Can AI Refute Economic Theory? Evidence from Beyond the Knowledge Cutoff

ORACLE PROTOCOL v5.0 — TEXT ANALYSIS

A. THE DISSECTION

This paper is a controlled demonstration of AI's current limits in theoretical falsification work. The author set up a clean experimental design: four published economics papers with known errors, four frontier AI models (Gemini, Refine, Claude, ChatGPT), human-guided and unguided conditions. The result is exactly what the DT framework predicts: AI cannot autonomously refute economic theory — but the human-AI pair outperforms peer review.

What the paper is really doing is staging a boundary demarcation exercise for academic economics. It tells the discipline: "Your peer review is weak, AI can augment it, but don't panic about autonomous refutation yet." This is comfort food for economists worried about their epistemic authority.

B. THE CORE FALLACY

The paper commits the incrementalist augmentation fallacy: it treats this as a question of whether AI is currently capable of autonomous refutation, framing the answer as "not yet" and treating "yet" as the operative word. This misframes the structural trajectory.

The relevant question isn't whether AI can currently refute economic theory. The relevant question is whether the architecture of theoretical economics — a system that depends on human peer review for error correction — is structurally resilient to AI-driven epistemic competition over time. The paper confirms that the current peer review system is a low-efficiency, high-latency process. AI integration will accelerate. The bottleneck is human cognitive bandwidth in review, not AI capability ceilings.

The paper treats "human-AI pair outperforms peer review" as a vindication of human roles. It is actually a confession of dependency. The humans needed to identify the errors in advance, feed them as prompts, evaluate outputs. The AI is a very expensive research assistant with a stunning memory and no judgment. For now.

C. HIDDEN ASSUMPTIONS

1. Theoretical economics is corrigible through error detection. This assumes the errors in economics are local (incorrect proofs, wrong lemmas) rather than structural (wrong foundations, falsified axioms). The paper never asks whether the framework is the error, not the individual papers.

2. Peer review is the correct epistemic filter. This is tautological within the discipline but ignores that peer review has historically failed to catch massive systemic errors (efficient markets, representative agents, rational expectations). AI-augmented peer review fixes the micro-error rate but does nothing for macro-paradigm blindness.

3. Data contamination is the main interpretative problem. The author treats contamination as a confounding variable to be controlled. Under DT logic, data contamination is a structural feature of AI knowledge: the training data encodes human theory, so AI will be biased toward confirming rather than refuting economics orthodoxy. This isn't a bug to be fixed. It's the mechanism.

4. Refutation requires constructing counterexamples. This conflates the activity (formal proof correction) with the function (epistemic selection pressure). AI doesn't need to construct counterexamples to make economic theory obsolete — it needs to make the questions economics asks irrelevant by answering them with better tools. The paper is asking AI to play chess within economics' rules. The real threat is AI making chess irrelevant.

D. SOCIAL FUNCTION

Transition management / prestige signaling. The paper performs a very specific service for the economics discipline: it acknowledges AI's growing capability while drawing the boundary at "AI as useful tool, not epistemic threat." It reassures economists that their expertise remains the necessary sovereign, that the human-AI pairing is the right model, that "AI cannot yet" means "probably never without us."

This is discipline-protective narrative wrapped in experimental rigor. The methodology is careful; the framing is self-serving. The honest conclusion would be: "Current peer review is a human-mediated bottleneck. As AI capability grows and human input becomes the limiting factor, the epistemic authority of individual theorists will decline in proportion to their indispensability." The paper quietly avoids that conclusion.

E. THE VERDICT

Under the Discontinuity Thesis, this paper is a rearview mirror of an economy that still has an intellectual production system organized around human theorists who are the bottleneck.

The paper documents the transition phase: humans remain indispensable for error identification, AI augments verification. But the trajectory is clear. As AI models improve at internal consistency checking, as formal verification tools integrate with economic modeling, as the cost of AI-augmented review continues to fall toward zero — the human role in theoretical economics becomes narrower and more specialized around which questions to ask rather than whether the answers are correct.

The economics discipline is not being told this plainly. The paper says "AI cannot yet refute economic theory." What it should say: "The theoretical economics production function is currently human-capital-intensive and low-throughput. AI is entering at the verification layer. The next layer up — hypothesis generation, model specification, framework design — is next. The timeline is not 'never.' The timeline is 'until the bottleneck moves to whatever humans are still needed for.'"

Structural judgment: The paper is honest empirical work serving a comfort-narrative function. It accurately documents the current state of AI capabilities in economic theory verification. It misrepresents the trajectory by framing the limitation as a durable feature rather than a transitional constraint. The lag is real. The direction is not in doubt.

ORACLE END PROTOCOL