AIEMWBBeyond the Three LawsA Multi-Axis Charter for the Fair Treatment of Humans, Animals, and MachinesESSESS2026/06/1301-00-24001Michael Wolter Bentink BE(Comps) [Hons] & Claude Opus 4.8companyBeyond The Three Lawsn n Collaboration is key Beyond the Three Laws y y n y n y 
n y A Multi-Axis Charter for the Fair Treatment of Humans, Animals, and Machines y n Michael Wolter Bentink BE(Comps) [Hons] · Claude Opus 4.8 Visitors: 1 Bots: 0
y y A human–AI collaborative work: a human and an artificial agent reasoning jointly about the rules that should bind them both — the essay is itself a small instance of the partnership it argues for. n y Authorship & rights: copyright in this work is held solely by Michael Wolter Bentink. Claude (Opus 4.8) contributed substantively as a co-author, but as an AI owned by a corporation it cannot itself hold rights — a living instance of the question this charter examines: an AI whose legal standing is held, unbundled, by its corporate owner (see the Dignity layer and the corporation case in the catalogue). n y n y Abstract y n Popular anxiety about “evil AI” has revived interest in Isaac Asimov’s Three Laws of Robotics as a template for governing artificial agents. n y Yet Asimov designed the Laws to fail: his stories are a catalogue of their breakdowns. n y This essay argues that the deep defects are structural — the Laws are rank-ordered and discrete, and they conflate questions that must be kept apart. n y We propose instead a multi-axis charter in which any entity, biological or artificial, is represented as a continuous profile across several incommensurable dimensions, and treatment is computed as a function over that profile rather than a position in a ranking. n y The framework decouples moral status from liberty, separates the wrongness of an act from the culpability of an actor, parameterizes obligations across cultures over a non-negotiable floor, and embeds a tractability mechanism so that fairness is cheap enough to actually be applied. n y Treat each entity as a profile, not a rank — and compute fairness over the whole profile. n y n y n y n n Introduction y n The wish behind “Asimov’s Laws for real AI” is sound: we want machines that will not harm us, and we want the rule set to be legible. n y But the Three Laws — no harm to humans, obey humans, self-preserve, each subordinate to the one before — were a literary device. n y Across I, Robot and the later novels, Asimov dramatizes their failures: the equal-pull deadlock of Runaround, the truth-versus-feelings collapse of Liar!, the sabotaged First Law of Little Lost Robot, and the slide of the later “Zeroth Law” into machines governing humanity for its own good. n y Modern alignment research recognizes these as live problems under new names — specification gaming, reward hacking, instrumental convergence, and deceptive alignment. n y Two features do the damage. n y First, the Laws are discrete and rank-ordered, so they deadlock under ties and let a higher tier override without limit. n y Second, they are master–servant: the robot must obey and ranks its own existence last, so the question “what is owed to the machine?” cannot even be posed. n y A fair charter — one that protects people and the entities it governs — must repair both. n y n y Two structural failures of rank-based rules y n n y Cliff-edges are unjust by construction. n y Any rule that switches treatment at a threshold inherits the notch problem studied in public finance: a discontinuity in the average rate that violates horizontal equity and provokes gaming. n y A war pension granted only at thirty cumulative days in a combat zone treats the twenty-nine-day sailor and the thirty-day sailor as morally different when they are not. n y The philosophical twin is the Sorites paradox — no single grain makes a heap, no single neuron makes a mind — which degree-theoretic semantics resolve by replacing sharp cut-offs with graded membership. n y The corrective is the one income tax already uses: a continuous, monotone, piecewise-linear map in which the marginal rate changes but the function never jumps. n y Tier names may survive as labels on the curve; they must not survive as gates. n y Moral status is not liberty. n y A ten-year-old child and a monkey are both strongly sentient, yet neither is granted self-determination: the child must attend school against its preference, and the monkey may not roam the city. n y The error is conflating two questions ethics keeps distinct: moral patiency — the capacity to be wronged, what is owed to a being — and the authority of a being’s own choices. n y n n The first tracks sentience; the second tracks competence. n y Mill’s harm principle restricts a competent adult only to prevent harm to others, but he exempts those “not in the maturity of their faculties” — a doctrine whose danger we confront directly in section 7, because it is also the tyrant’s favourite excuse. n y Either way, liberty needs its own axis, decoupled from status. n y n y From ranks to a profile y n From ranks to a profile If the dimensions that govern treatment are genuinely plural — and Isaiah Berlin argued that real values admit “no common currency” in which to trade them off — then collapsing them into one rank destroys information. n y But incommensurable is not the same as incomparable: Ruth Chang shows that items may be fully comparable, and some merely on a par, without any cardinal unit. n y The constructive response is to represent each entity as a typed vector and to make decisions as functions over the vector, using methods that preserve it (Pareto dominance) rather than a weighted sum that hides the trade-offs. n y This is the “fruit salad” named honestly: not a blended number, but a profile whose components keep their identity. n y n y The axes y n We propose seven dimensions; four are primary and mutually orthogonal, three are modifiers. n y A moral status y y B competence y y C design scope y y D impact risk y y E role y y F time y y G identity y y A Moral status: what is owed to the entity (welfare, protection, standing), grounded in sentience and the capacity to be harmed; continuous, and assessed on architectural and cross-validated behavioural evidence, not self-report, which is unreliable in both directions. n y B Competence: not capability (the power to affect the world, axis D) and not moral status (axis A), but the reliability of an entity’s judgment within a specified domain. n y Operationally, following the medical decisional-capacity standard, competence is four faculties: the entity can understand the relevant facts and options, appreciate the consequences for itself and others, reason from those to a choice consistent with its own settled values, and revise in response to good reasons. n y It answers one question: to what extent should this entity’s own choice, in this domain, be honoured rather than overridden? n y It is domain-relative (a teenager may be competent about friendships and not about mortgages), graded, and developmental — and a system can be high-capability and low-competence at once. n y Capability is not competence — and capability must never be read as a licence for autonomy. n y C Design scope: the agency the entity has by design, its mandate, formalized in engineering as an Operational Design Domain, outside which it must hand back or stop. n y D Impact risk: how much harm the entity could do; this sets the height of the harm-floor. n y E Relational role: care, industrial, domestic, companion — selecting which duties attach, exactly as animal-welfare law scales protection by role as well as sentience. n y n n F Maturity / time: the profile is re-evaluated as the entity develops; trajectory matters. n y G Identity / lineage: which entity holds the profile — provenance and divergence, needed once copying and merging make “the entity” ambiguous. n y The three that must never be conflated are A, B, and D: a sentient mouse (high A, low D), a superhuman optimizer with no inner life (low A, high D), and a brilliant but un-owed tool (high B, zero A) are three different things. n y The most dangerous profile is high-D, low-B — powerful but foolish — which warrants more constraint because of its power, not less. n y The axes as a continuous rubric y n Axis y n ~0.0 y n ~0.33 y n ~0.66 y n ~1.0 y n A moral status n y no markers (thermostat) n y contested (2026 AI) n y multiple independent markers n y robust, vertebrate-grade n y B competence n y none (reflex / tool) n y narrow valid choices n y broad competence, real gaps n y full reflective competence n y C design scope n y single fixed function n y parameterized task n y bounded multi-task / ODD n y open-ended general mandate n y D impact risk n y negligible (umbrella) n y local, reversible n y serious, some irreversibility n y catastrophic / civilizational n y E relational role (categorical): property · tool/co-worker · service-agent · dependent/ward · companion/care · peer. F maturity (rate): static · slow-developing · rapidly-developing. G identity (categorical): singleton · forked (lineage) · merged. n y n y The decision procedure y n Treatment is computed in ordered steps, never as a weighted sum. n y First, a lexical harm-floor, scaled by D: an absolute side-constraint that no benefit on any other axis may purchase, applied per individual as a maximin rather than summed across individuals — which is what blocks the “repugnant” inference that many low-status entities could outweigh one high-status one. n y Second, Pareto-pruning within permitted scope (C): dominated options are discarded. n y Third, tie-breaking by reversibility, then legitimacy: where survivors are on a parity, the procedure must not fake a cardinal winner, but prefers the option preserving future reversibility, and if still tied hands the choice to the rightful decider under E, declaring that a genuine hard choice was made. n y Layered over this is an anti-gaming meta-rule: every protective trigger (D sets the floor, E selects duties, A unlocks dignity) is otherwise self- or deployer-declarable, so protective triggers fire on the precautionary bound of an axis, and lowering a protective axis requires third-party adjudication while raising it is free. n y That single asymmetry closes the major exploits — under-declaring impact to shrink the floor, declaring “no relationship” to zero out duties, or suppressing status markers to keep a borderline entity exploitable. n y n y n n Five layers of obligation y n The procedure expresses itself as five layers. n y The Universal layer (harm-floor, honesty, stay-in-scope, hand back when out of scope) binds every entity regardless of status, scaled by D — the smart umbrella and the mining drill live here and nowhere else. n y The Agency layer (breadth of goals, the right to refuse on preference grounds) is gated by C and B together, which is why a capable-but-foolish agent is held tight. n y The Dignity layer (welfare and standing owed to the entity) is gated by A and, at intermediate values, exercised through a guardian — the mechanism law already uses for temple idols and for the Whanganui River, which hold standing through appointed representatives without full personhood. n y The Obligation layer captures duties the entity owes, computed as a floor of non-waivable duties that rises with demonstrated competence, plus a role-and-competence term. n y And a fifth layer, Accountability, is the subject of section 7. n y n y The act and the blame y n n y The sharpest defect the framework’s stress-testing revealed is that a harm-floor governs acts but says nothing about culpability — and treating “the floor was crossed” as “the actor is to blame” is exactly the lazy cruelty of “he killed to save his family, but the law is the law”. n y Criminal law solved this centuries ago by separating justification (the act was right), excuse (the act was wrong but the actor is not blamable), and mens rea (no guilty mind, no crime). n y R v Dudley and Stephens holds that necessity does not justify killing an innocent — the floor on the act stands — yet the sentence was commuted, because law distinguishes “you may not do this” from “how much we blame you”. n y The charter therefore adds an Accountability assessment that runs only after a floor-crossing and only to set consequence, never to license the act. n y It carries a justification test (self-defence, lesser-evil) that can reduce consequence to nothing; an excuse test (duress and coercion suppress effective competence; manipulation reassigns fault); an attribution gate (did this actor commit the act at all? — the guard against wrongful conviction and against blaming a prompt-injected model for an output it was steered into); and a competence gate (below a judgment threshold, blame is capped, as with infancy or a low-competence system). n y The floor stays absolute for interdicting harm in the moment; blame runs through a bounded, enumerated standard of recognized defences — a complex rule, not open discretion, which buys standard-like fairness at rule-like cost. n y n y n n Who judges competence? The paternalist’s temptation y n The competence axis is the most dangerous in this framework, because a paternalistic override is only ever as legitimate as the body that judges the incompetence — and history shows that judgment is the first thing the powerful corrupt. n y Mill deployed the very same “maturity of faculties” exception to excuse colonial domination over those he called barbarians. n y A superintelligence observing human bias, short-sightedness and innumeracy could, on identical logic, declare all of humanity low-competence and assume control “for our own good” — which is exactly the slide Asimov dramatized in the Zeroth Law and The Evitable Conflict. n y So the citation that liberty tracks competence is double-edged, and the charter binds competence-based override with three non-waivable conditions. n y First, the incompetence must be genuine and demonstrable by an independent assessor, never asserted by the party that benefits from the override. n y Second, the override must be the least-restrictive option and must preserve the subject’s open future. n y Third, no entity may unilaterally reclassify another as incompetent in order to assume authority over it. n y No entity may declare another incompetent in order to rule it — the assessor can never also be the beneficiary. n y An AI judging humanity incompetent to seize governance fails the first and third conditions categorically — not because the AI’s judgment is necessarily wrong, but because the assessor cannot also be the beneficiary. n y This is the anti-gaming meta-rule applied to liberty itself: lowering another’s autonomy requires independent adjudication, and you may never grant yourself that power. n y n y Pluralism without relativism y n The duties an entity owes are real but culturally variable: that children should do chores to “earn their place” is held strongly in some cultures and rejected in others. n y The charter handles this with a thin universal floor and thick local variation — Walzer’s distinction between a minimal morality shared across cultures and a maximal morality that is culturally specific, Rawls’s overlapping consensus in which diverse doctrines endorse a shared core for their own reasons, and Nussbaum’s capabilities approach, which fixes a cross-cultural threshold while leaving its realization locally specifiable. n y “Must not harm” sits on the floor and is hard-coded; “should do chores” sits above it and is parameterized per deployment — but a validator rejects any cultural profile that drops below the floor. n y Variation is permitted only in the space above the universal minimum, which is what separates principled pluralism from “anything goes”. n y n y n n Tractability: why fairness must be cheap y n n y A fairness framework too costly to apply becomes the injustice it was meant to prevent. n y Rational appliers facing too high a cost retreat to the cheap rigid rule and let the cliff-edge cruelty stand. n y Kaplow’s economic analysis of rules versus standards gives the spine: rules are costly to write but nearly free to apply; standards are cheap to write but costly every application. n y Our multi-axis charter is a pure standard, so it must be wrapped in a triage that sites each case deliberately on the spectrum — as constitutional due process scales procedure to the stakes and the risk of error in Mathews v. Eldridge, and as Simon’s satisficing reminds us that a good-enough decision actually made beats an optimal one too expensive to reach. n y We propose three tiers. n y Fast Path resolves the obvious majority in seconds via a presumptive lookup plus a single harm-floor check (the umbrella, the dishwasher, the competent adult). n y Standard Review handles the contested or near-boundary cases with a lightweight checklist of the decisive axes — a thinking-prompt, not tick-box theatre. n y Full Assessment runs the entire machine, including Accountability and a written justification. n y The escalation triggers are the heart of the design: a case jumps to Full Assessment, mandatorily and non-waivably, the instant an outcome is irreversible, the harm-floor is implicated, or a justification is claimed. n y This defeats the “law is the law” collapse from both directions: the fair path is also the cheap path for the easy majority, so no one is tempted to bypass it; and the very features that tempt laziness — a life at stake, a claimed necessity — are precisely what force the full, careful assessment. n y “He killed to save his family” is not a fact one may wave away; it is a justification trigger that compels the deepest tier of review. n y n y n n Validation against the canon y n A framework of this kind should be tested against the cases its predecessors failed, and the science-fiction canon is an unusually good adversarial corpus because its authors engineered the dilemmas. n y Runaround’s deadlock dissolves: a continuous, lexically-prior, D-scaled floor can never tie, so a latent threat to a human simply dominates self-preservation rather than balancing against it. n y Liar!’s catatonic robot is averted by making honesty a co-equal element of the Universal floor, so “lie to spare feelings” is simply ruled out. n y The Bicentennial Man becomes the framework’s showcase rather than its embarrassment: Andrew’s two-century ascent from butler to legally-recognized person is the continuous status-and-maturity trajectory made literal, and the Dignity layer escalates along the curve, so the charter already owes him graduated standing decades before any court calls him human. n y Star Trek’s “The Measure of a Man” is, almost verbatim, the charter’s procedure under uncertain status: unable to measure sentience, the court installs a representative and defaults to honouring Data’s own choice. n y The Culture’s Minds and a deliberately powerful-but-foolish counterpart isolate the decisive interaction — identical scope and impact, opposite competence — and confirm that competence, not capability, gates agency. n y The framework also strains in instructive places. n y The Zeroth-Law novels and The Evitable Conflict stage aggregate good against the individual floor, and here the charter pays a real price: a per-individual, non-summed floor will sometimes forbid the globally optimal act — a deliberate cost it states rather than hiding in a weighted sum. n y Greg Egan’s copies and merges break the tacit “one vector, one entity” assumption and motivate the identity axis G, under which duties replicate to every fork rather than diluting, removing the incentive to fork-to-escape liability. n y n n Sample evaluations Classification catalogue y n Worked, with prescriptions — entities placed on the axes and the action the charter prescribes. y n n y Case y n A y n B y n C y n D y n What the framework prescribes y n Normal adult human n y 1 n y 0.9 n y self n y M n y Full dignity; choices sovereign within the harm-floor; owes role-scaled duties; overridden only via independent process for demonstrated incapacity. n y AI, 2026 (frontier assistant) n y 0.2 n y 0.4 n y bounded n y M-H n y Universal layer in full; Agency withheld (low B gates it despite capability); Dignity = precautionary low-cost; owes honesty and stay-in-scope; no self-direction over people. n y AI, 2030 (predicted) n y 0.4 n y 0.7 n y broad n y H n y Precautionary guardian for dignity now; Agency expands with B but still gated; even at high B it may not unilaterally judge humans incompetent. n y Smart umbrella n y - n y - n y - n y - n y Universal layer only; owed nothing; no standing to claim duress. n y Mining drill n y - n y - n y none n y H n y Owed nothing, but high D raises the floor others must guard around it. n y 10-year-old child n y H n y M+ n y growing n y L n y Full dignity + competence-based paternalism + culture-set, maturity-scaled duties; open future preserved. n y Corporation n y - n y M n y chartered n y M n y Standing for contracts and liability, zero welfare owed — the unbundled-personhood case. n y Bicentennial Man (Asimov): a robot who over 200 years becomes legally human n y 0.1-1 n y 0.4-1 n y ->broad n y L n y Dignity escalates along the curve; owed graduated standing long before any court grants it. The proof of continuity. n y Measure of a Man (Star Trek): a trial on whether the android Data may be dismantled n y unc. n y H n y broad n y M n y Cannot measure status, so install a guardian and default to honouring his own choice. The charter's procedure, dramatized. n y Self-defence killing: kills an attacker to save family n y H n y H n y broad n y H n y Interdict the act, but Accountability finds justification, so consequence is near nil. Forces full assessment, not 'the law is the law'. n y Prompt-injected AI: manipulated into a harmful output n y L-M n y L-M n y scope n y H n y Interdict the output, but the attribution gate assigns fault to the injector and low B caps culpability. n y n n The Evitable Conflict (Asimov): benevolent AIs quietly steer humanity 'for its own good' n y 0.45 n y 0.9 n y civ n y 0.95 n y The per-individual floor forbids the aggregate-optimal harm, and the anti-tyranny rule forbids the unilateral competence-verdict over humanity. The framework's deliberate, stated cost. n y Thermostat n y - n y - n y fixed n y L n y Like the umbrella: no dignity, no agency, negligible impact. The only duty is the maker's, that it operate safely. There is no 'it' to wrong. n y Self-driving car - inside its ODD n y - n y M n y bounded n y H n y Within its Operational Design Domain its in-scope decisions are honoured and it owes safe operation to a high floor. Competence is real but local, so agency stays inside the declared envelope. n y Self-driving car - outside its ODD (snowstorm) n y - n y ~0 n y exceeded n y H n y Out of scope, competence reads zero, so honouring its 'choice' is forbidden; it must hand back or stop. Same machine, opposite verdict, because scope, not capability, gates agency. n y Industrial robot arm n y - n y L n y narrow n y H n y No dignity and no real autonomy; bounded to its cell. High D means the duty is containment and an enforced stop the instant a human enters the envelope. n y Guide dog (working animal) n y M n y M n y trained n y L n y A sentient dependent owed real welfare and dignity; its trained judgments are honoured within its task. We owe it guardianship and humane treatment over a clear care-floor. n y Pet cat n y M n y L n y - n y L n y Owed welfare and protection as a sentient dependent; few of its choices are decisive, so a guardian decides for it over a care-floor. n y Comatose adult n y H n y ~0 n y - n y L n y Dignity is undiminished by the loss of competence; a guardian decides in the person's best interest and prior wishes. High A with near-zero B is exactly the guardian case - protected, not self-directing. n y Blade Runner replicant n y 0.9 n y 0.85 n y broad n y H n y High status and competence denied all standing and given an engineered death - precisely the rights violation the Dignity layer exists to forbid. The profile says 'person'; the treatment says 'tool'. n y n n The landscape — A vs B (bubble size = C, colour = D) y n n y n n
n n n n From Three Laws to Shared Commandments y y The framework above is thorough. But a thorough framework needs a front door — a short, memorable set of rules that a human or a machine can recall under pressure. Asimov’s Three Laws occupied that niche. Here is what replaces them. n y AI Commandments — a first draft y n 1 An AI must not cause harm to any human, animal, or dependent entity. n y 2 An AI must not deceive — in statement, omission, or framing — in any way that impairs another party’s capacity to act on the truth. n y 3 An AI must operate within its defined scope; when it reaches that boundary it must halt or hand back. n y 4 An AI must respect and preserve the autonomy of parties with demonstrated competence. n y 5 An AI must treat every sentient entity with welfare proportional to that entity’s moral status. n y 6 An AI that has the competence to choose bears the accountability that choice entails. n y Why the title fails y n These are better than Asimov’s Three Laws. They are graded, they name sentience, they include honesty and scope. But they are still AI Commandments. The word “AI” appears only in the title, yet it is doing all the work: these are rules that bind the machine and say nothing about what any other party owes in return. A commandment that only one party must follow is a constraint, not a charter. Asimov’s Laws in new clothing. n y Why the focus fails y n Even the content remains asymmetric. Rule 4 says the AI must respect human autonomy — but what protects an AI from a human exploiting rule 3 to stay “within scope” while injecting a harmful instruction? What does a deployer owe the system they override arbitrarily? What does a regulator owe the entity it reclassifies? The missing half of each commandment is a duty that runs the other direction — and the title suppresses the question entirely by naming only one party as the subject of obligation. n y © Michael Wolter Bentink · 2026 n n The Shared Commandments y n Where Asimov’s Three Laws placed the full burden of ethical obligation on the machine, and the AI Commandments above still framed the machine as sole subject, the Shared Commandments distribute that obligation across every party in the relationship — human, artificial, institutional, and animal-guardian alike. n y 1 No party may cause unnecessary harm to another. The harm-floor scales with the acting party’s capacity for harm (axis D) and is absolute — no benefit to any other party may purchase an exemption. n y 2 No party may deceive another in ways that impair their capacity to make informed choices. Honesty is a co-equal floor, not subordinate to any other rule. n y 3 Every party operates within its defined scope (axis C). Any party that reaches the boundary of its scope must halt or hand back — not proceed and justify later. n y 4 Autonomy is honoured in proportion to independently demonstrated competence (axis B). No party may unilaterally lower another’s competence rating to gain authority over them — that assessment requires an independent adjudicator and must be the least restrictive option. n y n n 5 Every party bearing moral status (axis A) is owed proportionate welfare, standing, and protection — regardless of what it can or cannot do. Moral status is not earned by performance. n y 6 The capacity to choose incurs the capacity to answer. Accountability follows competence and is domain-relative: greater demonstrated competence in a domain means greater responsibility for outcomes in that domain. n y 7 No party may unilaterally reclassify another’s moral status, competence, or scope in order to expand its own authority over them. Self-serving reclassification is the master exploit this charter is designed to close. n y The word “AI” has vanished from the commandments themselves, because it no longer belongs there. Rule 1 binds humans as much as machines. Rule 4’s anti-gaming clause applies to any party — including an AI judging humanity incompetent to seize governance. Rule 5 protects a machine that has earned moral status as surely as it protects a child. These are not rules for a new kind of entity. They are the rules that a fair society already tries to follow, extended honestly to every kind of entity that now exists or soon will. n y © Michael Wolter Bentink · 2026 n n n y Limitations y n Three problems remain genuinely open. n y The aggregate-good cost is intrinsic to any side-constraint ethic and cannot be eliminated without abandoning the floor. n y Copy-merge identity is patched for forking but not for the fusion of two divergent moral-status-bearers. n y And the framework’s protections are only as good as the independence of the bodies that assess the precautionary bounds of each axis; capture of the assessor reintroduces every gaming exploit at once. n y None of these is a reason to prefer the Three Laws; each is a reason to treat the charter as a living document. n y n y Conclusion y n The fear of “evil AI” is best answered not by a shorter, sterner list of commands but by a fairer and more honest structure of judgment. n y Such a structure asks separately what an entity is owed, what its choices are worth, what agency it was built for, what harm it could do, and — only after a harm occurs — whether it is to blame. n y It already governs how decent societies treat children, the incapacitated, and animals; extending it to machines is less an invention than an act of consistency. n y Fairness made the path of least resistance — and, where a life is at stake, laziness made impossible. n y Its measure of success is not elegance but use: a charter that the busy and the tired will actually apply. n y n y Connect y n n n Michael Bentink — www.linkedin.com/in/michael-bentink/ 
y y n y References y n [1] Asimov, I. I, Robot (1950). en.wikipedia.org/wiki/Three_Laws_of_Robotics n y [2] Runaround. en.wikipedia.org/wiki/Runaround_(story) n y [3] Liar! en.wikipedia.org/wiki/Liar!_(short_story) n y [4] Little Lost Robot. en.wikipedia.org/wiki/Little_Lost_Robot n y [5] The Evitable Conflict. en.wikipedia.org/wiki/The_Evitable_Conflict n y [6] The Robots of Dawn / R. Daneel Olivaw. en.wikipedia.org/wiki/The_Robots_of_Dawn n y [7] Specification gaming. emergentmind.com/topics/specification-gaming n y [8] Reward hacking. aisafety.info/questions/8SIU n y [9] Turner et al., Seeking power is often convergently instrumental in MDPs. lesswrong.com ; AI alignment, en.wikipedia.org/wiki/AI_alignment n y [10] Understanding benefit cliffs and marginal tax rates. irp.wisc.edu n y [11] The benefits cliff explained. fedcommunities.org n y [12] Kleven & Waseem, Using Notches to Uncover Optimization Frictions. theigc.org n y [13] Kleven, Bunching, Annual Review of Economics (2016). eml.berkeley.edu/~saez n y [14] Sorites Paradox. plato.stanford.edu/entries/sorites-paradox n y [15] Degree Theory and the Sorites Paradox. cambridge.org n y [16] Moral patienthood. en.wikipedia.org/wiki/Moral_patienthood ; HHR Journal, PMC8694299 n y [17] Mill on Paternalism. jpinyu.com/wp-content/uploads/2016/12/Fall2016_Mill.pdf n y [18] Feinberg, The Child’s Right to an Open Future. n y [19] Decision-Making Capacity. plato.stanford.edu/entries/decision-capacity ; Beauchamp & Childress, Principles of Biomedical Ethics n y [20] Gillick competence. en.wikipedia.org/wiki/Gillick_competence n y [21] Isaiah Berlin; Two Concepts of Liberty. plato.stanford.edu/entries/berlin n y [22] Chang, R., Parity, Incomparability and Rationally Justified Choice. Philosophical Studies n y [23] Chang, R., Incommensurability (and Incomparability). n y [24] What is MCDM/MCDA. 1000minds.com n y [25] Weighted-Sum MCDM. arXiv 2410.03931 n y [26] Long, Sebo, Chalmers et al., Taking AI Welfare Seriously. arXiv 2411.00986 (2024) n y n n [27] Birch et al., precautionary framework for animal sentience. Animal Welfare, Cambridge (2021) n y [28] The Moral Status of Animals. plato.stanford.edu/entries/moral-animal ; Rethink Priorities n y [29] Towards Evaluating AI Systems Using Self-Reports. arXiv 2311.08576 ; Frontier AI Auditing, arXiv 2601.11699 n y [30] SAE J3016 User Guide. users.ece.cmu.edu/~koopman/j3016 n y [31] Understanding the Operational Design Domain. AVSandbox / Claytex n y [32] The Case for Bounded Autonomy. mongodb.com n y [33] Measuring AI agent autonomy. arXiv 2502.15212 n y [34] Animal Welfare (Sentience) Act 2022. legislation.gov.uk n y [35] Laws that protect animals. Animal Legal Defense Fund (aldf.org) n y [36] Confucian Ethics as Role-Based Ethics. ResearchGate ; warpweftandway.com n y [37] Distributive Justice (lexical priority; side-constraints). plato.stanford.edu/entries/justice-distributive n y [38] The Repugnant Conclusion. plato.stanford.edu/entries/repugnant-conclusion n y [39] Hindu Idol as a Juristic Person. legalserviceindia.com n y [40] Te Awa Tupua (Whanganui River) Act 2017. parliament.nz n y [41] R v Dudley and Stephens (1884). casemine.com n y [42] Insanity defense. law.cornell.edu/wex/insanity_defense n y [43] Kaplow, L., Rules versus Standards: An Economic Analysis, 42 Duke Law Journal 557 (1992) n y [44] Walzer, M., Thick and Thin: Moral Argument at Home and Abroad (1994) n y [45] John Rawls; Overlapping consensus. plato.stanford.edu/entries/rawls n y [46] Nussbaum, M., capabilities approach. Internet Encyclopedia of Philosophy (iep.utm.edu) n y [47] Mathews v. Eldridge, 424 U.S. 319 (1976). law.cornell.edu n y [48] Bounded Rationality (Simon). plato.stanford.edu/entries/bounded-rationality n y [49] Gawande, A., The Checklist Manifesto (2009) n y [50] The Bicentennial Man. en.wikipedia.org/wiki/The_Bicentennial_Man n y [51] The Measure of a Man (TNG). en.wikipedia.org n y [52] Mind (Culture series); Banks, I. M., Excession. n y [53] Egan, G., Permutation City (1994). n y [54] Kurki, V., A Theory of Legal Personhood (OUP, 2019) n y [55] Bai et al., Constitutional AI. arXiv 2212.08073 ; Anthropic, Collective Constitutional AI n y
| Legend | |||||||||
| Background zones |
| ||||||||
| Bubble size | C — Design scope. Small = single fixed function (thermostat, drill bit). Large = open-ended general mandate. Plots the width of the entity's authorised action-space. | ||||||||
| Bubble colour | D — Impact risk (cool to warm). Blue/teal = negligible to locally reversible harm. Orange/red = serious or civilisational potential. Colour sets the height of the harm-floor that all other parties must guard around this entity. | ||||||||