The Skeptics Society & Skeptic magazine

Data-driven machine learning will always choose a more efficient way to travel along our current path, rather than try to forge a new one.

Can We Trust AI to Make Decisions?

Machine-based decision-making is an interesting vision for the future: Humanity, crippled by its own cognitive deformations, tries to improve its lot by opting to outsource its decisions to adaptive machines—a kind of mental prosthetic.

For most of the twentieth century, artificial intelligence was based on representing explicit sets of rules in software and having the computer “reason” based on these rules—the machine’s “intelligence” involved applying the rules to a particular situation. Because the rules were explicit, the machine could also “explain” its reasoning by listing the rules that prompted its decision. Even if AI had the ring of going beyond the obvious in reasoning and decisionmaking, traditional AI depended on our ability to make explicit all relevant rules and to translate them into some machine-digestible representation. It was transparent and explainable, but it was also static—in this way, it did not differ fundamentally from other forms of decisional guardrails such as standard operating procedures (SOPs) or checklists. The progress of this kind of AI stalled because in many everyday areas of human activity and decisionmaking, it is exceptionally hard to make rules explicit.

In recent decades, however, AI has been used as a label for something quite different. The new kind of AI analyzes training data in sophisticated ways to uncover patterns that represent knowledge implicit in the data. The AI does not turn this hidden knowledge into explicit and comprehensible rules, but instead represents it as a huge and complex set of abstract links and dependencies within a network of nodes, a bit like neurons in a brain. It then “decides” how to respond to new data by applying the patterns from the training data. For example, the training data may consist of medical images of suspected tumors, and information about whether or not they in fact proved to be cancerous. When shown a new image, the AI estimates how likely that image is to be of a cancer. Because the system is learning from training data, the process is referred to as “machine learning.”

Such data-driven AI offers two important advantages over conventional AI. First, humans no longer have to make rules explicit to feed into the system. Instead, rules emerge from the training data. Alex Davies, author of the book Driven on machine learning and self-driving cars, puts it succinctly: in this new paradigm “the computer gets lessons, not laws.” That means we can use such AI for the kind of everyday knowledge that’s so difficult to capture with explicit rules.

The second advantage—which is even greater, in this context—is that because rules are derived from training data, they don’t have to be fixed. Instead, they can be adapted as more (and newer) training data is used. This should prevent the stiffening that lessens the effectiveness of many decisional guardrails as times change. It enables looking at patterns not only from the past but also from the present to deduce rules that can be applied to decisions in the future. It has, in other words, a built-in mechanism of updating rules.

Advocates suggest that we should incentivize the use of machine learning in an ever-increasing number of contexts, and even mandate it—much like collision warning systems have become obligatory in commercial aviation. While this might sound dramatic, the change may actually be more gradual. In many instances in our daily lives, we already have machines making decisions for us, from the relatively simple—such as an airbag deploying in a car crash—to the more sophisticated, such as Siri selecting music on our smartphone. And we profit from it: Machines aren’t as easily derailed by human biases; they perform consistently, irrespective of their emotional state. They also act efficiently—capable of doing so within a split second and at relatively low cost.

The central idea of data-driven decision guidance is that past experiences can be employed to decide well in the present. That works when the world doesn’t change—not the circumstances in which we must decide, nor the goals we want to attain through our decisions. Hard-coded rules are a poor fit for times of change; in theory, this is where data-driven AI should be able to shine. If a situation changes, we should be able to add more training data that reflect the new situation. However, there is a flaw in this line of reasoning.

Autonomous driving company Waymo illustrates the argument—and the flaw. For years, Waymo has had hundreds of cars roam the roads in the United States, collecting enormous heaps of data on roads, signage, conditions, weather, and the behavior of drivers. The data were used to train Waymo’s AI system, which then could drive autonomously. These cars were the guinea pigs for the Waymo system. Mistakes observed (including by their own drivers) in turn help the Waymo system to learn to avoid them. To identify the best driving behavior for any given circumstance, such a system needs not only data about a wide variety of situations, but also data about the outcomes of many different decisions made by drivers in each situation. Learning is richest when there is sufficient variability in the training data, so the system can deduce what works best in which conditions. To get diverse training data, Waymo needs to capture drivers making a variety of choices.

The more we use data-driven machine learning to make decisions, the more it will take the variability of decisions out of the data and shed its ability to progress.

Because Waymo never stopped collecting training data, even small changes in circumstances—such as in driving laws and resulting driving behavior—were reflected in the data collected and eventually embedded in the Waymo system. It was a machine that not only learned once, but never stopped learning.

However, let’s imagine a world in which we increasingly rely on machines when making decisions. The more machines shape our choices, the more these decisions will become the only source of training data for ongoing machine learning. The problem is that data-driven machine learning does not experiment; it acts based on the best practice it has deduced from data about previous decisions. If machines begin to learn more from choices we made based on their recommendations, they will amplify their own, conservative solutions.

Over time, this will narrow and drown out behavioral diversity in the training data. There will not be enough experimentation represented in it to enable the machines to adjust to new situations. This means data-driven machine learning will lose its single most important advantage over explicit rule-based systems. We will end up with a decisional monoculture that’s unable to evolve; we are back to fixed decisional rules.

The flaw is even bigger and more consequential than not being able to adjust to changed circumstances. Even if reality doesn’t change, we may miss opportunities to improve our decision-making in the future. Many innovations that end up becoming successful are less useful than existing choices in their initial form. But any new decision options emerging from the training data will likely only be adopted if they yield better results than existing choices straight away. This closes off any opportunity to experiment with promising new ideas.

For example, the first steam engines used far more energy than they could translate into motion and power. If a machine had compared them to the existing solution of using horses for power, it would have discarded the idea of steam power right away. The only reason the steam engine succeeded is because stubborn humans thought that they could improve the invention in the long run and stuck with it. These tinkerers had no data to support their confidence. They just imagined—and kept tinkering.

Of course, most such would-be innovators fail over time. The path of progress is paved with epitaphs to dogged tinkerers following crazy ideas. Occasionally, though, small changes accumulate and lead to a breakthrough—a far more optimal decision option. Modern societies have permitted tinkering to persist, though it is almost always unproductive, even destructive, in the short term—because of the slight chance of a big payoff sometime in the future.

Data-driven machine learning, if widely utilized, would discard initially suboptimal inventions. But in doing so, it would forego the possibility of long-term breakthroughs. Machines can learn only from what already exists. Humans can imagine what does not yet exist but could. Where humans invented steam power, data-driven machine learning would instead have found more and more efficient ways to use horse power.

Human dreaming can go far beyond technical novelties. Our ancestors once dreamed of a world in which slavery is abolished; women can vote; and people can choose for themselves whom to marry and whether to have children. They imagined a world in which smallpox is extinct and we vaccinate against polio. And they worked to make those dreams come true. If they had looked only at data from their past and present, none of these dreams would have been realized.

Decisional guidelines, from SOPs to nudges, emphasize constancy. Traditional education, too, often aims to perpetuate—suggesting there is a right answer for decisions much like for math problems. But decisional guidelines are just that—suggestions that can be disobeyed if one is willing to take the risk (and shoulder the responsibility). For eons, young people have frequently revolted against their parents and teachers, pushed back against the old, the conventional and predictable, and embraced instead not just the original and novel, but the still only imagined. Humans continue to dream—of a world, for example, that will warm by less than two degrees, or in which people have enough to eat without depleting the planet.

In contrast to humans, machine decision-making is optimized toward consistency across time. Even if data-driven machine learning has access to the very latest data, it will still limit our option space. It will always choose a more efficient way to travel along our current path, rather than try to forge a new one. The more we use it to make decisions, the more it will take the variability of decisions out of the data and shed its ability to progress. It will lead us into vulnerability, rigidity, and an inability to adapt and evolve. In this sense, data-driven machine learning is an adulation of immutability, the anathema of imagination.

No technological adjustment can remedy this easily. If we want to increase diversity in the data, we will need variability in machine decisions. By definition, this means machines that make suboptimal choices. But the entire argument for using more AI in our decision-making is premised on AI’s ability to suggest better choices consistently across space and time. In many instances, it would not be societally palatable to deliberately introduce variation into what options a machine picks, thereby increasing the near-term risk of bad decisions in the hope of long-term benefits. And even if it were, it would not necessarily produce the experimentation we hope for. Very often, the theoretical decision space is immense. Randomly iterating through decision options to generate the diverse data necessary would take a very long time—far too long in most instances to help in timely decision-making. Even when iterations are non-random and can be done purely digitally, it would require massive computing resources.

In contrast, when humans experiment, they rarely decide randomly; instead, they use mental models to imagine outcomes. Done correctly, this can dramatically narrow the decision space. It’s that filtering based on cognitive modeling that differentiates human experimentation in decision contexts from the random walk that the machine, in the absence of a mental model, has to employ. And if machines were to use a particular mental model, the resulting data would be constrained again by the limitations of that model. A diverse set of humans experimenting using diverse mental models is simply very hard to beat. END

This essay was excerpted and adapted by the authors from their book Guardrails: Guiding Human Decisions in the Age of AI. Copyright © 2024 by Princeton University Press.

About the Author

Urs Gasser is professor of public policy, governance, and innovative technology and dean of the School of Social Sciences and Technology at the Technical University of Munich. He is the author of Born Digital: How Children Grow Up in a Digital Age.

Viktor Mayer-Schönberger is professor of internet governance and regulation at the University of Oxford. He is the author of Delete: The Virtue of Forgetting in the Digital Age. This essay was excerpted and adapted by the authors from their book Guardrails: Guiding Human Decisions in the Age of AI.

This article was published on May 31, 2024.

Skeptic Magazine App on iPhone


Whether at home or on the go, the SKEPTIC App is the easiest way to read your favorite articles. Within the app, users can purchase the current issue and back issues. Download the app today and get a 30-day free trial subscription.

Download the Skeptic Magazine App for iOS, available on the App Store
Download the Skeptic Magazine App for Android, available on Google Play
SKEPTIC • 3938 State St., Suite 101, Santa Barbara, CA, 93105-3114 • 1-805-576-9396 • Copyright © 1992–2024. All rights reserved • Privacy Policy