The Human Element: Machine Learning Cannot Replace Critical Thought

In the past decade, technology has become indispensable for crunching mind-boggling volumes of information. The number of businesses using machine learning — and the amount of data their computers process to enable that learning — has exploded.

But Dokyun “DK” Lee, Assistant Professor of Business Analytics at the Tepper School of Business, warns that businesses must always remember that machine learning works best as a tool, and that the art of human interpretation and reasoning cannot be replaced by artificial intelligence.

At issue is the need for human interpretation of artificial intelligence outputs to make them easier to understand and to investigate further for more robust “causal inference.”

“Machine learning, if sufficiently transparent, is great for helping humans discern patterns and understand the nuance of complex big data,” Lee explains. “But inherently, making connections about how the world works and creating hypotheses to find answers that stand the test of time — we (humans) are the only ones who can do that, at least for now.”

For few decades at least – and especially around 2010 when a sub-area of machine learning called deep learning began to thrive and scale – businesses quickly recognized the potential power of harnessing large and raw data sets and began implementing systems that at first blush appeared to be a revolutionary panacea. The algorithms on which these systems are based learn by example, not by rules or data features that were engineered by humans.

Instead, the algorithm recognizes patterns in data and makes decisions based on those patterns. But the machine’s decision-making process is not something people can always understand, because they are occurring outside of human control, and the method the machine is learning for drawing its inferences is not readily comprehensible to people.

The result is that the computer’s decision could be wrong. Lee offers the example of a vision algorithm that looked at raw images of animals and classified them as either a husky dog or a wolf. The algorithm incorrectly interpreted a husky as a wolf whenever it was photographed in snow, because during its training data, the wolf image included snow.

But sometimes, the rationale for the decision is more difficult to deduce, and the computer’s misinterpretation can be harmful. For example, in 2008, Google introduced a tool that tracked flu outbreaks based on people searching for flu-related topics. For a few years, it worked well, creating accurate predictions weeks earlier than the U.S. Centers for Disease Control — until it didn’t. The tracker missed the peak of the 2013 flu season by 140 percent and was quietly retired.

Algorithms can incorporate racial bias into its functions and discriminate against women, misdiagnose diseases, and promote the wrong products. And because the way the machine got from point A to point B is not transparent to humans, trying to undo the faulty learning could be difficult and costly, Lee points out. 

These problems will worsen if businesses continue to make decisions based solely on predictive analytics, which are often opaque. Lee suggests instead that people must scrutinize the artificial intelligence outputs first and then test them with theory-driven causal inference when the stakes are not negligible— a process that requires “explainable AI” in the first step for inspecting AI outputs.

“Many businesses do not bother with algorithmic transparency and causal inference because they are often harder and slower than opaque predictive analytics. Sometimes that could be okay,” Lee says. “But in many cases that cheaper shortcut hurts in the long run. Including the algorithmic transparency and human guidance might sacrifice some profit or predictive power which results in inefficiencies, but it is definitely worth the effort to avoid the consequences.”