Why AI Is a Terrible Human Replacement

In a recent interview, Nvidia CEO Jensen Huang remarked that it made little sense for AI, in its current state, to already be responsible for widespread job losses. He criticized CEOs who attribute retrenchments to AI as being "just too lazy," arguing that many of these workforce reductions were likely planned long before generative AI became genuinely useful. He's not the only tech leader (and at least one economist) to make similar comments, some of whom have since tempered earlier predictions of a looming "jobs apocalypse".

It's tempting to simply hotswap manpower with technology, and go from managing people to feeding prompts. This assumes that an employee's value lies solely in the most visible task they perform, a.k.a. the doorman fallacy. A doorman appears to open doors, yet he also provides wayfinding, security, reassurance, local knowledge, and a welcoming first impression. Despite doors being automated for decades, doormen still exist in high-end establishments to greet and coordinate assistance the very moment guests arrive.

This phenomenon was actually captured in 1997 by the human factors researchers at The Ohio State University. The "substitution myth"* is the pervasive and flawed corporate belief that a robot can seamlessly substitute for a human. Clearly there's more than meets the eye when it comes to transforming the workplace using AI.

Rick and Morty, Season 8, Episode 7. Source

Humans and intelligent technologies should not be viewed as separate actors. Real-world work emerges from Joint Cognitive Systems, networks of people, technologies, and organizations continuously coordinating toward shared goals. A surgeon does not perform surgery alone. Even the most skilled experts depend on others to monitor, anticipate, communicate, cross-check, coordinate, and adapt as conditions change. Organizations do not succeed because individual tasks are completed efficiently. They succeed because people coordinate those tasks effectively toward shared goals. AI may increasingly contribute computation, but coordination remains the harder challenge.

Successful teams also reflect high reciprocity, the ability of one actor to recognize when another is approaching the limits of their capacity and to adjust accordingly. The scrub nurse quietly monitors the surgeon in action and only offers the next instrument when the surgeon appears ready. As things get complicated, the assisting surgeon may jump in and take on more tasks to help an overwhelming lead surgeon. High-performing teams continuously redistribute work, attention, and resources in response to changing demands. Can AI appropriately adapt their behavior as their human are overloaded, uncertain, or struggling?

This also forces us to rethink what expertise actually means. Expertise is often perceived as the ability to produce the right answer, diagnose the problem, or execute a task. In reality, we appreciate experts for their ability to recognize when situations are unusual, anticipate downstream consequences, adapt when plans unravel, and help others make sense of ambiguity. Experts are not necessarily those who calculate the objectively best answer every time. In time-sensitive and uncertain situations, they excel by recognizing patterns quickly and generating context-sensitive responses before conditions deteriorate. AI will increasingly contribute answers. Achieving true expertise, however, remains fundamentally tied to helping a larger system function under uncertainty.

In our partnership with AI, I often describe AI as a teammate from a foreign land, or an alien from another planet. It speaks our language, yet often reasons differently. It can be brilliant in situations that stump experts and astonishingly naïve in situations a novice would navigate with ease. The irony is that as AI grows more capable, human expertise grows more important. Someone must still provide context, recognize when the AI is wrong, and ensure its contributions make sense within the larger system. AI's greatest limitation may not be intelligence, but its inability to fully participate in the coordination that makes teams successful.

Source

Perhaps the core argument supporting AI adoption in today’s workplace is its ability to complete tedious computerized tasks in seconds or minutes. Employees are then freed to focus on activities that require judgment, creativity, relationship-building, and other distinctly human capabilities. In theory, this should reduce manpower, or does it? The proliferation of self-checkout counters and self-service kiosks did not trigger a wave of job losses. Instead, organizations often responded by raising expectations, expanding services, handling greater volumes of work, or pursuing opportunities that were previously out of reach.

Far from making human expertise redundant, AI will make it more important. This creates a learning paradox as the routine tasks most suitable for automation are often the same tasks through which people acquire the experience needed to handle more difficult situations. Think about automated driving and parking cars, and their potential impact on the future generation's driving skills and abilities to respond quickly on the road. Genuine expertise develops in environments where recurring patterns can be encountered repeatedly and where decisions generate timely feedback. Yet these regular, repetitive environments are precisely the first to be automated.

Like the internet before it, generative AI will transform how work is done. Aunties who once staffed cashier and registration counters benefited from deliberate job redesign and gaining the competencies to oversee multiple kiosks, assist customers who needed help, resolve exceptions, and manage the small frictions that automation could not handle on its own. AI adoption deserves the same care. Organizations should examine its place within the broader joint cognitive system, redesign jobs around the new division of work, and enhance the pathways through which people acquire and sustain expertise.

Borrowing Prof. Teo Eng Kiong’s analogy, it would be misguided to ask why deliveries have not become dramatically faster after handing a delivery driver a Ferrari to navigate the roads of Singapore. The performance of any technology depends on the wider system in which it operates. The goal should not simply be to automate more work, but to build systems in which humans and AI remain capable of handling the situations that automation alone cannot anticipate.

Source

*Random HFES fun fact: the substitution myth was first proposed by Nadine Sarter, when she was completing her Ph.D. under David Woods

Read my previous commentary that featured more concepts surrounding human factors and automation, or check out The Ohio State University's Cognitive Systems Engineering Lab.