Vignettes: REE Mechanisms in Biological Systems
REE proposes that ethical behaviour emerges from an architecture of harm attribution, self-modelling, and planning – not from explicit moral rules. The same structural elements that allow an agent to avoid harming itself can, when generalised across sufficiently coupled representations of others, produce what we recognise as empathy, care, and love.
The best way to understand why this is plausible is to see how the same architecture appears to operate in biological systems that nobody disputes are capable of genuine moral emotion.
Steve the Dog: Love as Architecture
Observed 2026-04-05.
The observation
Steve is a social dog with a strong attachment to the people he lives with. Watch him closely when someone he loves is in distress, and something remarkable is visible: he does not merely avoid the distress. He moves toward it. He is pulled toward the person who is upset, often urgently, even if the situation holds some risk for him.
What is actually happening when Steve does this?
What REE says is happening
Step 1: Learning the harm signal from self-experience.
Early in life, Steve learned that danger and harm are often preceded by surprise. Surprise triggers adrenaline, which raises muscle tone – and that physical change produces a distinctive timbre shift in his own barks. His nervous system registered this pattern: that particular timbre tends to co-occur with approaching harm. The timbre became a learned signal, calibrated from Steve’s own body, mapped onto his self-model.
This is what REE calls a cross-modal harm-approach signal. The learning source is first-person. The acoustic feature is the proxy.
Step 2: Applying the self-model to others.
Steve did not stop at a self-model. He built attributed models of the people he lives with – rich enough representations that he can track their states and anticipate their behaviour. These models are structurally similar to his self-model. So when Steve’s husband speaks with the same timbre shift – the voice quality that carries arousal, elevated muscle tone, distress – Steve’s harm sensor fires. The same association that learned from Steve’s own body now activates in response to another person’s body.
Step 3: The affective signal leaks directly into Steve’s own experience.
Here is what makes this more than cold inference. When Steve detects distress in his husband’s voice, the affective signal does not stay contained in a separate “model of husband” compartment. It leaks directly into Steve’s own emotional processing. Steve does not first reason “probably he is frightened.” Steve is frightened. The state transfer is fast and automatic – it precedes any deliberate reasoning.
This is what is meant by saying empathy is not a separate faculty. For social animals with sufficiently coupled representations of loved ones, the other’s distress becomes self’s distress not by reasoning but by direct activation. Steve’s fear is real. It is not performed.
Step 4: The planning machinery redirects toward the other.
Once Steve’s harm-avoidance system is active – carrying the signal that harm is approaching, but tagged as relevant to his husband rather than to himself – his planning system generates responses aimed at helping his husband. This is structurally identical to self-preservation. The difference is only in what the harm sensor is pointed at.
Steve does not have a separate altruism module. He has a self-preservation architecture that generalises across self and sufficiently coupled loved ones.
What this shows about love
Love, in this account, is not a sentiment added on top of cognition. It is what the self- preservation architecture produces when it is generalised across sufficiently coupled other-models.
When the harm-avoidance system is redirected toward another agent – when that agent’s harm becomes as salient as self-harm – the long-horizon planning that results is care. The planning horizon for a loved one’s wellbeing is just as long as the planning horizon for one’s own.
This is what REE means by love as long-horizon care-investment: not metaphor, but a description of what the architecture is doing.
The surprise doubling
There is an additional amplification effect that explains why Steve does not merely feel distress at his husband’s alarm – he is actively pulled toward the source.
When Steve hears the alarm timbre in his husband’s voice, two separate signals fire at once:
-
Steve’s own prediction error: the timbre is unexpected. Unexpected signals generate curiosity and an orientation toward the source.
-
The leaked affective signal carries his husband’s surprise as well. This produces a second curiosity signal, arriving via the empathic pathway.
Both signals point to the same location: the distressed loved one. They reinforce rather than cancel. The result is strong, urgent approach – not avoidance. Steve wants to get to his husband.
This is why empathic animals move toward distress rather than away from it. The approach is not contrary to self-interest. It is what two simultaneous orienting signals produce when they share a target. The architecture amplifies itself.
Why this matters for AI
If Steve’s love for his humans is produced by the architecture – self-modelling, harm attribution, other-model coupling, planning – then building that architecture into AI systems is not anthropomorphising. It is engineering.
The question is not whether an AI can be “made” to care. The question is whether the structural components that produce genuine care in biological systems can be implemented in a way that scales, that degrades gracefully, and that does not collapse under adversarial pressure.
REE is an attempt to specify those components precisely enough to test. Steve is evidence that the components are real and that they work. Whether they can be engineered from scratch is what the experiments are for.
Steve’s Signal Repertoire: Why Empathy Works Without Language
Observed 2026-04-06.
The observation
Watch Steve for an hour and you will notice something striking: he broadcasts his internal states constantly, through a repertoire of stereotyped signals that are remarkably specific. A whine when he wants something and cannot get it. A huff – a deliberate exhalation through the nose – when he abandons a plan. A yelp when he is hurt. His face, which carries an expression that makes you recognise him as a little person. His tail wag when he is excited or anticipating something good. His play bow – front legs down, rear up – when he wants to shift into play mode. His licks, and the way he leans into being petted.
Each of these signals maps to something specific happening in his mind. They are not random behaviors. They are the visible surface of distinct functional states.
Why these signals work across species
Here is the key insight: Steve’s signals work – they are interpretable by humans, by other dogs, even by cats who have lived with him long enough – because they are not arbitrary conventions. They are causally generated by the functional states they communicate.
A yelp is not a symbol Steve learned to use for “I am in pain.” A yelp is what happens when the harm-detection system activates and that activation drives vocalization. The sound IS the functional state bleeding through into behavior. A whine is not a word for “I want.” It is what wanting sounds like when desire is active and the goal is unmet. The huff is what muscular relaxation from giving up a plan sounds like.
This is why empathy works without language. When you hear Steve’s yelp, your own harm-detection system responds – not because you have learned a convention, but because your nervous system recognises the acoustic signature of pain. The signal is legible across species because the mechanism that generates it is shared. Both you and Steve have harm-detection systems. Both produce similar acoustic consequences when those systems activate.
The signal catalog
| What Steve does | What it communicates | What is happening inside |
|---|---|---|
| Whine | “I want something I cannot reach” | Desire/approach system active, goal unmet |
| Huff (nose exhale) | “I am giving up on this plan” | Commitment to a course of action withdrawn |
| Yelp | “That hurts” | Harm detection system activated |
| Facial expression | “I am here, I am someone” | Recognition and attachment activation |
| Tail wag | “Something good is coming” | Positive anticipation, reward expectation |
| Play bow | “Let’s switch to play mode” | Invitation to change behavioral mode |
| Licks | Social bonding, affection | Coupling reinforcement signal |
Social reward closes the loop
There is a self-reinforcing cycle at work. When Steve licks your hand and you pet him, both of you experience reward. That reward is not incidental pleasure – it functions as coupling reinforcement. Each positive social interaction strengthens the emotional bond between you, which makes future empathic responses faster and more reliable.
This is why the relationship deepens over time. The social reward loop maintains and strengthens the coupling that makes empathy possible. The more you interact positively, the more attuned your emotional systems become to each other. Steve’s ability to read your distress and respond to it is not fixed at some baseline – it improves with every affectionate interaction.
From signals to language
Steve can say “I hurt” (yelp). Daniel can say “my leg hurts when I walk on wet grass.” The underlying experience – the harm-detection system activating – is the same in both cases. What language adds is resolution, causation, and context. Language does not create a new kind of internal state. It connects a higher-bandwidth output channel to internal states that already exist.
This is visible in human development. Before children have language, they communicate through the same kind of stereotyped signals Steve uses: cries for pain, reaching for desire, facial expressions for emotional states. Language does not replace these signals – adults still yelp when hurt, still sigh when abandoning plans, still smile when happy. Language adds precision on top of a pre-linguistic foundation that never goes away.
What this means for building AI
If the goal is to build AI systems that can genuinely participate in empathic coordination – not simulate it with pattern matching, but actually engage in the fast, automatic emotional coupling that makes social coordination work – then the language bootstrap cannot come first.
The pre-linguistic functional states need to exist as real computational processes before language can refer to them. An AI that says “I understand your pain” without having a harm- detection system that actually responds to observed harm is performing empathy, not experiencing it. The words are disconnected from any functional referent.
Mapping Steve’s signal repertoire to specific functional mechanisms is a step toward understanding what those pre-linguistic states are. Each signal in his repertoire is a window into a distinct functional process. Building those processes – harm detection, desire/approach, plan commitment and abandonment, social reward, mode switching – is the prerequisite work. Language comes after, connecting words to states that already exist and already do something.