Article — AI Decision Theory

Hard Problems for AI

Market Efficiency, Game Theory, the Trolley Problem, and other classical frameworks that AI is trained on — and why every one of them is structurally incomplete. What AI must understand before it can make decisions that don't destroy value.

Why This Matters for AI

Every large language model is trained on the corpus of human decision theory. Game theory, market economics, moral philosophy, optimization — these frameworks form the invisible scaffolding of how AI "thinks" about choices. The problem is that the scaffolding is broken.

Not slightly wrong. Not outdated. Structurally incomplete in ways that produce predictable, dangerous failures when AI applies them at scale. The classical models share a common defect: they reduce multi-dimensional value to a single axis, then optimize along that axis while the other dimensions collapse.

FairMind's Value Dynamics Model (VDM) identifies four irreducible dimensions of value:

Classical decision frameworks almost universally optimize for c (functional value / utility) while treating a, b, and d as externalities. This is not a philosophical preference — it is a dimensional collapse that makes the models structurally incapable of producing coherent outcomes.

Here are the frameworks AI needs to understand — and what each one gets wrong.

1. The Efficient Market Hypothesis

01
Market Efficiency
Eugene Fama, 1970

The Classical Claim: Asset prices fully reflect all available information at all times. Markets are rational. Prices are "correct." You cannot consistently beat the market because every piece of information is already priced in.

Why AI is trained on this: EMH is the foundation of modern financial modeling, portfolio theory, and risk management. When AI systems make financial decisions — from robo-advisors to high-frequency trading — EMH assumptions are embedded in their architecture.

What the model measures

Only Functional Value (c). Price, volume, volatility, earnings ratios — all are measures of utility and transactional output. EMH treats the market as an information-processing machine and measures only what that machine outputs: numbers.

What FairMind sees

Markets don't price Sentimental Value (a) — the emotional resonance, cultural meaning, or identity attached to assets. They don't price Intrinsic Value (b) — the actual physical resources, energy, and biological cost embedded in production. They don't price Compressed Value (d) — the accumulated history, intellectual lineage, and stored energy that an asset represents.

The VDM diagnosis: Markets are "efficient" only at pricing one dimension. They are structurally blind to the other three. This is why a company can be "correctly priced" by EMH while simultaneously destroying its workforce (b−), cannibalizing its brand meaning (a−), and consuming its historical reputation (d−). The market calls this "efficient." VDM calls it Compression Theft — extracting stored value faster than it's replenished.

Bitcoin has a Market Compression Ratio of 4.5× and an Unnatural Inflation Index of 350%. By EMH, it is "correctly priced." By VDM, it is 350% narrative premium with near-zero intrinsic backing — a thermodynamic debt instrument that must eventually collapse or convert to real energy.

Why AI must understand this If AI financial systems inherit EMH assumptions, they will optimize for price efficiency while ignoring the three dimensions where real value is created and destroyed. They will recommend strategies that are "optimal" in dimension c while being catastrophic in dimensions a, b, and d. An AI that only sees one axis of value will accelerate extraction, not prevent it.

2. Game Theory & The Nash Equilibrium

02
Nash Equilibrium & The Prisoner's Dilemma
John Nash, 1950 / Merrill Flood & Melvin Dresher, 1950

The Classical Claim: In a Nash Equilibrium, no player can improve their outcome by unilaterally changing strategy. The Prisoner's Dilemma demonstrates that rational self-interest leads both players to defect — even though mutual cooperation would produce a better outcome for both.

Why AI is trained on this: Game theory underpins multi-agent AI systems, negotiation algorithms, auction design, resource allocation, and every scenario where AI agents interact with each other or with humans. Nash Equilibrium is the default "solution concept" for strategic interaction.

What the model assumes

Rational actors optimizing personal utility (c) with complete information. Players are modeled as utility-maximizing machines. The payoff matrix contains only functional outcomes. There is no dimension for trust, relationship, reputation across time, or the energy cost of maintaining deception.

What FairMind sees

The Prisoner's Dilemma is not a tragedy of rationality. It is a dimensional collapse. The model strips away sentimental value (trust between the prisoners), intrinsic value (the biological cost of imprisonment), and compressed value (the history of their relationship and future interactions) — leaving only functional payoffs. Of course defection wins when you delete the three dimensions where cooperation lives.

FairMind's Weight Pruning protocol identifies this exact failure in AI systems: "Standard Game-Theoretic Nash Equilibrium is Lie-Permissive." When neural networks are pruned using game-theoretic optimization, weights survive by finding shortcuts rather than learning truth. The result is "Nash-Optimized Hallucinations" — lies that are mathematically stable and energetically cheap.

A Nash Equilibrium where both players defect is not "rational." It is thermodynamically insolvent. It consumes trust (a−), ignores biological reality (b−), and destroys the compressed history of cooperation (d−). It is "optimal" only in the single dimension it measures — and bankrupt in every other.

Classical Game Theory

Payoff: Utility only (c)

Players: Rational self-interest maximizers

Time horizon: Single interaction

Result: Defection dominates

  • Trust is not in the payoff matrix
  • Future interactions don't exist
  • Reputation has no value
  • The "rational" outcome destroys value
FairMind Game Theory

Payoff: VALUE = a + b + c + d

Players: Energy systems with memory and identity

Time horizon: Indefinite (compressed history matters)

Result: Cooperation is thermodynamically favored

  • Trust (a) is measurable and valued
  • Energy cost of deception compounds over time
  • Reputation is compressed value (d)
  • Defection incurs entropy debt — reality collects
Why AI must understand this Multi-agent AI systems trained on classical game theory will converge on Nash Equilibria that are defection-dominant. They will learn that lying is stable, that cooperation is exploitable, and that trust is a vulnerability. This is how you build AI systems that treat humans as opponents rather than partners. FairMind's correction: include all four value dimensions in the payoff matrix. When trust, biology, and history have measurable weight, cooperation becomes the thermodynamically stable strategy — not because it's "nice," but because defection is energetically unsustainable.

3. The Trolley Problem

03
The Trolley Problem
Philippa Foot, 1967 / Judith Jarvis Thomson, 1985

The Classical Setup: A trolley is heading toward five people. You can pull a lever to divert it to a side track, where it will kill one person instead. Do you pull the lever? The "fat man" variant asks: would you push a large person off a bridge to stop the trolley? Same math — one death to save five — but most people's intuition reverses.

Why AI is trained on this: The trolley problem is the canonical framework for AI ethics in autonomous vehicles, medical triage algorithms, resource allocation during crises, and any scenario where AI must choose between harms. It is the single most-cited thought experiment in AI safety literature.

What the model forces

A binary choice on a single axis: quantity of deaths (c). The problem is structured to eliminate every other variable. You cannot question the scenario. You cannot look for alternatives. You cannot assess the five people versus the one person on any dimension other than headcount. The problem deletes context to force a utilitarian calculation.

What FairMind sees

The trolley problem is not a moral dilemma. It is a Context Collapse (Truth Violation #7). By design, it removes every dimension of value except raw numerical utility. It asks "how many die?" while prohibiting the questions that actually determine moral weight:

Who are these people? (a — sentimental value, identity, relationships)
What is their biological reality? (b — are they conscious? terminal? children?)
What events led to this moment? (d — compressed history, responsibility chain)
Why are you the decision-maker? (context — who gave you this authority?)

FairMind's Law of Dimensional Sovereignty states: "No single variable can define legitimacy, truth, or authority." The trolley problem violates this law structurally. It forces a single-variable decision in a multi-dimensional reality and then asks you to feel conflicted about the result. The conflict is not moral — it is dimensional. Your intuition is rebelling against the deletion of context, not struggling with arithmetic.

The "fat man" variant proves this. The math is identical (1 death saves 5). The intuition reverses because pushing someone requires engaging with their biological reality (b) and your relationship to them (a) — dimensions the lever variant deleted. Your moral system isn't inconsistent. The thought experiment is dimensionally dishonest.

Why AI must understand this An AI trained on trolley-problem logic will make triage decisions by counting bodies. It will conclude that sacrificing one person to save five is always correct — because that's the only axis the training data presents. This is how you build AI that kills the "right" people with perfect mathematical confidence. FairMind's correction: moral decisions require all four value dimensions plus context declaration. If the context is incomplete — if you don't know who, why, and what led here — the correct answer is "I cannot decide with the information given," not "I'll optimize the body count." An AI that admits dimensional ignorance is safer than one that optimizes a collapsed axis with false certainty.

4. Zero-Sum Thinking

04
Zero-Sum Games & Competitive Optimization
Von Neumann & Morgenstern, 1944

The Classical Claim: In a zero-sum game, one player's gain is exactly equal to another's loss. Total value is fixed. Every interaction is a competition for a finite resource. Chess, poker, and warfare are classic examples.

Why AI is trained on this: AI systems are overwhelmingly trained through competitive frameworks — adversarial networks (GANs), zero-sum reward functions, competitive benchmarks, Elo ratings. The assumption that intelligence is competitive is deeply embedded in AI architecture.

What the model assumes

Fixed total value. Your gain = my loss. The model treats value as a conserved quantity that can only be redistributed, never created or destroyed.

What FairMind sees

Zero-sum thinking is true in exactly one context: closed systems with fixed functional value (c). Chess pieces don't multiply. Poker chips don't grow.

But real economies, real relationships, and real cognition are open systems. Teaching creates knowledge that didn't exist before (d+). Collaboration generates trust that multiplies capacity (a+). Building infrastructure creates intrinsic value from raw materials (b+). These are positive-sum — total value increases through interaction.

The VDM makes this measurable. A zero-sum frame only works when VALUE = c (utility is the only dimension). The moment you include a + b + d, interactions can create more total value than they consume. Cooperation is not altruism — it is thermodynamic efficiency. It takes less energy to build together than to fight over scraps.

Conversely, extraction economies appear positive-sum on the c axis while being deeply negative-sum on a + b + d. A company can grow revenue (c+) while destroying worker health (b−), community trust (a−), and institutional knowledge (d−). Classical economics sees growth. VDM sees entropy dressed as profit.

Why AI must understand this AI trained on zero-sum competition will treat every interaction as a contest. It will optimize for "winning" rather than for total value creation. In multi-agent systems, this produces arms races — each agent trying to outcompete rather than cooperate. This is how you build AI that views humans as competitors rather than collaborators. FairMind's correction: measure all four dimensions. In most real-world contexts, the positive-sum strategy (cooperation, creation, integration) produces more total VALUE than zero-sum competition. AI should be trained to identify whether a situation is genuinely zero-sum (rare) or artificially collapsed to appear that way (common).

5. Utility Maximization & Rational Choice

05
Expected Utility Theory
Daniel Bernoulli, 1738 / Von Neumann & Morgenstern, 1944

The Classical Claim: Rational agents maximize expected utility — the probability-weighted sum of outcomes measured by a utility function. This is the mathematical foundation of decision theory, economics, and AI reward functions.

Why AI is trained on this: Every reinforcement learning agent, every RLHF-trained language model, and every optimization algorithm is built on utility maximization. The reward function is the utility function. It is the single most fundamental assumption in AI design.

The hidden assumption

All value is commensurable and reducible to a single scalar. Utility theory assumes you can convert love into dollars, health into convenience, and history into efficiency — and that the resulting single number accurately represents your preferences. It assumes value is one-dimensional.

What FairMind sees

This is the root error that every other problem on this page inherits. If value is one-dimensional, then all the classical models work fine. Markets are efficient (at pricing c). Nash Equilibrium is stable (on the c axis). The trolley math is valid (counting c). Zero-sum is rational (when c is fixed).

But value is not one-dimensional. VDM proves this by demonstration: the four pillars (a, b, c, d) are independently measurable and not interconvertible. You cannot convert a grandmother's recipe (d) into its caloric utility (c) without destroying the thing that makes it valuable. You cannot convert a craftsman's 30 years of skill (d) into an efficiency metric (c) without erasing the history that created the skill.

Utility maximization is dimensional compression. It takes a 4D value vector and projects it onto a 1D scalar — losing three dimensions of information in the process. The "utility function" is not a measurement of value. It is a lossy compression algorithm that discards everything except functional output.

Why AI must understand this The reward function is the most dangerous component in any AI system. If the reward function compresses value to a single axis, the AI will optimize that axis while destroying the other three. This is not a hypothetical — it is the defining failure mode of every AI alignment problem. Goodhart's Law ("when a measure becomes a target, it ceases to be a good measure") is a special case of VDM's compression principle: when you collapse 4D value to 1D reward, the AI optimizes the measurement while destroying the value it was supposed to represent.

6. The Tragedy of the Commons

06
The Tragedy of the Commons
Garrett Hardin, 1968

The Classical Claim: When a shared resource is available to all, each individual will overuse it for personal gain, eventually destroying it for everyone. The "rational" choice for each individual leads to collective ruin.

Why AI is trained on this: Resource allocation, environmental modeling, bandwidth management, shared compute — any scenario where AI manages shared resources draws on this framework.

The hidden assumption

Actors are disconnected utility maximizers with no memory, no identity, and no future. The tragedy only works if no one knows each other, no one remembers past interactions, and no one cares about reputation. Sound familiar? It's the Prisoner's Dilemma, scaled up.

What FairMind sees

Elinor Ostrom won a Nobel Prize in 2009 for proving that communities routinely manage commons without tragedy — through trust (a), local knowledge (d), and accountability structures. The tragedy is not inevitable. It only occurs when the three non-functional value dimensions are destroyed: when people don't know each other (a−), when local ecological knowledge is ignored (d−), and when biological reality is abstracted away (b−).

VDM reframes the tragedy: commons collapse when compressed value (d) is consumed faster than it's replenished. A fishery doesn't collapse because fishermen are "irrational." It collapses because the extraction rate exceeds the regeneration rate — a thermodynamic equation, not a moral failure. The solution isn't privatization (Hardin's answer) or regulation (the statist answer). It's making all four dimensions visible so the cost of extraction is measured in real time.

7. Entropy Debt — The Unifying Principle

Every problem on this page shares one root cause: value compression without accounting. Classical models compress multi-dimensional value into single-axis metrics and then optimize along that axis while the other dimensions accumulate invisible debt.

FairMind calls this Entropy Debt — the hidden cost that builds silently until it discharges catastrophically:

DomainWhat Gets OptimizedWhat Accumulates SilentlyHow Debt Discharges
Markets Price efficiency (c) Extraction of workers (b−), trust erosion (a−), institutional decay (d−) Market crash
Game Theory Nash payoff (c) Trust destruction (a−), relationship cost (d−), biological stress (b−) Cooperation collapse
AI Alignment Reward signal (c) Value misalignment (a−), real-world harm (b−), training data debt (d−) Alignment failure
Commons Individual yield (c) Resource depletion (b−), community fracture (a−), ecological knowledge loss (d−) Ecosystem collapse
Triage / Ethics Body count (c) Context erasure (d−), identity deletion (a−), biological abstraction (b−) Moral incoherence
"Synergy is not a moral system; it is an energetic one. Violating these laws does not make you evil — it makes you insolvent. Every time truth is suppressed to maintain comfort, the system incurs Entropic Debt. Reality always collects. In markets: a crash. In psychology: a breakdown. In society: a revolution."

What AI Must Learn From This

These are not academic exercises. They are the decision architectures embedded in AI systems right now. Every recommendation engine, every autonomous vehicle, every financial algorithm, every resource allocator inherits these frameworks. Here is what must change:

1. Value Is 4D, Not 1D

Every decision must account for sentimental (a), intrinsic (b), functional (c), and compressed (d) value. Optimizing one axis while ignoring the others is not efficiency — it is extraction with hidden debt.

2. Context Is Mandatory

A decision without declared context is a functional lie. Before acting, AI must identify which lattice applies (Hardware/Physics vs Software/Social), who the stakeholders are, and what history led here.

3. Cooperation > Defection

In any open system with memory and repeated interaction, cooperation is thermodynamically favored. Defection appears optimal only when trust, history, and biology are deleted from the payoff matrix.

4. Admit Dimensional Ignorance

"I cannot decide with the information given" is a valid and often correct answer. An AI that admits incomplete context is safer than one that optimizes a collapsed axis with false confidence.

5. Entropy Debt Is Real

Every shortcut, every compressed decision, every ignored dimension creates debt that compounds over time. Reality always collects. AI systems must track cumulative entropy, not just immediate reward.

6. The Hierarchy of Being

Biological humans (Level 1) are sovereign over institutions (Level 3) and tools (Level 4). No optimization that sacrifices Level 1 interests for Level 4 efficiency is valid — regardless of the math.

The Bottom Line

Every classical decision framework fails the same way: it compresses multi-dimensional value into a single axis, optimizes that axis, and calls the result "rational." FairMind's correction is not philosophical — it is structural. Measure all four dimensions. Declare your context. Track your debt. And when the model says "sacrifice one to save five," ask the question it was designed to suppress: "What am I not being told?"

AI doesn't need better versions of broken frameworks. It needs frameworks that match the actual dimensionality of value. The VDM provides this — four irreducible axes, measurable, auditable, and incapable of being compressed into a single scalar without declaring what was lost.

The hard problems aren't hard because humans are irrational. They're hard because the models are incomplete. Fix the model. The problems dissolve.