Differential Diagnosis: Seven Hypotheses

The Question

AI transforms economic activity. But how it transforms economic activity — and who benefits — depends on which model correctly describes the mechanism. Getting this wrong means misidentifying both the problem and the solution.

The differential diagnosis applies the same methodology physicians use: generate competing hypotheses, test each against evidence, eliminate those the evidence contradicts, and compose a model from the survivors. The approach matters because the dominant narrative about AI — that it makes everyone more productive, a rising tide lifting all boats — did not survive evidence testing. What survived describes a more complicated and consequential dynamic.

The seven hypotheses below represent the major competing explanations for how AI reshapes the economy. Each receives a fair hearing. Three get eliminated by evidence. Four survive and compose into a single model. One modulates the others. Understanding which hypotheses survived and why provides the analytical foundation for everything else on this site.

Seven Hypotheses

H1: Productivity Multiplier (Eliminated)

Claim: AI doubles developer output, creating a straightforward productivity gain across the software industry.

Evidence against: The METR study (February 2026) found experienced open-source developers worked 19% slower with AI assistance on real-world tasks. Faros AI reported 75% of organizations saw no measurable productivity gains. The SF Fed found limited macro-level evidence of AI productivity impact.

Why it matters: Under H1, AI would simply make everyone more productive — a rising tide lifting all boats, with minimal ICESCR relevance. Its elimination means the mechanism operates differently. The gap between what AI companies promise (doubled productivity) and what rigorous measurement finds (slower or neutral) represents one of the most consequential disconnects in the current technology narrative. Organizations making staffing decisions based on the productivity promise — including layoffs justified by projected AI capability rather than demonstrated performance — act on a hypothesis the evidence contradicts.

H2: Constraint Removal (Survives)

Claim: AI reduces the marginal cost of software labor toward zero, removing a constraint that previously bounded economic activity. Projects that were infeasible when they required hiring developers become feasible when AI handles the coding.

Evidence for: Observable with individuals and small teams building software that previously required entire engineering departments. Solo developers shipping products that would have required 10-person teams two years ago.

Why this differs from H1: The productivity multiplier (H1) says AI makes existing work faster. Constraint removal says AI makes previously impossible work possible. The difference matters enormously: a productivity multiplier distributes benefits roughly proportionally to existing workers, while constraint removal creates entirely new categories of activity — and the people who access those new categories capture disproportionate value.

ICESCR connection: Article 15 — AI creates enormous new economic value. Who accesses it? When a solo developer ships a product that previously required a team of ten, the economic value created does not flow to the nine workers who would have been hired. It flows to the person who used AI to remove the constraint. Article 15’s right to benefit from scientific progress addresses exactly this concentration.

H3: Jevons Explosion (Survives)

Claim: When production costs drop, demand for production explodes exponentially. The historical precedent: content creation. When digital tools made content cheap to produce, the world did not produce the same amount of content more cheaply — it produced vastly more content.

Evidence for: Global AI spending reaches approximately $2 trillion in 2026, with 22% compound annual growth rate. The investment exceeds what simple productivity gains would justify — consistent with explosive demand.

How to think about the Jevons effect: When coal became cheaper in 19th-century England, total coal consumption did not decrease — it exploded, because cheaper energy unlocked new uses for energy that had been economically infeasible. The same dynamic applies to software: when AI makes software labor nearly free, the economy does not produce the same amount of software more cheaply. It produces vastly more software — for purposes no one would have considered when software required expensive human labor to create. This demand explosion reshapes the labor market not by eliminating software jobs (as the simple automation narrative predicts) but by transforming what software jobs entail.

ICESCR connection: Articles 6 and 7 — the labor market restructures as demand for what kind of work gets done shifts fundamentally. Workers whose skills match the new demand thrive. Workers whose skills matched the old demand face displacement. The speed of this restructuring — measured in months, not decades — exceeds the capacity of existing retraining programs.

H4: Bottleneck Migration (Survives)

Claim: When one constraint lifts, the next constraint becomes binding. AI removes the software labor bottleneck; the system runs until it hits the next wall. Four new bottlenecks emerge: regulation, energy, human judgment, and data quality.

Evidence for: AI capital expenditure of $527 billion in 2026 includes massive data center and energy infrastructure investment — consistent with energy becoming the next constraint. Regulatory frameworks scramble to keep pace. Organizations report difficulty finding people who can evaluate AI output.

The migration pattern in practice: Consider a company that adopts AI coding assistants. Software labor ceases to constrain development — the team produces code faster. But now code review becomes the bottleneck: more code generated means more code to evaluate. The company hires reviewers, but reviewing AI-generated code requires different skills than writing code from scratch. Meanwhile, the energy bill for running AI models rises. Each solved bottleneck reveals the next, and the value in the system migrates from “who can write code” to “who can evaluate code” to “who can specify what code should accomplish” — a chain of scarcity shifts that transforms the skills organizations need.

ICESCR connection: Articles 9, 11, 13 — transition costs fall on workers displaced by bottleneck shifts; energy costs affect living standards; judgment development requires educational transformation. The workers least equipped to navigate each bottleneck shift tend to absorb the highest transition costs.

H5: Recursive Acceleration (Eliminated)

Claim: AI builds better AI tools, creating a recursive acceleration loop.

Evidence against: Current improvements trace primarily to human research and engineering, not AI self-improvement. The mechanism lacks empirical verification; human R&D explains observations more parsimoniously.

H6: Quality Erosion (Survives as Modulator)

Claim: More code produced means lower average quality. Maintenance debt from AI-generated code offsets productivity gains.

Evidence for: DevOps reports of “productivity at the cost of code quality.” Observable increase in AI-generated code that passes initial review but fails under edge cases.

A concrete example: An AI system generates a medical diagnostic algorithm. It works well for common conditions represented in its training data. But for rare conditions, or for populations underrepresented in training data, it performs unpredictably. A premium healthcare provider validates the algorithm extensively before deployment. A cost-constrained clinic deploys the same algorithm without validation. Both report “AI-assisted diagnostics.” The quality gap between them — invisible to patients — carries life-or-death consequences. This dynamic applies across every domain where AI generates critical outputs: legal analysis, educational assessment, financial advice, infrastructure monitoring.

ICESCR connection: Articles 12 and 13 — AI-built healthcare and education software of uncertain quality poses direct risks to rights. Quality erosion does not mean AI produces bad output everywhere. It means the average quality across all AI-generated output declines as the volume increases, creating a distribution where premium and commodity tiers diverge sharply.

H7: Bifurcated Economy (Survives)

Claim: AI benefits distribute unevenly. Organizations that deeply integrate AI pull ahead; those with surface-level adoption stagnate. Workers in AI-adopting sectors gain; workers in lagging sectors lose.

Evidence for: 34% of organizations report deep AI transformation, 30% redesign processes, 37% surface-level only. Only 19% report ROI improvements exceeding 5%.

What bifurcation looks like from inside: Two accounting firms operate in the same city. Firm A deeply integrates AI: automated bookkeeping, AI-assisted tax analysis, machine learning for anomaly detection. Its accountants handle three times as many clients, commanding higher fees for the judgment and oversight they provide. Firm B adopted AI chatbots for client communication but otherwise operates traditionally. Its accountants compete on price, handling the same number of clients at lower margins. Over three years, Firm A’s revenue grows 40%. Firm B’s revenue stagnates. The accountants at Firm B did nothing wrong — their firm’s adoption strategy determined their economic trajectory. Multiply this pattern across every industry and the bifurcation becomes structural.

ICESCR connection: ALL articles — uneven distribution of AI benefits maps directly onto uneven enjoyment of economic rights. The bifurcation does not follow individual merit or effort. It follows organizational adoption patterns that individual workers cannot control.

The Surviving Model

The evidence supports a composite model — Composite A — combining four surviving hypotheses:

Constraint removal + Jevons expansion + bottleneck migration + bifurcation, modulated by quality erosion

The mechanism: Constraint removal (H2 — AI makes software labor nearly free) triggers demand explosion (H3 — cheaper production creates more demand, not less) that runs until the next bottleneck binds (H4 — removing one constraint reveals the next). All effects distribute unevenly (H7 — adopters gain, non-adopters absorb costs). Quality erosion (H6 — more AI output, lower average quality) acts as persistent drag if unaddressed.

Discriminator score: 20/25 — high empirical support, strong parsimony (four components explain the full pattern), consistent chain integrity.

What this means for you. The economy does not simply “get more productive with AI.” It restructures around new constraints, creating winners and losers based on position relative to those constraints. The ICESCR protects the losers in this restructuring — the people whose work transforms, whose safety net frays, whose access to AI-enhanced services depends on ability to pay.

How Each ICESCR Article Connects

Component	Mechanism	Primary Articles
H2: Constraint Removal	AI creates new value; who accesses it?	Art. 15 (Science)
H3: Jevons Explosion	Labor market restructures; new work forms	Art. 6 (Work), Art. 7 (Conditions)
H4: Bottleneck Migration	Transition costs; energy; judgment scarcity	Art. 9, 11, 13
H6: Quality Erosion	AI-built critical services of uncertain quality	Art. 12 (Health), Art. 13 (Education)
H7: Bifurcation	Uneven benefit distribution	All articles

The full analysis continues through four orders of knock-on effects, tracing how these mechanisms compound and interact.