Leaning into our Humanity in the Age of AI (Part 1)

How AI Can Boost Productivity, Revive Middle-Skill Jobs, and Recenter the Human Advantage.

Dædalus and the Birth of Automation, as imagined by OpenAI's DALL·E.

Welcome to our new series on AI & Labor—a space to explore how intelligent machines are reshaping the very nature of work.

Over the coming weeks, we’ll share insights on how AI is transforming labor markets—led by our Labor Policy predoctoral researcher, Revana Sharfuddin.

“Leaning into our Humanity in the Age of AI” is Part 1 of this series. Here’s a preview of our upcoming posts:

1. A Market-Driven Approach to AI and Workforce Transformation (Part 2)

2. Who Uses AI, and How? What Millions of Anonymized Conversations Say About AI’s Role in Evolving Nature of Work

3. AI Adoption at Work: Early Evidence from a National Survey

In the ongoing debate within labor economics, a central question is whether AI will ultimately benefit or harm workers. The answer hinges on a crucial distinction: will AI be deployed primarily to automate jobs—by completely replacing human workers—or to augment them, enhancing their productivity and capabilities? The balance between automation and augmentation will determine whether AI acts as a threat to job stability or a catalyst for new opportunities in labor.

We should not be helpless bystanders in this technological shift—we can influence AI’s direction before it defines ours. We can do this by examining early experimental research to gain insights that shape the conversation proactively.

The three key takeaways from research:

1. AI can significantly boost worker productivity—especially for lower-skilled or less experienced employees—by serving as a real-time assistant, not a replacement.

2. Even as automation advances, human traits like judgment, empathy, and creativity remain irreplaceable—and may become even more valuable in an AI-enhanced workplace.

3. If deployed thoughtfully, AI can reverse job polarization by empowering non-experts to perform higher-skill tasks, helping to rebuild the middle-skill workforce hollowed out by past technological shifts.

Is AI Augmenting Human Jobs or Automating Them: The Best Evidence Yet

A recent study by Brynjolfsson, Li, and Raymond finds that a GPT-3-powered AI tool augmented workers' performance, increasing productivity by 15%. Their paper, Generative AI at Work, was just published in the Quarterly Journal of Economics, one of the top journals in economics. This paper offers some of the best causal evidence to date on how generative AI affects workplace productivity and job quality—an area where most prior research has been more predictive than causal. While many studies have estimated AI’s potential impact by categorizing jobs and tasks, establishing true causal effects requires either experimentation, randomization, or well-designed observational studies that leverage quasi-experimental variation. That’s exactly what Brynjolfsson, Li, and Raymond provide.

Their study examines the rollout of a GPT-3-powered AI assistance tool among 5,172 customer service agents at a large (but unnamed) Fortune 500 business software firm. The researchers leverage a staggered implementation—often referred to in empirical work as “staggered rollout”—to compare workers who gained access to AI at different points in time. This allows for a natural difference-in-differences approach, creating a quasi-experimental setting.

Customer service is a strong test case for AI: high turnover, steep learning curves, and pressure from frustrated customers make it ripe for AI support—particularly with real-time response suggestions. Importantly, this wasn’t automation—agents weren’t required to use the AI’s suggestions, and they retained discretion over how to respond. This distinction matters: when AI performs a task that a human otherwise would have done, it forces us to ask whether that task truly requires human input at all. The broader question, then, is not just how AI augments work, but what work remains for humans when AI can handle specific tasks

Example of a real-time response suggestion generated by the GPT-3 chatbot after analyzing a customer complaint.

Here is the breakdown of the 15% productivity gains:

• Agents spend 8.5% less time per chat.

• AI allows for better multitasking, increasing chats per hour.

• Event studies show an immediate boost after adoption.

• After the second month, productivity gains persist for at least five months (likely also end of study observation period).

And here is the most interesting finding: The least skilled agents (bottom quintile pre-AI) saw a 36% productivity jump, while top-performing agents saw almost no improvement. Similarly, new hires (with less than a month of experience) saw an approximately 34% improvement, while workers with more than a year of tenure showed no significant productivity gains.

What do these results tell us? AI doesn’t boost productivity evenly—it provides the biggest lift to those who need it most – less-experienced and lower-skilled workers. This is likely because AI serves as a real-time assistant, filling knowledge gaps, providing decision support, and reducing cognitive overload. Meanwhile, highly skilled and experienced workers, who already operate at peak efficiency, see little to no change.

This is exciting because it suggests AI could help reduce inequality—a topic I will explore more in a later section.

Typically, discussions around AI and labor focus on the risks to low-skilled workers, but this research flips that narrative. Rather than replacing them, AI appears to enhance their productivity, potentially increasing their value in the workforce. That’s a striking shift—one that could create new opportunities for those who have historically been the most vulnerable to technological change.

Of course, this also raises questions about how AI interacts with highly skilled, experienced workers. If their productivity remains largely unchanged while lower-skilled workers become more efficient, we may see some adjustments in wage structures over time. But the broader takeaway is clear: AI has the potential to level the playing field, making economic mobility more accessible.

Strength in our Humanness: Tacit Knowledge, Intuition, Empathy

In 1900, 41% of American workers were in agriculture; by 2000, that dropped to just 2% due to automation and mechanization. Innovations like tractors, assembly lines, and digital tools consistently replaced manual labor with machine precision. Yet, despite this trend, overall employment hasn't fallen. Why? Because automation may reduce labor per task, but it often increases the value of human work elsewhere.

The key lies in the strengths humans and machines each bring. Automation excels at routine, codifiable tasks, while humans thrive with problem-solving, adaptability, and creativity. Tasks requiring judgment or common sense remain hard to automate. A narrow focus on what automation replaces overlooks a fundamental economic dynamic: automation raises the value of tasks only humans can do. That’s why median wages have grown more than tenfold since 1820.

Economic progress has always depended on the interplay between labor and capital, intellect and muscle, precision and judgment. These elements don’t just coexist—they reinforce each other, shaping the evolving landscape of human productivity. When one improves, the others often become more valuable. As automation handles routine tasks, human traits like creativity, empathy, and adaptability matter more than ever.

Think of a disorganized kitchen—orders get mixed up, ingredients aren’t prepped, and dishes come out late— even the best chef can’t succeed there. But when the team runs smoothly, the chef’s expertise shines. Similarly, when AI improves routine tasks, it boosts the value of the human work that remains. When one part of a process becomes more efficient, the remaining tasks often become more valuable as they operate within a more optimized system. This is how AI, when integrated thoughtfully, can drive greater demand for labor.

A great example comes from banking. One study found that ATMs didn’t eliminate tellers—in fact, teller employment rose from 1980 to 2010. By lowering branch operating costs, ATMs indirectly increased demand for tellers: the number of tellers per branch fell, but urban bank branches grew (partly to deregulation on banking branches). Crucially, while routine tasks were automated, tellers shifted to “relationship banking”—advising customers and offering products using empathy and judgment.

Even when productivity shrinks a sector, overall employment may not fall—because income saved is spent elsewhere. When cars replaced horses, jobs didn’t disappear; new industries emerged, like motels and roadside diners. Rising incomes also boost demand in “technologically lagging” sectors, such as hospitality and personal services, which don’t directly benefit from automation. As technology lowers costs, spending shifts to new goods and services. Indeed, from 1940 to 1980, jobs shifted from physically demanding, low-paid work to skilled blue- and white-collar work (graph below).

Some specific takeaways over this time period:

• Agricultural employment fell by nearly 4 percentage points per decade

• Professional and technical jobs grew by 2.5 points

• Managerial roles grew by 0.5.

• Middle-tier jobs remained stable—clerical and sales roles expanded, while manual labor declined.

The data reinforce a clear pattern: as technology advances, employment doesn’t disappear—it evolves.

After the 1970s, however, job polarization emerged as automation increasingly displaced middle-skill occupations involving explicit, routine, and codifiable tasks. Computers excelled at these "routine tasks," reducing employment in clerical, administrative, and repetitive manual roles. Yet the substitution was limited: many human tasks remain inherently tacit—understood intuitively, impossible to fully codify, and thus difficult to automate.

David Autor’s take on Polanyi’s paradox highlights a key challenge for automation: we often perform tasks we can’t explain—like cracking an egg or recognizing a bird. These intuitive, tacit skills are hard to codify, making them difficult to automate. Computers handle logic well but struggle with common sense, judgment, and flexibility. That’s where AI shows promise—not by replacing these human abilities, but by amplifying them. As we saw in the earlier study, AI offered real-time suggestions, but human customer services agents ultimately decided how—and whether—to use them.

Augmenting human capabilities with technology creates vast new possibilities, as illustrated by Erik Brynjolfsson below. Machines perceive what humans cannot, act in ways humans physically cannot, and grasp concepts humans find incomprehensible. More profoundly, technologies that help humans invent better technologies not only enhance our collective capabilities, but also accelerate the rate at which these capabilities grow.

For instance, while AI may outperform radiologists at reading mammograms, it can’t handle the other 26 tasks required for the role—like comforting patients or coordinating care. In fact, across 950 occupations, Erik Brynjolfsson and coauthors found that although AI could perform certain tasks in nearly all jobs, there was no occupation that machines could fully handle alone. David Autor’s research also shows that as tech augments work, it creates new roles: across eleven of twelve broad occupational categories increases in augmentation-oriented innovations strongly predict the emergence of new occupational titles (figure below). About 60% of today’s jobs didn’t exist in 1940—a reminder of technology’s profound capacity to reinvent and expand human labor

Can AI Be a Force for Good in the Battle Against Job Polarization?

Although literature of income inequality in the US lack methodological consensus, it is clear from the patterns discussed in the sections above that technological shifts since the 1970s have contributed significantly through job polarization. Earlier waves of technological change automated a broad range of middle-skill occupations—jobs in administrative support, clerical work, and blue-collar production—pushing many of the approximately 60% of American adults without a bachelor's degree into low-paying, low-skilled service jobs.

A central reason for this widening gap is that information alone is merely an input for a much more valuable economic activity: decision-making. Since 1970s decision-making power has increasingly become concentrated among a relatively small group of highly educated professionals, typically those with college or graduate degrees. The widespread availability of cheap computational power amplified this concentration, dramatically raising the economic value of elite expertise at the expense of middle-skill workers.

Yet the unique promise of AI lies in its ability to reverse this pattern and broaden decision-making beyond a narrow elite. By combining information, explicit rules, and learned experience, it can amplify human judgment and enable workers with basic training to take on tasks once reserved for doctors, lawyers, or engineers. Rather than replacing expertise, AI can democratize it—revitalizing the middle-skill, middle-class core hollowed out by automation and globalization.

While some fear AI will replace human expertise, historical and economic patterns suggest otherwise. Like calculators or machines, AI tends to amplify, not substitute, human skill. Indeed, this mirrors the shift from artisanal work to mass production in the 18th and 19th centuries. Though craftspeople were displaced, new roles emerged—electricians, machinists, telephone operators—each requiring new forms of expertise built on literacy, numeracy, and training. As tools evolved, so did the value of mastering them.

However, the rise of the Information Age and the widespread use of digital computers after WWII reduced demand for “mass expertise”—the middle-skill jobs that powered the industrial age. Computers automated routine tasks, hitting middle-skilled roles hardest. Meanwhile, high-paid professionals like doctors and engineers remained insulated, as their work relied on judgment and creativity—qualities machines couldn’t replicate, as Polanyi’s Paradox suggests. This shift boosted elite productivity but displaced the middle class. Ironically, low-paid service jobs—like cleaning and caregiving—survived, deepening job polarization and income inequality.

AI holds real promise to reverse this trend. Unlike past automation, AI can augment human judgment instead of just replacing routine tasks. For example, an MIT experiment showed that giving writers tools like ChatGPT significantly improved both productivity and quality—especially among lower performers. Similar results appeared in the earlier customer-service study: AI boosted lower-skilled workers' output, narrowing gaps with experienced peers.

Crucially, AI didn’t replace expertise—it complemented it, requiring workers to apply judgment alongside AI’s suggestions. Thus, rather than deepening inequality, AI could democratize expertise, allowing a broader range of workers to participate meaningfully in tasks involving judgment, creativity, and decision-making—ultimately revitalizing the middle-skill workforce.

This snippet from the Macro Musings podcast with Dean Ball highlights how AI could challenge credentialism in academia. Ivy League professors (we know from literature are disproportionately drawn from university-educated families and higher socio-economic backgrounds) often have an army of research assistants, allowing them to outproduce skilled professors at mid-tier institutions (who probably do not share the same background characteristics). But what if all professors had access to their own AI-powered RAs? And what if human RAs had AI RA of their own? Deep research is exactly doing that. AI has the potential to level the playing field, expanding research capacity beyond elite institutions and democratizing academic productivity.

Conclusion

The future of work will not be defined solely by the power of AI, but by how we choose to use it. If deployed to augment rather than automate, AI can unlock human potential, expand access to opportunity, and rebuild the middle of the labor market. The evidence is clear: when paired with human judgment, AI improves outcomes—especially for those who have historically lacked access to expertise or upward mobility.

But realizing this promise will require intentional choices—from businesses, policymakers, and society as a whole. ​Rather than fearing disruption, we must embrace augmentation, foster resilience, and remain optimistic about the inclusive opportunities it presents.​ If we do, AI won't just change how we work—it will elevate what it means to work at all.