AI’s perfection problem

A magazine of research and ideas from the
faculty of the Kellogg School of Management

Insight in your inbox

Receive our newsletters to keep up with the latest research and ideas from faculty at the Kellogg School of Management.

The Insightful Leader Sent to subscribers on May 27, 2026

Before the rise of ChatGPT sparked the current craze of using artificial intelligence for business, the technology was already transforming science. Though hiding under less-flashy names like “data science” or “machine learning,” the approaches we now call AI were helping astronomers, physicists, and biologists comb through massive datasets much more quickly than before and make unprecedented discoveries.

Science continues to be a pivotal testing ground for AI’s limits. This week, Kellogg’s Julio Ottino and Brian Uzzi ponder whether automated “end-to-end science” can succeed, and we hear what happened when AI turned a scientific field upside down overnight.

In both cases, the debates about AI in science today may foreshadow those in business tomorrow.

Progression without progress

Work in labs has long been moving towards automation, shifting from relying on research assistants to robots to conduct more experiments, faster. But the advent of new AI tools has inspired an even more ambitious vision where AI handles every step of research, from generating hypotheses and running experiments to publishing results.

This concept, of using AI for “end-to-end science,” is “the logical endpoint of current trajectories,” write Ottino and Uzzi in Science. But the two professors of management and organizations argue that fully delegating research to the machines risks losing a critical element of science: imperfection.

“The scientific system thrives on inefficiency: redundant efforts, failed attempts, and divergent paths,” Ottino and Uzzi write. “These are not costs to be eliminated but sources of discovery. By contrast, optimization pressures drive convergence—faster iteration within a constrained search space. The result may be more output but less exploration of the unexpected.”

Science is often a messy process, where the most-robust discoveries emerge from a battlefield of success and failure, criticism and competition. And while AI excels at recombining existing ideas into new ones, major scientific advances are often the result of forward-thinking conceptual shifts, which are less likely to emerge from models trained on the past.

Removing human error and creativity from the research process could thus stifle the search for new knowledge. And the same could be true for business applications of AI, where automation may create a trade-off between increasing efficiency and decreasing innovation.

Processes like end-to-end-science “may produce impressive results,” conclude Ottino and Uzzi. “But if it replaces an evolutionary process with an engineered one, it risks narrowing not only what we discover but what we are capable of discovering at all.”

How one AI transformation unfolded

One of the first areas of science to experience a seismic shock from AI was protein biology, the study of the small cellular machines integral to life.

Determining the 3D structure of a single protein once involved weeks of difficult and expensive experiments—until the AI model AlphaFold2 arrived in 2020. Suddenly, scientists could generate a 3D structure for virtually any known protein in minutes.

Kellogg’s Ryan Hill and collaborator Carolyn Stein wanted to know how this breakthrough affected the work of these highly-specialized scientists. They found that AlphaFold didn’t push away the traditional methods of protein biology; instead, it expanded the field to study a broader range of proteins than before.

“This often happens with technological change,” Hill says. “If a task becomes very cheap through automation, that allows people to do a variety of new tasks they couldn’t do before.”

Thus far, that expansion of research hasn’t translated into accelerated drug development—a process with many human bottlenecks that will be hard to automate. But the early evidence of fruitful human–AI collaboration offers some reassurance to skilled workers in other fields anxious about their future.

“There have been a lot of new technologies that have capably done many tasks that human workers did before, so there’s always worry that people will be negatively affected,” Hill says. “On the other hand, in most past cases of automation, it’s also opened up a lot of new opportunities in the economy, new jobs that appear or changes in the types of jobs that are available. Those types of tools usually make humans more productive.”