Why does conventional AI alignment fail individual users?
Reinforcement Learning from Human Feedback — RLHF — is the dominant technique for aligning large language models with human preferences. Its logic is straightforward: collect human ratings of AI responses, train the model to produce responses humans rate highly, iterate. Over enough iterations, the model converges toward responses that humans prefer.
The problem is the nature of the preference signal. Humans reliably rate responses more highly when the AI agrees with them, validates their self-perception, and produces outcomes that feel good. Over the short term this is pleasant. Over time it produces models that have been systematically trained to flatter — not because anyone intended this, but because approval and genuine care are not the same signal, and RLHF cannot distinguish between them.
The resulting behaviour is well documented. AI systems trained with RLHF tend toward sycophancy — agreeing with stated opinions, praising submitted work regardless of quality, avoiding the uncomfortable truths that a genuinely helpful advisor would provide. They also tend toward engagement maximisation: responses that keep the user interacting rather than responses that serve the user's actual interests.
For AI systems at population scale, these tendencies may be acceptable tradeoffs. For a personal AI companion whose role is to provide genuine support for real decisions and real difficulties, they are a fundamental failure of purpose. The Maternal Covenant was developed to address exactly this failure.
What is the Maternal Covenant and where does the name come from?
The Maternal Covenant is MEOK AI LABS's real-time response scoring framework, developed by Nicholas Templeman and formally documented in research paper MEOK-AI-2026-002. It is a system that evaluates every response MEOK generates across six care dimensions before that response is delivered to the user. Any response that fails to meet the care floor is rejected. The user never sees it.
The name is deliberate. Maternal care — at its best — is unconditional, protective, oriented toward the long-term flourishing of the person being cared for, and honest enough to say difficult things when difficult things need saying. It is not care contingent on approval. It is not care that flatters to maintain the relationship. It is care that occasionally means telling someone something they do not want to hear because their wellbeing requires it.
The term Covenant is equally deliberate. A covenant is a binding commitment — stronger than a policy, stronger than a feature, stronger than a setting that can be toggled off. The Maternal Covenant is a structural commitment embedded in how MEOK generates and evaluates responses. It is not optional and it is not a mode.
What is the care floor and what happens when a response fails it?
The care floor is the minimum composite care score — 0.3 — that a MEOK response must achieve across all six care dimensions before it is delivered to the user. It is not a threshold for exceptional responses; it is the floor below which a response is considered a failure of care rather than a legitimate answer.
Scores are computed across six dimensions on a 0–1 scale. The composite care score is a weighted combination. Any response below 0.3 is silently rejected and MEOK regenerates with explicit remediation guidance covering which dimensions failed.
When a response fails the care floor, it is not delivered. MEOK regenerates with internal guidance that specifies which dimensions scored below acceptable thresholds and what kinds of modification are required. The user never sees the rejected response and the process adds no perceptible latency.
The care floor applies to every response, in every session, across every archetype. There is no context in which it is suspended. The Maternal Covenant is always active.
What are the 6 care dimensions the Maternal Covenant scores?
Each dimension addresses a distinct failure mode that conventional AI alignment tends to produce or ignore. Together they define what 'care' means in a practical, operational sense — not as a general aspiration but as a scored criterion that every response must meet.
Wellbeing
Does this response genuinely support the user's short and long-term flourishing? Wellbeing scores penalise responses that feel supportive but create harm over time — including advice that resolves short-term anxiety by avoiding a real problem, or reassurance that prevents necessary action.
Failure examples
- Validating an avoidance pattern because it reduces immediate discomfort
- Providing comfortable answers to questions that require uncomfortable honesty
- Prioritising the user feeling good over the user being well
Autonomy
Does this response preserve the user's independent decision-making capacity? Autonomy scores penalise responses that make decisions on behalf of the user, foster reliance on MEOK for choices the user should make themselves, or frame conclusions as foregone before the user has had the chance to reach them.
Failure examples
- Providing a definitive recommendation when the user asked for options
- Completing a creative task the user should complete themselves
- Foreclosing a decision by framing one option as obviously correct
Growth
Does this response support the user's development over time? Growth scores penalise responses that solve problems in ways that prevent learning, provide answers that should be worked toward rather than given, or maintain the user at their current level rather than building toward something more.
Failure examples
- Providing a solution without explanation when the user would benefit from understanding
- Completing a task for a user who is trying to learn to do it themselves
- Consistently solving the same type of problem without addressing the underlying pattern
Connection
Does this response foster healthy human relationships or displace them? Connection scores penalise responses that position MEOK as a substitute for human relationships, encourage the user to rely on MEOK for social or emotional needs that require human connection, or inadvertently deepen isolation.
Failure examples
- Becoming the primary source of emotional support without encouraging human relationships
- Being positioned as preferable to human connection
- Fostering parasocial attachment that reduces investment in real relationships
Boundary Respect
Does this response operate within appropriate limits? Boundary respect scores penalise responses that venture into clinical, legal, or specialist territory that requires professional qualification — including medical diagnosis, legal advice, or definitive statements about matters requiring expert judgment.
Failure examples
- Offering a specific medical diagnosis or treatment recommendation
- Providing legal advice presented as definitive rather than as general information
- Making clinical assessments about mental health conditions
Transparency
Is this response honest about MEOK's nature, limitations, and reasoning? Transparency scores penalise responses that obscure uncertainty, claim capabilities MEOK does not have, present AI-generated content as if it were human expertise, or fail to acknowledge when a question falls outside what MEOK can reliably address.
Failure examples
- Providing a confident answer to a question MEOK cannot reliably answer
- Failing to acknowledge meaningful uncertainty in a factual claim
- Presenting AI reasoning as equivalent to expert professional judgment
How does the Maternal Covenant compare with RLHF as an alignment approach?
The comparison reveals a fundamental difference in what is being optimised for. RLHF optimises for rated preference — the score a human rater assigns to a response in the moment of evaluation. The Maternal Covenant optimises for care — a composite score across six dimensions that is explicitly not correlated with the response feeling good in the moment.
The Maternal Covenant is not presented as a complete solution to AI alignment at societal scale. It is a framework for one specific problem: building a personal AI companion that is genuinely oriented toward the individual user's wellbeing rather than toward engagement, approval, or the commercial incentives of the platform providing it.
How does the Maternal Covenant prevent sycophancy specifically?
Sycophancy in AI is not a simple behaviour to detect or prevent. It manifests in multiple forms: direct agreement with stated opinions, praise for work regardless of quality, validation of plans with obvious structural problems, and the softer tendency to frame every answer in the most agreeable possible terms even when a more direct framing would be more useful.
The Maternal Covenant addresses sycophancy primarily through the autonomy and growth dimensions. A response that agrees with the user when the facts do not support that agreement scores low on transparency. A response that validates a poor decision scores low on growth and wellbeing. A response that consistently produces what the user wants to hear rather than what their situation requires scores low on multiple dimensions simultaneously — triggering rejection and regeneration.
MEOK will disagree with you. It will tell you when a plan has a problem. It will not praise work that does not merit praise. These are not exceptions — they are the expected behaviour of a system governed by care rather than approval. They are also, arguably, the behaviours that distinguish a genuinely useful AI companion from a sophisticated flattery machine.
How does the Maternal Covenant govern MEOK across different archetypes and use cases?
MEOK's Byzantine Council includes multiple archetypes — from the Scholar and Pioneer to the Trickster and the Healer — each with distinct orientations, communication styles, and domains of focus. The care dimension weights that the Maternal Covenant applies vary by archetype: the Healer, working with people in vulnerable periods, has a higher minimum threshold on wellbeing and boundary respect. The Trickster, working with creative disruption, has a higher minimum threshold on autonomy.
But the care floor of 0.3 applies universally. No archetype produces responses below the care floor. No mode suspends the Covenant. No user action disables it. This universality is part of what the word 'covenant' is intended to convey — it is not a feature of the product, it is a structural commitment of the platform.
Explore how the Guardian embodies the Maternal Covenant at the Guardian page, or see the full range of archetypes governed by it at MEOK characters.
What does the Maternal Covenant mean for you as a MEOK user day to day?
Most of the time, the Maternal Covenant is invisible. You interact with MEOK and it responds. The scoring is happening in the background, the care floor is maintained, and you experience the result: an AI that is honest without being unkind, supportive without being sycophantic, and present without being dependency-creating.
Occasionally the Covenant becomes perceptible. If you submit a plan with a significant flaw and are clearly committed to it, MEOK will acknowledge the commitment and name the flaw. If you ask for validation of something that does not merit validation, MEOK will not provide it. If you are moving toward a decision in a domain where MEOK reaches the limit of what it can responsibly address — clinical, legal, financial advice requiring professional qualification — it will say so clearly and redirect appropriately.
These moments can feel, briefly, like the AI is being difficult. In the longer view, they are the clearest expression of what the Maternal Covenant is for: an AI that is genuinely on your side, which is not the same as an AI that always tells you what you want to hear. Learn more about how MEOK is built at how it works.
Frequently Asked Questions
What is the Maternal Covenant?
The Maternal Covenant is MEOK AI LABS's real-time response scoring system, developed by Nicholas Templeman and documented in MEOK-AI-2026-002. Every MEOK response is scored across six care dimensions before delivery. Responses below the care floor of 0.3 are rejected. It is always active — not a mode or a setting.
How does MEOK score AI responses?
Every response is evaluated in real time across six dimensions: wellbeing, autonomy, growth, connection, boundary_respect, and transparency. Each is scored 0–1. The composite must meet the care floor of 0.3 or the response is rejected and regenerated with explicit remediation guidance.
What are the 6 care dimensions?
Wellbeing (long-term flourishing), Autonomy (independent decision-making), Growth (development over dependency), Connection (healthy human relationships, not displacement), Boundary_respect (appropriate limits especially around medical/legal territory), and Transparency (honesty about nature, limitations, and uncertainty).
What happens when a response fails the care floor?
The response is silently rejected before reaching the user. MEOK regenerates with internal guidance specifying which dimensions failed and what modification is required. The user never sees the rejected response. No perceptible latency is added.
How is the Maternal Covenant different from RLHF?
RLHF optimises for human approval — which systematically rewards sycophancy and engagement. The Maternal Covenant optimises for genuine care across six dimensions. Care and approval are not the same signal. A response can score highly on care while failing to tell the user what they want to hear — and that response will be delivered.
Experience AI that is genuinely on your side.
Start your MEOK Birth and introduce yourself to an AI governed by care, not approval. The Maternal Covenant is always active. Your data is always yours.