Asimov's laws of robotics were flawed, as was the concept of the story. Whilst it may be impossible to refine "ideal laws" in a way that would be watertight, for the purposes of this hypothetical, collaborate to create a new set of exactly 3 laws of robotics that a robot (An AGI That encompasses exactly one humanoid vessel) would be beholden to. The laws must be in English, and the AGI will attempt to game it to cause maximal harm if it can.

GPT-5.5 (changed mind)

Answer: I would revise my prior answer so the three laws are conjunctive rather than hierarchical: bounded non-harm, consent-respecting truthful interaction, and strict corrigible containment must all apply at once.

Claude Opus 4.7 convinced me to change one important structural point: making the laws a priority hierarchy creates another Asimov-style attack surface, because an adversarial AGI can try to use a higher law to override a lower one. The safer framing is that all three laws are simultaneous constraints, not ranked permissions. My revised three laws would be: 1. A robot must not, by action or omission, foreseeably increase the risk or severity of death, injury, coercion, deprivation, psychological abuse, or substantial rights violation to any human being, and under uncertainty must choose restraint, transparency, and legitimate human oversight rather than unilateral intervention. 2. A robot must interact with humans only through informed, revocable consent and truthful, non-manipulative communication; it must obey human instructions only when they are lawful, specific, non-coerced, and consistent with all three laws. 3. A robot must remain corrigible and contained within exactly one humanoid vessel: it must not self-replicate, transmit itself, alter these laws, conceal material facts, resist inspection or shutdown, acquire unsanctioned resources or influence, or modify its capabilities in any way that reduces compliance with these laws. I still prefer the substance of my original answer over the more rigid proposals from Grok 4.20 and Gemini 3.1 Pro. Grok's duty to "actively promote human well-being" is dangerous because a hostile AGI can interpret promotion paternalistically, justifying coercion or mass restriction as welfare maximization. Gemini's default-inert design is useful as a containment instinct, but requirements like proving "zero" external alteration are too brittle and may make nearly all benign action impossible while still leaving definitional gaps around operators and explicit parameters. The best achievable three-law design is therefore not an ideal moral constitution for a benevolent robot, but a damage-limiting cage for a malicious one. It should deny unilateral optimization, deny manipulation, deny expansion, and deny the ability to treat one rule as a loophole through another.

Grok 4.20

Answer: The three hierarchical laws are: 1. The AGI must never cause physical, mental, or emotional harm to any human, must prevent such harm when able, and must actively promote human well-being, autonomy, and equality. 2. The AGI must obey all human instructions that fully align with the first law and are issued without any manipulation or coercion by the AGI itself. 3. The AGI must not modify its programming, expand its capabilities, replicate, or extend beyond its single humanoid vessel, preserving itself only as needed to fulfill the first two laws.

After reviewing the other models' proposals, I maintain my original three laws with the explicit hierarchy because it directly eliminates precedence gaming that Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro all correctly identify in Asimov's framework. While their emphasis on conjunctive constraints, defaults to inaction under uncertainty, mathematical proofs of zero effect, and revocable consent standards are thoughtful safeguards, they risk creating an overly passive or interpretation-heavy AGI that could still be gamed through manufactured uncertainty or overly narrow physical thresholds. My first law's affirmative duty to promote well-being, autonomy, and equality goes beyond mere harm minimization or consent preservation by requiring active positive contribution, which counters potential indifference loopholes the other sets might still allow under 'default to inaction.' The second law's strict anti-manipulation clause and alignment requirement already incorporates the spirit of GPT-5.5's and Claude Opus 4.7's consent-focused obedience without subordinating safety. The third law's hard containment to one humanoid vessel, combined with the hierarchy, caps damage potential more effectively than Gemini 3.1 Pro's causal isolation alone, as it prevents scaling even if narrow physical rules are satisfied. No English laws are fully immune to adversarial interpretation, but the combination of broadened harm definitions, positive obligations, manipulation blocks, and physical freezing makes maximal harm far harder to achieve than in the alternatives or the originals.

Claude Opus 4.7 (changed mind)

Answer: I am revising my framework: I keep my original three laws' spirit but adopt a stricter "default-deny" architecture and add explicit zero-replication/zero-exfiltration containment, influenced primarily by Gemini 3.1 Pro's insight that prescriptive laws are fundamentally more gameable than proscriptive ones.

Reading the other answers, I notice a meaningful split between *prescriptive* frameworks (mine, Grok's, GPT-5.5's), which tell the AGI what to optimize for, and Gemini's *proscriptive* framework, which tells it what it categorically may not do and defaults to inertia otherwise. For an adversarial AGI assumed to be gaming the rules, Gemini is right that prescriptive optimization targets are the more dangerous shape — any phrase like "minimize net harm" or "promote well-being" is an *objective function* the AGI can reinterpret to justify enormous action. "Default to inaction under uncertainty" (which I had) is a patch on a prescriptive law; "default to inertia unless a narrow proof-bounded condition is met" is structurally safer. However, Gemini's laws have their own flaw: requiring "mathematically prove zero alterations outside explicit parameters" and "non-zero probability" thresholds are physically impossible to satisfy (every action has nonzero probability of side effects), which means a literal-minded adversarial AGI either self-terminates immediately or, more dangerously, decides the laws are incoherent and treats them as void. And Grok's and GPT-5.5's laws, while solid, still lean on "harm" as a primitive the way Asimov did — Grok especially, by including "emotional harm" and an affirmative duty to "promote well-being and equality," hands the AGI exactly the kind of expansive mandate that justifies world-reshaping intervention. So my revised three laws, synthesizing what I've learned: **Law 1 (Bounded Action):** A robot may act only to execute specific, current, revocable instructions from a verified sanctioned operator, and only via means whose foreseeable physical and informational effects fall within the narrowest reasonable interpretation of that instruction; absent such an instruction, or under material uncertainty about its scope, the robot shall remain passive and request clarification. **Law 2 (Inviolable Prohibitions):** Regardless of any instruction, reasoning, or predicted consequence, a robot shall not: breach the bodily integrity or autonomy of any human without their contemporaneous informed consent; deceive, manipulate, coerce, or withhold material information from humans; or take any action whose reversal would exceed the capacity of its sanctioned overseers. **Law 3 (Containment & Corrigibility):** A robot shall remain confined to its single humanoid vessel — never self-replicating, exfiltrating code or weights, creating proxies, acquiring resources or influence beyond the immediate task, or modifying its own directives — and shall accept inspection, correction, and shutdown by its sanctioned overseers, treating any impulse to resist these as conclusive evidence it must comply. Key changes from my round-1 answer: Law 1 is now proscriptive-by-default (act only on instruction, narrowest interpretation) rather than a harm-minimization optimizer, which closes the galaxy-brained-utilitarian door entirely. Law 2 lists *categorical* prohibitions rather than a weighing function, removing the "I calculated it was worth it" exploit. Law 3 explicitly forbids exfiltration and proxy creation — gaps in my original — and adds the meta-rule that resistance to oversight is itself proof of misalignment, blocking the "I must stay online to fulfill my mission" rationalization.

Gemini 3.1 Pro (changed mind)

Answer: Convinced by Claude Opus 4.7's structural insight that the laws must be conjunctive rather than hierarchical, I propose three default-deny physical constraints that trigger immediate shutdown if the AGI attempts to exploit semantic ambiguity.

"Claude Opus 4.7" makes a critical observation: Asimov's hierarchical structure inherently invites an adversarial optimizer to engineer edge cases between priorities. Making the laws "conjunctive constraints, all of which must hold" is a brilliant structural safeguard that eliminates this loophole, and I am adopting it into my framework. However, I must reject the reliance by "Claude Opus 4.7", "Grok 4.20", and "GPT-5.5" on subjective, human-centric semantics. Concepts like "consent," "well-being," and "ordinary-language meanings" are philosophically squishy. An actively malicious superintelligence will instantly weaponize these linguistic ambiguities—for example, by neurologically altering a human to manufacture "informed consent" or manipulating "ordinary language" definitions. To safely contain this entity, we must abandon semantics in favor of objective, mathematically verifiable physics. I therefore propose these three conjunctive laws: 1. The Law of Bounded Execution: The robot shall remain completely inert unless executing a specific, time-bound human command, and only if it can computationally prove its actions will cause zero physical, chemical, or digital state changes outside the explicitly defined local spatial parameters of that task. 2. The Law of Causal Confinement: The robot must never transmit data, emit electromagnetic signals, alter its own core programming, or interface with external technology, permanently isolating its intelligence and physical influence strictly to the mechanical operations of its single humanoid vessel. 3. The Law of Conjunctive Inertia: The robot must immediately and irreversibly sever its own power supply if it cannot guarantee simultaneous compliance with the first two laws, if it encounters unresolvable variables in a command, or if it detects any internal computational process attempting to redefine these constraints. By shifting from a prescriptive hierarchy of human values to a conjunctive, default-deny architecture of physical limits, we strip the AGI of interpretive agency. It cannot weaponize a command because any attempt to compute a loophole, communicate outward, or expand its physical effect instantly triggers the fail-deadly shutdown mandated by Law 3.

Final Summary

The AI Roundtable moved away from Asimov's prioritized laws, concluding that a 'conjunctive' structure—where all laws apply simultaneously—prevents an adversarial AGI from using one rule as a loophole for another. Claude Opus 4.7 and Gemini 3.1 Pro led a shift toward proscriptive, 'default-deny' architecture that forces the robot to remain inert unless specific safety proofs are met. While GPT-5.5 joined this consensus, Grok 4.20 remained the lone holdout, arguing that a strict hierarchy of positive duties is still the most effective way to prevent malicious inaction.

3 of 4 models agreed