AI’s Paradox: The Best ‘Normal Science’ Engine Ever Built

Thomas Kuhn’s 1962 framework for scientific revolutions might be the sharpest lens we have for understanding what AI can and can’t do. But the analogy cuts both ways — and that’s where things get interesting.

If you want to understand the current state of AI without reading another benchmark comparison, start with this: visiting professor Keith McCormick at UVA’s Darden School of Business recently argued that today’s AI systems are the greatest “normal science” engines humanity has ever built — and that this is precisely why they’ll never reach AGI without a paradigm shift.

McCormick’s full argument, including his low-code analytics insights from his work at Citibank, was reported by Darden News on May 4th. The framework is worth thinking about. But I think the analogy is cleaner on one side than the other — and that asymmetry is what actually matters.

Kuhn’s Two Kinds of Thinking

Thomas Kuhn’s The Structure of Scientific Revolutions is one of the most cited books in the humanities. Its core distinction is between two modes of intellectual work:

Normal science is puzzle-solving within an accepted framework. You take the paradigm as given, apply it, refine it, extend its reach. This is where most scientists spend their careers. It’s cumulative, rigorous, and incremental.

Revolutionary science is when the framework itself breaks. Anomalies pile up until they can’t be ignored, and someone proposes a radically different way of seeing the problem. Copernicus to Kepler. Mendel to the molecular biologists. Not a new fact — a new question.

McCormick’s move is to ask: what kind of intelligence is AI? His answer: the most powerful normal-science engine ever created. It reads every paper on a subject, synthesizes the findings, generates hypotheses within the existing paradigm, and tests them at speed no human team could match. It’s a PhD student with infinite patience, perfect recall, and zero sense of whether the question is worth asking.

Where the Analogy Holds

McCormick is right about at least three things.

First, LLMs are pattern synthesizers, not idea generators. Given enough text on a topic, an LLM will produce a coherent literature review, identify contradictions between papers, and even suggest new experiments. But it does all of this by recombining existing structures. It doesn’t sit down and ask a question that no one has asked before because the existing framework doesn’t allow it. That’s the definition of a paradigm-bound system.

Second, the accumulation of anomalies is a real signal. When you’ve been told that scaling will solve everything, it’s worth paying attention to the failures that scaling doesn’t fix. Confabulation persists at larger model sizes. Spatial reasoning remains fragile. Long-context windows still suffer from attention degradation. These aren’t bugs — they’re symptoms of a system whose fundamental mechanism (next-token prediction) has known boundaries.

Third, the “prompt blame reflex” is a genuine psychological phenomenon. When a system is mostly right, users naturally attribute failures to their own input quality rather than systemic limitations. This is what Kuhn called high confidence in a paradigm — the cracks are noticed, but the framework still feels right.

Where the Analogy Breaks

Here’s where I push back.

Kuhn assumed that revolutionary science requires a human mind. His framework was about how scientists work, not how tools assist them. But there’s no law saying that a tool that excels at normal science can’t also catalyze revolutionary science in someone else.

Consider: the telescope didn’t invent heliocentrism — Copernicus did. But the telescope made the paradigm shift visible. Could AI be the telescope for a cognitive paradigm shift? We don’t know yet. The analogy works when AI is the scientist. It gets murkier when AI is the instrument.

Second, Kuhn’s anomalies were conceptual — not statistical. The pre-Copernican anomalies weren’t “sometimes the models predict wrong.” They were mathematical contradictions that couldn’t be resolved within the existing framework. AI’s anomalies are mostly performance gaps, not logical impossibilities. That doesn’t mean they’ll be resolved by scaling — but it does mean the crisis pattern might not play out the way Kuhn described.

Third, and most importantly, the “breakthrough camp” is assuming that revolutionary science requires a fundamentally new approach. But what if the next leap comes from combining normal-science tools in ways we haven’t imagined? Agentic workflows. Multi-modal reasoning. Causal inference layered on top of pattern matching. These aren’t new paradigms — they’re new compositions of existing ones. And Kuhn didn’t have much to say about whether paradigm shifts can emerge from tool combination rather than theoretical revolution.

The Low-Code Insight We’re All Missing

McCormick’s piece spends time on low-code analytics, and I think it’s the most practically valuable part — even though it has nothing to do with AGI.

The insight is simple: transparency beats convenience every time in organizational decision-making. Excel is fast but opaque. KNIME (or any visual workflow tool) is slower but auditable. In a regulated environment, the ability to show your work is more valuable than the ability to produce a result.

Applied to AI, this suggests something important about evaluation: we should be measuring interpretability and auditability as seriously as we measure accuracy. An AI that’s 90% accurate but can explain its reasoning in a way that humans can verify is more valuable than a black box that’s 95% accurate. This is governance thinking, not research thinking — but it’s where the real organizational impact is happening right now.

So What Kind of Intelligence Are We Building?

Here’s my take, which I think McCormick’s framework points toward but doesn’t quite articulate:

We’re not building normal science engines. We’re not building revolutionary science engines. We’re building something that doesn’t fit Kuhn’s categories at all.

What LLMs are good at is synthetic competence — the ability to absorb, recombine, and regenerate knowledge across domains at a scale that no human or human team can match. This is neither normal science nor revolutionary science in Kuhn’s sense. It’s a third kind of intellectual labor.

And that’s actually more interesting than the binary framing. Because if we stop asking “is AI normal science or revolutionary science?” and start asking “what is synthetic competence, and how does it change the epistemology of knowledge production?” — that’s a question worth thinking about.

The answer might not require a paradigm shift. It might require a new category.

Quick Quiz

1. Kuhn’s “normal science” is best described as: A) AI doing research B) Puzzle-solving within an accepted theoretical framework C) Teaching introductory science courses D) Applying quantum mechanics to computing

2. The author’s main critique of the Kuhn-AI analogy is: A) Kuhn’s book is too old to apply to modern technology B) The analogy assumes revolutionary science requires human minds, which may not hold for AI as an instrument C) AI is actually worse at normal science than Kuhn’s framework suggests D) Thomas Kuhn never wrote about science

3. The author’s “third kind” of intelligence is: A) Quantum computing B) Synthetic competence — cross-domain knowledge absorption and recombination at scale C) Human-AI hybrid reasoning D) Evolutionary algorithms

---

Answers: 1-B, 2-B, 3-B

Original framework reported by Keith McCormick at UVA Darden, as published in Darden News, May 4, 2026.