AI Is Great at “Normal Science,” but Breakthroughs Still Belong to People

Artificial intelligence can translate menus, analyze massive datasets, and even outperform humans on certain scientific tasks. But according to visiting University of Virginia Darden School of Business professor Keith McCormick, it isn’t likely to invent the next big idea.

McCormick, the 2025–2026 Bodily Bicentennial Professor of Analytics at Darden, has spent more than three decades building predictive analytics models across industries. He has authored seven books on the effective use of predictive analytics and taught LinkedIn Learning courses reaching more than one million learners.

In a recent capstone presentation, he offered a grounded perspective on what AI can and cannot do today. Today’s systems are powerful, he said, but their limits are becoming clearer. According to McCormick, AI excels at what science philosopher Thomas Kuhn calls “normal science,” but may fall short of reaching artificial general intelligence (AGI) without some new scientific breakthrough that changes the current paradigm.

The Hidden Value of Low-Code Tools

Keith McCormick is the 2025–2026 Bodily Bicentennial Professor of Analytics at Darden.

Before turning to AI’s future, McCormick focused on an example of how organizations are using AI right now and often misunderstanding what it’s valuable for.

One unexpected insight from his work comes from low-code or no-code tools, or using visual tools like flow charts to generate code and workflows. These tools are often dismissed, says McCormick, as “not for serious data scientists.”

But McCormick pointed out an underrated use for low-code tools: governance.

McCormick discovered this underappreciated use of low-code or no-code tools in his work with Citibank and a low-code tool called KNIME Analytics Platform. The traditional wisdom might be that a corporation like Citi would frown on the use of low-code tools, out of concern for “shadow IT.” But by training the finance team on KNIME, McCormick and team found that the tool actually reduced risk and unauthorized IT work-arounds.

In practice, it turns out, these tools can replace chaotic, hard-to-audit Excel processes that are common in business units trying to solve problems without enough technical support.

“It’s not really a choice between some low-code tool that the data team is unfamiliar with or Python. It’s either Excel—and they really don’t tell anybody how they did it—or it’s a tool like this,” said McCormick.

The low-code tool creates a shared language between business teams, data scientists, and managers, making it easier to understand and trust how decisions are made.

“Who knew that low-code/no-code was basically a data governance play?” said McCormick.

The AGI Question: Scale or Breakthrough?

Zooming out, McCormick addressed one of the most debated questions in AI today: how—or whether—we will reach AGI.

While opinions vary widely, he argues the debate boils down to two camps. One believes that scaling up current models will eventually lead to general intelligence. The other believes that scale alone is not enough, and that a new scientific breakthrough is required.

McCormick falls into the second camp.

To explain why, he draws on Thomas Kuhn’s seminal work “The Structure of Scientific Revolutions,” which distinguishes between “normal science” and paradigm-shifting breakthroughs.

Normal science, he explained, is the day-to-day work of improving existing ideas—testing, refining, and scaling what already works. “It’s really important, and it’s what scientists spend most of their time doing,” he said.

Today’s AI systems are exceptionally good at this kind of work. They can process vast amounts of information, identify patterns, and generate useful outputs across a wide range of tasks. But McCormick is skeptical that the current paradigm of AI has what it takes to achieve what might be called “revolutionary science” in Kuhn’s language, or intuitive and creative leaps that lead to solving problems in entirely new ways.

Signs of a Limit

If the current approach is reaching its limits, how would we know?

McCormick points to the accumulation of “anomalies,” or problems that should be solvable within the current paradigm, but aren’t.

“You’re going to know that you have an anomaly when there are these kind of mysterious puzzles, and they start to accumulate and get distracting,” he said.

Some of these limitations are already familiar to everyday users. Large language models can produce impressive results, but they often forget earlier parts of a conversation, struggle with spatial reasoning, or generate answers that sound confident but miss the mark.

Even more telling, McCormick noted, is how users respond to these failures.

“When something goes wrong, we tend to blame ourselves,” he said. “We think, ‘I must have written a bad prompt.’”

That instinct, he suggested, reflects a moment when confidence in the current approach is still high, even as cracks begin to show.

He also pointed to emerging skepticism among leading AI researchers, some of whom have begun to question whether scaling alone will continue to deliver major advances. The implication is not that AI is failing, but that it may be approaching the limits of its current design.

What Comes Next

For McCormick, this does not diminish the significance of today’s AI systems. If anything, it highlights how far they have come. Systems that can perform “normal science” at scale are already transforming industries.

But the next leap forward, he believes, will require something different.

“There will be enough things that it can’t do that we’ll need something new,” he said.

That breakthrough may come in the next few years, or it may take longer.

For now, McCormick’s position is that AI is powerful, but it is not complete. And the kind of insight that drives true innovation—the ability to approach problems in entirely new ways and imagine new possibilities—remains uniquely human.

Watch the full presentation here: