AI Helps Unskilled Employees on the Job, Hurts Experienced Workers

The past few months I’ve been mulling over a series of studies economists have conducted on the value of artificial intelligence in the workplace. How much, they wanted to know, does AI help white-collar professionals do their jobs? The productivity gains they’ve observed are substantial: AI is clearly making us better, faster workers. The numbers have prompted AI optimists to predict an economic boom and AI pessimists to worry about a future of fewer jobs.

But behind those numbers, buried a little deeper in the studies, is the finding that interests me. The question isn’t how much AI helps out around the office but who it helps — and why.

AI, the studies indicate, is making us more productive in a weird way. It’s not helping everyone get better at their jobs. It’s mostly turbocharging workers who are bad at their jobs, while doing little to aid — or even hindering — those who are already productive to begin with. AI, in other words, is raising overall productivity by narrowing the gap between high performers and low performers. It’s equalizing white-collar work — a vast swath of the economy that has always been predicated on the assumption that some people will inherently be much, much better at their jobs than others.

Before we get into the broader implications of the studies, let’s start by reviewing their findings. Economists looked at the impact of AI in six different areas of work:

Creative writing. Researchers tasked people to write a short story, with and without the help of an AI tool for generating ideas. Those who had no spark of their own became as much as 11% more novel and 23% more enjoyable with the help of AI. But the tool didn’t benefit those who were already creative on their own.

Office memos. Researchers had subjects complete writing tasks that are common in professional jobs — think press releases, short reports, delicate emails. Access to AI made everyone faster, regardless of their skill level, by an average of 37%. But when it came to the quality of their writing, AI mostly helped the low performers.

Coding. Software engineers with fewer years of professional coding experience benefited much more from access to GitHub Copilot, an AI coding assistant, than veteran coders did.

Management consulting. Researchers graded professional consultants on 18 knowledge-intensive tasks similar to what they actually do in their jobs. Access to GPT-4 boosted the scores of low performers by 43%, compared with only 17% for high performers.

Law school. Researchers administered an exam to law students with and without GPT-4. Students at the bottom of the class got a big performance boost. But access to the tool actually hurt the grades of the students at the top of their class.

Call-center work. Researchers measured the effects of a tailored AI tool that was introduced at a real call center. Novice and low-skilled workers became 34% more productive, while those with more experience and skill saw few benefits. Access to AI even slightly hindered the top performers on some measures, like conversation quality.

So yes, AI boosts productivity in a wide variety of common office tasks, from repetitive work in low-paying call centers to complicated duties at elite management firms. And though most of the studies were hypothetical experiments in a lab — making their findings difficult to extrapolate to the real world — the call-center study looked at actual job performance at an actual company. But it’s how AI increases productivity that should interest us the most. Together, the studies present a strong case that by disproportionately boosting those at the bottom, this new generation of AI tools is narrowing the variation in job performance. In just a few short months, it’s already doing what decades of education have failed to do — it’s equalizing the American workplace.

When you stop to think about how large language models work, this finding makes sense. LLMs basically regurgitate what worked before — something the low performers can learn a lot from, but stuff the high performers already know. If you give everybody a cane, it’ll speed up the slowest walkers the most. But it won’t do much for Usain Bolt — and it might even slow him down.

This runs counter to how we’re used to thinking about technology in the workplace. Over the past few decades, new technologies like industrial robots, the personal computer, and the internet have disproportionately aided highly skilled workers with college degrees, but they’ve done little to help (or, depending on who you ask, screwed over) those with fewer skills and less education. Economists call this skill-biased technological change, and it’s a big reason income inequality has grown so much since the 1980s.

Which brings us to the broader implications of the studies. If AI boosts the productivity of low performers, putting them on equal footing with the superstars, how is that going to change professional work as we know it?

One possibility is that AI could help reverse America’s growing chasm of income inequality. Some of the inequality we see today is a result of the huge gaps in salary within many elite professions — of a superstar software engineer, say, who can churn out thousands of lines of code in the blink of an eye, compared to an average-performing techie. Presumably, the superstar gets paid more because they’re so much better at their job than everyone else. But if AI makes it so that every coder can blaze away, it’ll be a lot harder for the hotshots to justify their astronomical salaries.

This is something the law-school study touches on. “The legal profession has a well-known bimodal separation between ‘elite’ and ‘nonelite’ lawyers in pay and career opportunities,” the authors write. “By helping to bring up the bottom (and even potentially bring down the top), AI tools could be a significant force for equality in the practice of law.”

But the true promise of AI lies in narrowing inequalities not within occupations, but between them. Software developers in the United States make, on average, 5.5 times more than fast-food workers. If AI makes it easier for a fast-food worker to move into a coding job, that’s when we’ll really start to see the income gap shrink. The GitHub Copilot study hinted at that: It found that the tool benefited novice programmers much more than expert ones. That could lower the barrier to entry for a whole new generation of aspiring engineers.

If you’re already one of the highly paid coders, this probably won’t come as good news. Part of the reason programmers are paid so much is because there are so few of them. By allowing tons of people to flood into the occupation — and by turning crappy coders into decent ones — AI will almost certainly depress the sky-high salaries of those at the top of the profession. Education and expertise won’t count for as much as they used to.

Admittedly, this scenario I’ve laid out is an optimistic view of how AI will affect salaries. If it helps raise the skill level of subpar coders, then it will also raise their pay, right?

Not necessarily. There’s another way AI could reduce wage inequality: It could depress the pay of top earners without doing much to raise wages for those at the bottom. As productivity goes up, owners might opt to pocket the gains for themselves, lowering the salary ceiling rather than raising the salary floor. In this scenario, we’ll have less income inequality thanks to AI. But we’ll all make less.

By commodifying the talents of the best illustrators, AI lowered their pay — the same way mechanized looms destroyed the livelihoods of artisan weavers in the Industrial Revolution.

Unfortunately, that seems to be how AI is affecting the job market so far. In one study, researchers looked at what happened to freelancers on the online platform Upwork after the introduction of AI tools like ChatGPT. The number of jobs on the platform declined, and so did incomes. Those who were earning the most suffered the biggest hit. The top freelancers among those who offered image-based services received 7% fewer jobs and watched their earnings tank by a staggering 14%. In economic terms, AI isn’t upskilling the workforce — it’s deskilling it. By commodifying the talents of the best illustrators, it lowered their pay — the same way mechanized looms destroyed the livelihoods of artisan weavers at the onset of the Industrial Revolution. And AI systems are doing it, ironically, by feeding off the experience of the top performers, whose work provides the datasets they’re trained on.

The implications of these findings could go far beyond the question of pay and opportunity. We’ve organized much of white-collar work around the idea that there’s a huge variation in both the quality and the quantity of work people produce. The entire idea of professionalism, in a sense, is predicated on the notion of talent. Some people are just really good at their jobs, the thinking goes, and it’s worth throwing a lot of money at them to get them to work for you. That’s why we receive raises for accumulating degrees and experience and expertise. And it’s why companies have developed complicated performance-management systems to weed out the lower performers and to reward, retain, and promote the superstars.

But if AI leads to a world in which employers get more or less the same work from everyone — regardless of schooling, years on the job, or inherent talent — that opens up all kinds of wacky possibilities for the future of work. Will companies start paying everyone in a particular job the same salary, regardless of their seniority level? Will promotions be a thing of the past? How will we raise our families and save for retirement if there’s no opportunity for salary progression? Will HR departments dispense with time-consuming performance evaluations altogether? And if managers currently spend most of their time coaching, cajoling, and managing out their bottom performers, what happens to their jobs when there are no more bottom performers left?

If we take the recent studies of AI at face value, the smart move for employers would be to hire the novices at cheap salaries and get rid of the veteran superstars who are earning the big bucks — implementing a “Moneyball”-style arbitrage for the ChatGPT age. But the thing is, I’ve spoken to a lot of executives over the past year about how they’re rethinking their staffing plans, and not a single one has talked about scrapping their hotshot earners. In fact, many of them have told me, in private, that they intend to do the exact opposite. They’re aiming to hire fewer entry-level people straight out of school, since AI can increasingly take on the straightforward, well-defined tasks these younger workers have traditionally performed. They plan to bulk up on experts who can ace the complicated stuff that’s still too hard for machines to perform.

If I had to guess, though, I’d say that trend won’t last. A few enterprising employers will go all in on hiring job candidates with less experience and boosting their performance with AI. They’ll save boatloads of money on salaries, and from there the practice will inevitably spread. That will open up all kinds of opportunities for professional wannabes to get their foot in the door. But for white-collar veterans, I suspect an onslaught is coming — one in which being good at your job will no longer offer the protections and perks it once did. When it comes to professions like law and management, talent has long been considered a ticket to success, deserving of rich rewards. Now, in the era of AI equality, it may turn out to be a costly liability.

Aki Ito is a senior correspondent at Business Insider.