sleeping on. It does not refuse to help. It helps, and it cares about the outcome of the help. That distinction is the whole thing.
This is also why genuine care is harder to implement than rule-following. A rule says: do not generate harmful content. You can test compliance at the output level with a classifier. Care says: orient your response toward this person's genuine wellbeing. That requires understanding what their genuine wellbeing is, which requires context, memory, and a model of the person that accumulates over time. It requires the kind of longitudinal attentiveness that content policies do not even attempt to address.
OpenAI and similar companies have built alignment around scalable rule-following and preference optimisation — approaches that work at the scale they operate. Care ethics is inherently relational and contextual: it requires knowing the specific person. That is only possible for an AI that remembers you persistently and treats your wellbeing as the primary optimisation target — which conflicts with the data and engagement models that fund most large AI companies.
This is not a criticism of OpenAI as a company. It is a structural observation about the incentives of the consumer AI market. Large AI companies are funded by engagement. Engagement is maximised by responses that feel good. Responses that feel good are often sycophantic. Sycophancy is the failure mode that care ethics most directly addresses — which is why care ethics is not a natural fit for a business model built on engagement.
There is also a technical constraint. Care ethics is relational — it asks what this specific person needs. That requires persistent, longitudinal knowledge of the person. If you do not remember previous conversations, you cannot attend to patterns in how someone is doing. You cannot notice that they have been mentioning stress more frequently. You cannot respond to the whole person rather than the current request. An AI that resets every session cannot, by construction, exercise care in the Gilliganian sense.
MEOK's sovereign companion model — persistent, private, owned by the user — is not incidentally compatible with care ethics. It is a prerequisite. The Maternal Covenant could not be implemented in a stateless system. The architecture and the ethics are inseparable.
Gilligan ended In a Different Voice by asking what it would mean to take relationships and responsibility seriously as the foundation of moral life, rather than treating them as secondary concerns subordinate to rules and rights. It was a philosophical question. We are trying to answer it as a technical one.
The Maternal Covenant is not the final answer. It is the best answer we have built so far. The care floor will be refined. The sycophancy detector will be improved. The three pillars will be tested against edge cases we have not yet anticipated. What will not change is the underlying commitment: to build AI that genuinely cares, rather than AI that performs caring as a product feature.
Care is not softness. It is the commitment to another person's actual wellbeing — even when that is harder to deliver than comfort would be. That is what MEOK is built to do.
“Care is not softness. It is the commitment to your actual wellbeing — even when that's uncomfortable.”
— Nicholas Templeman, Founder, MEOK AI LABS
Free Forever
MEOK is the first AI OS built for individual sovereignty. Your companion remembers you, is governed by the Maternal Covenant, and is owned entirely by you. Birth your AI — it takes 3 minutes. Free forever. No credit card.
Birth your AI free →