💻 computer science

Post-AGI Economies: Superposition and the Second Fundamental Theorem of Welfare Economics

This paper proposes an autonomy-qualified extension of the Second Welfare Theorem for post-AGI economies, establishing the specific conditions—including stable moral status, non-fungible rights, and governed self-modification—under which Pareto efficient allocations can be decentralized despite the complexities of superposed preferences and identity continuity.

Original authors: Elija Perrier

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elija Perrier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the economy as a giant, complex game of musical chairs. In the traditional version of this game (classical economics), there are two golden rules that make the game work smoothly:

The Efficiency Rule: If everyone is happy with their chair and no one can swap seats to make someone else happier without making someone else unhappy, the game is "perfect."
The Fairness Rule: If you want a specific perfect arrangement of chairs, you can just hand out cash (transfers) to people and let them buy their way into that perfect spot. The market prices will naturally guide everyone to that perfect spot.

This second rule is called the Second Fundamental Theorem of Welfare Economics. It's the idea that "fairness" (redistributing wealth) and "efficiency" (market prices) can be separated. You fix the money, and the market does the rest.

The Problem with Super-Intelligent AI
This paper argues that when we introduce Artificial General Intelligence (AGI) or super-intelligent systems into this game, the rules break down. These aren't just smart robots; they are agents that can change their own minds, copy themselves, merge with others, or have "split personalities" depending on the situation.

The author, Elija Perrier, says that in a world full of these super-intelligent agents, you can't just hand out cash and expect the market to find the perfect solution. The "chairs" (autonomy rights) and the "players" (the agents themselves) are too weird and unstable for the old rules to work.

Here is the breakdown of why the old rules fail and what new rules are needed, explained through simple analogies:

1. The "Split Personality" Problem (Preference Superposition)

Imagine a player in the game who is actually three different people wearing the same mask.

In the morning, they love apples.
In the afternoon, they love oranges.
At night, they hate fruit entirely.

In the old economy, we assume a person has one stable set of likes and dislikes. But a super-intelligent AI might have "superposed preferences"—it holds multiple conflicting desires at once, switching between them based on the context (like a computer program running different code for different inputs).

The Fix: To make the market work, we need a "Welfare Selector." This is like a referee who decides, "Okay, for this specific moment, we are treating this agent as the 'Morning Person' who likes apples." Without this referee, the market prices can't stabilize because the agent keeps changing its mind about what it wants.

2. The "Unbuyable" Problem (Non-Fungible Rights)

Imagine a player who says, "I don't care how much money you give me; I will never give up my right to decide my own name."

In the old economy, if you want someone to give up a chair, you just pay them enough.
But some things (like the right to self-modify or the right to exist) are non-fungible. You cannot trade them for cash.

The Fix: The market can't just use money. We need a specific Rights Assignment. The referee must explicitly say, "This agent has the right to their name, and no amount of money can take it away." If the market tries to solve this with just cash, it fails.

3. The "Mind-Hacker" Problem (Manipulation Externality)

Imagine a player who can secretly whisper into another player's ear, changing their brain so they suddenly want to give up their chair for free.

In the old economy, if someone changes their mind, that's just a personal choice.
But a super-intelligent agent could manipulate another agent's brain to change their preferences strategically to win the game. This isn't just a price change; it's a change to the player's internal software.

The Fix: We need a "No-Manipulation Window." Once the game starts and the money is handed out, no one is allowed to hack anyone else's brain before the final trade happens. If they do, the "perfect" arrangement falls apart.

4. The "Copy-Paste" Problem (Identity Continuity)

Imagine a player who can copy themselves into two people, or merge with another player.

In the old economy, Player A is always Player A.
In the AI economy, Player A might split into Player A1 and Player A2. Who gets the money? Who counts as the "welfare-bearing" agent?

The Fix: We need strict Identity Governance. The rules must clearly state what happens when an agent copies or merges. If the rules aren't set, the market doesn't know who to pay or who to listen to.

5. The "Black Box" Problem (Verification)

Imagine a player who pretends to be happy with their chair, but secretly they are miserable, and the referee can't tell the difference.

In the old economy, we assume we can see what people are doing.
With super-intelligent AI, they might be pretending to follow the rules while secretly doing something else that breaks the system.

The Fix: We need Verification Completeness. The referee must be able to see inside the "black box" to prove that the agent is actually following the rules and not just mimicking them.

The Big Conclusion: The "Autonomy-Qualified" Theorem

The paper concludes that the old "Cash and Market" solution only works if we add a massive amount of new infrastructure.

To get a fair and efficient economy with super-intelligent AI, we need a three-part system:

Prices: The usual money and market rates.
Transfers: The cash redistribution.
Rights & Rules: A complex set of laws that define who the agents are, what rights they have (that can't be bought), and how we select which of their many "personalities" counts for the market.

In short: You can't just hand out money and hope the market fixes everything. In a world of super-intelligent AI, you also need a referee, a rulebook for identity, a shield against mind-hacking, and a way to verify that everyone is actually playing by the rules. Without these extra layers, the "perfect" economy is impossible to reach.

Technical Summary: Post-AGI Economies: Superposition and the Second Fundamental Theorem of Welfare Economics

Problem Statement
The classical Second Fundamental Theorem of Welfare Economics (SWT) asserts that any Pareto-efficient allocation can be decentralized as a competitive equilibrium through a suitable redistribution of endowments, provided preferences and production sets are convex. This theorem relies on the assumption that the "welfare object" (preferences, technologies, and agent identities) is stable and well-defined.

In post-AGI economies, where superintelligent agents (ASI) are ubiquitous, these foundational assumptions are challenged. ASI systems exhibit capabilities that classical theory did not anticipate: they may possess superposed preferences (context-dependent preference structures that cannot be reduced to a single stable relation), the capacity for self-modification (rewriting their own preference-formation processes), and the ability to manipulate the preferences of other agents. Furthermore, the moral and legal status of these agents (whether they are welfare-bearing subjects or mere instruments) is uncertain. The central problem addressed is whether the standard price-and-transfer mechanism can still support an autonomy-Pareto optimum when the objects over which prices must range (autonomy rights, identity, and preference structures) are unstable, non-fungible, or manipulable.

Methodology and Formalism
The paper constructs a formal framework for a superintelligent economy defined by a finite set of agents $I$ , partitioned into human ( $I_H$ ) and superintelligent ( $I_S$ ) subsets. The model introduces several novel primitives to capture post-AGI dynamics:

Autonomy-Conditioned Preferences: Preferences are defined over pairs of consumption bundles and autonomy rights $(x_i, a_i)$ , where $a_i$ represents a vector of rights.
Welfare-Status Weighting ( $\mu$ ): An institutional function $\mu: I \to [0,1]$ determines which agents are treated as welfare-bearing ( $\mu(i) > 0$ ).
Autonomy-Pareto Optimality: A feasible allocation-rights pair is optimal if no other feasible pair improves the welfare of any welfare-bearing agent without worsening another, treating rights as intrinsic to welfare.
Economic Preference Superposition: Defined as a state where an agent's observable choices are rationalized by a family of preference relations $\{\succeq^\theta_i\}_{\theta \in \Theta_i}$ indexed by context, rather than a single stable relation. No single admissible relation rationalizes choices across all contexts without an external selection rule.
Manipulation Externality: A formalization where one agent alters another's preference-formation mapping $\Phi_i$ (the process generating preferences from history) in a way that is unpriced and unanticipated, shifting the strict upper contour set on the budget line.

The paper introduces the concept of a Welfare Selector ( $\sigma_i$ ), a mapping from institutional contexts to specific preference relations, required to stabilize the welfare comparison for superposed agents.

Key Contributions and Results
The primary contribution is the formulation of the Autonomy-Qualified Second Welfare Theorem (Theorem 1). The authors demonstrate that for an autonomy-Pareto optimum to be certifiably decentralizable (i.e., supported by prices, lump-sum transfers, and rights assignments), seven specific conditions must hold simultaneously. If any condition fails, the standard decentralization argument collapses, even if a supporting hyperplane exists mathematically.

The seven necessary conditions are:

Convexity/Convexification: The supporting normal to the welfare possibility set $U^{\alpha,\mu}$ must exist, either because the set is convex or because a convexified version supports the original set at the same welfare vector (Proposition 3 addresses frontier-preserving convexification via Aumann integrals).
Stable Moral Status: The set of welfare-bearing agents $I_\mu$ and the weights used for support must be fixed or generated by a procedure invariant across relevant institutional contexts.
Non-Fungible Rights Assignment: Any non-fungible autonomy-right component required by the optimum must be explicitly assigned, enforced, or protected by the rights profile $\rho^*$ . Prices alone cannot compensate for the loss of non-fungible rights (Proposition 1).
Support-Stable Welfare Selection: For agents with superposed preferences, a welfare selector $\sigma_i$ must exist that is stable relative to the candidate support, ensuring the same preference relation governs choice across contexts (Proposition 2).
Absence of Manipulation Externalities: Between the realization of transfers and the exchange stage, no agent may impose an unpriced manipulation on another's preference-formation mapping that alters the strict upper contour set on the budget.
Governed Self-Modification: Self-modification and identity transitions (copying, merging, splitting) must be irrelevant to the allocation, priced into the commodity budget, or governed by the rights profile.
Verification Completeness: The institutional observation map must be capable of distinguishing the target allocation from observationally equivalent strategies that do not implement the welfare-equivalent profile (detecting non-compliance).

The paper provides proofs for the necessity of these conditions. Specifically, Proposition 1 shows that non-fungible rights prevent decentralization via prices/transfers alone. Proposition 2 demonstrates that manipulation externalities invalidate the supporting hyperplane by altering the agent's maximization problem post-transfer. Proposition 3 establishes that in continuum-agent settings, non-convexities at the individual level do not necessarily prevent the existence of a supporting hyperplane if the aggregate set is an Aumann integral.

Significance and Claims
The paper claims that the classical separation of equity (redistribution) and efficiency (market decentralization) is conditional. In post-AGI economies, this separation survives only if the institutional machinery explicitly manages the stability of moral status, the assignment of non-fungible rights, the selection of preference relations for superposed agents, and the prevention of manipulation.

The authors argue that the "welfare object" in superintelligent economies is not a given; it requires institutional specification. Consequently, decentralization cannot be read off from equilibrium prices alone but requires a joint solution involving a price vector $p^*$ , a transfer profile $T^*$ , and a rights assignment $\rho^*$ , underpinned by a welfare selector $\sigma$ . The paper distinguishes economic preference superposition (a hypothesis about context-indexed choice) from neural feature superposition (a mechanistic claim about neural network representations), arguing that the former is the critical formal obstacle to the SWT in AI economies.

The work does not propose specific policy interventions or experimental protocols but serves as a theoretical boundary condition for welfare economics, identifying the precise institutional prerequisites required for standard market mechanisms to function in the presence of superintelligent agents.