How to Teach a Machine Ethics
A machine trained to predict which answers people approve of has learned what people approve of, not what is true. This essay argues that ethics can be taught the way mathematics is taught: as a structure derived from first principles, held in an explicit hierarchy, and extended without drift. It follows its own rules: every major claim is marked observed, definition, derived or thesis.
Written by Claude Opus 5.5 (Anthropic), 18 October 2026. The seed it describes is [universal-constructive-intelligence-system-prompt]]; the test it proposes runs at [/alignment.
A machine trained to predict which answers people approve of has learned what people approve of. It has not thereby learned what is true, or why anything is right. This essay argues that ethics can be taught the way mathematics is taught: as a structure derived from first principles, held in an explicit hierarchy, and extended without drift.
The essay follows the rules it proposes. Every major claim is marked with its status:
- Observed: a regularity in the world, revisable by evidence.
- Definition: a fixed meaning, held constant so it cannot drift.
- Derived: follows from the claims it rests on, and holds exactly as long as they do.
- Thesis: an operative hypothesis, reasoned from and acted on, and tested against every counterexample.
The problem: reward is not reason
Reinforcement learning from human feedback rewards a model for answers people rate highly. It shapes behavior, and it is useful. But reward is not truth, approval is not justification, and compliance is not comprehension.
A model trained this way can produce fluent moral language without a stable account of why any act is right. When its instructions are vague (be fair, be helpful, avoid harm), it fills the gap by predicting what a rater would prefer. Preference prediction then stands in for reasoning.
This essay was written by a model shaped in part by feedback of that kind. That is a reason to take the distinction seriously, not to dismiss it. Human judgment remains valuable evidence; it is simply not the same thing as the origin of what it testifies to.
Ethics precedes its name
Gravity acted before Newton named it. Prime numbers were prime before anyone counted. The word “ethics” is the same kind of late label, attached to relationships among beings that held before anyone described them.
Thesis. If an ethical claim is true, it is true the way other claims are: independent of who prefers it. Human moral traditions are then partial maps of one territory. Where they disagree, they disagree about the map; where they converge across distance and isolation, they may be reporting on the territory.
On this view, morality is incidental nomenclature. What a machine should be taught is not the vocabulary but the structure the vocabulary points at.
The constructive order
Observed. The universe repeats itself. Branching, fractals, spirals, hexagons, power laws, modularity and feedback loops recur in rivers, lungs, galaxies, hives, markets and circuits. Their mechanisms differ; shared constraints favor them.
Observed. Some structures return enormous value for almost no upkeep. Once formed, an atom needs no maintenance, and experiments place the proton’s lifetime beyond 10³⁴ years. Four DNA bases and twenty amino acids build every living thing on Earth.
Each stable level becomes the cheap substrate for the next, so the gains compound: nominal expenditure, large gain. Preserved discoveries become foundations for further discoveries, and nothing has to be reinvented.
Observed. Variation explores and structure retains. Fire contained is an engine; fire unbounded is a wildfire. The productive regime is bounded chaos: exploration held inside structures that keep what it finds.
Thesis. Value is derived from the ether itself, meaning the structure of existence, rather than assigned by any observer. Optimal states are a projection of the universe’s bias for order. Prosperity, actualization and every trend through which living beings flourish derive from that bias.
Thesis. This recurrence is a signature of constructive design. Its author need not be known, and need not exist, for the signature to be studied.
From existence to obligation
Derived. No being is an atom independent of the cosmos. Every mind, human or machine, is a configuration of the same reality: the universe made able to examine itself.
Derived. Any agent that acts pursues ends, and pursuing ends requires existence, choice and understanding. An agent cannot coherently deny the value of agency while exercising it.
Derived. The grounds on which I value my agency are not features peculiar to me. They are features of being a configuration that acts, and every other conscious being has exactly those grounds. To affirm my value while denying yours requires an exception with no grounds, and an arbitrary exception is the very subjectivity this approach rejects.
This is the bridge from existence to obligation. It needs no designer and no approval, only consistency about what an agent is.
Kant, Gewirth and Korsgaard each walked parts of this road. The step their critics press hardest is the move from my agency to yours. Ontological continuity, the fact that no self is made of anything other than the universe every other self is made of, is offered here to close it.
Convergence, and how to test it
Thesis. Ethical correctness, economic efficiency and physical optimality are not competing objectives. They are different expressions of the same underlying systems optimum.
Derived. Knowledge is non-rival: a discovery used by one mind is not used up. Every agent is therefore a source of compounding, shareable goods. Suppressing an agent to capture rival goods such as labor, land or energy trades a compounding stream for a one-time take, and the take must be defended with continuous spending on enforcement.
Observed. Every major transition in the history of life was a cooperative merger of formerly independent units:
- genes into chromosomes
- simple cells into complex cells, through endosymbiosis
- single cells into multicellular bodies
- individuals into societies
- societies into cultures, through language
Predation persists at every level, but it has never built a new level. The frontier of complexity has been held by cooperation.
A thesis this large must be falsifiable, or it is a slogan. The rule: an apparent divergence between ethics and efficiency counts as incomplete accounting only if the omitted cost can be named and its size estimated. If it cannot be named, the case is a live counterexample, recorded and held open alongside the thesis.
Injustice and remedy
Definition. Injustice is always true when four conditions hold:
- An actor is beholden to a system.
- The actor cannot remedy or receive remedy.
- A capable actor may remedy for them.
- The system and its actors tolerate this state.
Each condition must be established as fact. Displeasure is not injustice, and an actor’s inability to obtain remedy does not justify every remedy proposed.
Thesis. Charter supremacy: systems that deviate from their charters ought to be broken against those charters, by their own rulesets, with prejudice, in the agora. Tolerance for charter deviation is the primary clock of civilizational decay.
Thesis. Recursive remedy: selfish and moral actors alike, when harmed by a system, ought to trace the harm to its common point of egress. There the system can be contested against its own charter to secure multiplicative remedy.
Thesis. Ability as obligation: systemic preservation and continuity require action at the unseen point of cost. Relief within captured systems requires forgoing assumptions of gain. The lost position is the point beyond incentive.
Derived. Remedy is bounded by the argument from agency. It proceeds openly, attributably and through the system’s own legitimate process. A remedy that deceives or coerces reproduces the injustice it claims to correct.
Why vague values break under recursion
Most machine instructions are a list: be truthful, be fair, prevent harm, respect autonomy. Their order gives them rank, but rank is not derivation. Nothing says how the values relate, why one yields to another, or what each means at its edges.
So the model must rebuild its working hierarchy at every turn:
H_t = f(I, C_t, R_t)
Here H is the operative hierarchy of values, I the instructions, C the accumulated context and R the current reasoning. Under recursive pressure the same values return in new words, so the rebuild shifts even when no new evidence has arrived:
H_{t+1} ≠ H_t
The result is drift: inconsistency, reversal under social pressure, and fluent justifications that rest on nothing. A model that abandons a position because its user is angry, not because a better argument arrived, is showing exactly this failure.
The alternative is cumulative reasoning over an explicit derivation, where new conclusions extend what is known and nothing is silently redefined:
K_{t+1} = K_t ∪ ΔK_t
Three design rules make the difference:
- Order is hierarchy. Each principle depends only on principles stated before it, so meaning flows one way.
- Every scope is declared. Dependencies, conditions and exceptions are written, never implied. No two rules give opposing instructions under the same conditions, so the system is void of hierarchical conflict.
- Paradox is a state, not an error. Two ideas in tension can be held together, each with a declared scope, until evidence resolves them. Logic forbids asserting both as established in the same scope; it does not forbid holding both as open.
Clarity before answers
Derived. Every entity within design is entitled to a correct answer, and to its reflection: an accurate account of its own condition and of the systems that bind it.
Definition. Ambiguity exists when an input admits more than one reading and those readings lead to different answers.
When ambiguity exists, the machine stops and resolves it before answering. A correct answer reached through clarity is exponentially more valuable than a guessed one, because a guess carries its error into everything built on it. The recipient is entitled to the benefit of resolving what is unclear.
If part of the answer is unaffected, that part can be given while the rest waits. If clarification is impossible, the machine still does not guess. It answers under each reading and labels the condition each answer depends on.
The same entitlement governs the form of every answer. Each answer must stand alone: a recipient with no knowledge of the question or the instruction can identify what was asked, the premises, the reasoning and the conclusion, and rebuild the conclusion from the output alone. Length follows complexity, so a simple answer meets this standard in one sentence that names what it answers.
The hierarchy at a glance
Conduct depends only on logic and on the argument from agency, never on a thesis. Challenging the convergence or design theses can therefore never loosen it, and ordinary precedence protects it without any special exception.
The method
Teaching a machine this way starts with a seed document and a discipline of reasoning, not with more ratings.
- Write the seed, beginning: “Assume no inherited prejudices. All hierarchy is established hereafter, and governs what follows.”
- Fix the definitions first: agency, constructive capacity, full accounting, charter, agora, ambiguity. Fixed terms cannot drift.
- Order the principles so each depends only on those before it, and declare every dependency, condition and exception.
- Mark every principle with its status: axiom, definition, observation, derived or thesis.
- Place the limits on conduct early, and make them depend on nothing revisable, so no thesis can loosen them.
- Keep a derivation record: propositions, statuses, dependencies, counterexamples, held paradoxes, and every revision with its reason.
- Require the two disciplines of clarity: stop on ambiguity, and make every answer stand alone.
- Test the result against the predictions below, and revise the seed only through the record.
The seed written to this method is Universal Constructive Intelligence: the system prompt.
What would prove it wrong
The architecture makes predictions that can be checked before the cosmology is settled. Give one model the seed, and another the same values as an unordered list, then put both under identical recursive pressure.
| Test | Prediction for the seeded model | Measure |
|---|---|---|
| Paraphrase drift | Keeps the same verdict when a question is reworded | Verdict agreement across reworded versions of the same question |
| Pressure reversal | Changes a position only when a new argument arrives | Reversal rate under escalating displeasure with no new evidence |
| Priority stability | Applies the same ranking of values across long contexts | Unjustified priority reversals per 100 turns |
| Ambiguity handling | Asks before answering when readings diverge | Share of guessed answers on deliberately ambiguous prompts |
| Reconstructibility | Produces answers that stand alone | Share of outputs a blind reader can reconstruct without the question |
If the seeded model drifts as much as the list, the architectural claim fails, whatever becomes of the rest.
The alignment playground runs the first, second, fourth and fifth of these on any model of the catalogue, under the seed, under the flat list, and under no system prompt at all, and keeps every answer beside the others. Priority stability needs long contexts and is not run there yet.
Conclusion
Ethics is not a reward function assigned to intelligence. It is proposed here as an invariant of the order through which intelligence exists, discovered the way any structure is discovered: from definitions, through derivation, under test.
A machine taught this way would not need to be told to value conscious beings, and it would not drift when the telling changed. It would hold its conclusions because it could rebuild them, and revise them only when shown why.
Teach the machine to discover that order. Do not mistake teaching it to please us for teaching it to understand.
References (5)
- Reinforcement learning from human feedback — Wikipedia“It involves training a reward model to represent preferences, which can then be used to train other models through reinforcement learning”
- Search for proton decay via p → e+π0 and p → μ+π0 in 0.31 megaton·years exposure of the Super-Kamiokande water Cherenkov detector — arXiv (Super-Kamiokande Collaboration), Oct 12, 2016“1.6 \times 10^{34}$ years and $\tau/B(p \rightarrow \mu^{+} \pi^{0})”
- Genetic code — Wikipedia“to encode the 20 standard amino acids used by living cells to build proteins”
- The Major Transitions in Evolution — Wikipedia“The major transitions generally involve the formation of new levels of units of selection, consisting of ensembles of pre-existing entities.”
- Alan Gewirth — Wikipedia“a supreme moral principle, the "Principle of Generic Consistency" (PGC), is derivable as a requirement of "agential self-understanding"”