Why I'm not worried about AI causing mass unemployment. By Timothy B Lee

www.understandingai.org

Why I'm not worried about AI causing mass unemployment

Timothy B Lee

18 - 23 minutes

In 2011, the venture capitalist Marc Andreessen published an essay that became a kind of manifesto for Silicon Valley during the 2010s.

“Software is eating the world,” Andreessen declared.

Computers and the Internet had already revolutionized a bunch of information-oriented businesses: books, movies, music, photography, telecommunications, and so forth. Software also played a major supporting role in more tangible industries. New cars had dozens of computer chips in them, for example, and the oil and gas industry made heavy use of software to discover new drilling sites.

But Andreessen, co-founder of the venture capital firm Andreessen Horowitz, argued that the software revolution was only getting started. “In many industries, new software ideas will result in the rise of new Silicon Valley-style start-ups that invade existing industries with impunity,” Andreessen wrote. “Companies in every industry need to assume that a software revolution is coming.”

At the time, Silicon Valley was abuzz with talk of the sharing economy. In 2011, Andreessen Horowitz invested in Airbnb, which was working to disrupt hotels by having people rent out their spare rooms. In 2013, the firm invested in Lyft, which was trying to revolutionize the taxi industry by letting people give rides in their personal vehicles.

When Bitcoin started to gain mainstream attention in 2013, Andreessen Horowitz jumped on that bandwagon too. Cryptocurrency was supposed to “eat” the financial world, rendering banks and other financial institutions irrelevant in much the same way Netflix made Blockbuster irrelevant and digital cameras bankrupted Kodak.

The venture capitalists who poured billions of dollars into startups like this during the 2010s expected some of them to eventually be as big as Google or Facebook. After all, they thought, industries like hospitality, transportation, and finance are huge. The rewards for disrupting them should be correspondingly large.

But it hasn’t worked out that way. Computers and smartphones have become ubiquitous across the economy. But this has led to only modest changes for established industries like health care, education, housing, and transportation.

Andreessen’s essay reflected a persistent blind spot in Silicon Valley thinking: a tendency to overestimate the power of information technology and underestimate the complexity of the physical world. In 2011, this led to excessive optimism about the economic impact of software startups. Today I suspect this same bias is distorting many people’s thinking about the likely impact of artificial intelligence.

Many technologists worry that AI will get so powerful that it will be capable of performing most of the jobs currently performed by humans, leading to mass unemployment. I don’t buy it. Certainly AI will have a significant impact on the economy—perhaps even bigger than the Internet. But there will also be significant sectors of the economy that see only modest changes as a result of AI. And there will continue to be plenty of work for human beings to do.

One way to evaluate Andreessen’s 2011 prediction is to look at the most successful software startups of the 2010s. With help from Connor Leech, the CEO of job search website Employbl, I made a list of successful Internet startups founded since 2009. They can be broken down into a few major groups:

There are social networks and messaging apps like Discord, Instagram, Slack, Snap, TikTok, Whatsapp, and Zoom. The success of these companies doesn’t really bolster the “software eating the world” thesis because they were entering fields already dominated by other tech companies.

Companies like Square, Stripe, Robinhood, and Venmo have thrived by offering modern interfaces for traditional financial services. They represent incremental progress in the finance sector, not a revolution.

In contrast, cryptocurrency companies like Coinbase and Circle are trying to build new payment rails that will eventually make conventional financial institutions irrelevant. But these firms have struggled, especially in the last year, and crypto-based financial products remain far from mainstream adoption.

The startups that best fit the “software eating the world” thesis are probably “sharing economy” companies like Bird, DoorDash, Instacart, Lime, Lyft, Uber, and WeWork. Each of these companies use software to offer services in the “real world”—taxi rides, scooter rentals, food delivery, lodging, office space, and so forth. They enjoyed a lot of hype in the mid-2010s, and most of them have struggled in the last few years.

Some of them have been total fiascos. WeWork failed to disrupt commercial real estate. Shares in the scooter startup Bird have lost 97 percent of their value since the company went public less than two years ago. Last year I drove for Lyft for a week and wrote about its difficulty in turning a profit. 

The two most successful “sharing economy” startups are probably Airbnb (founded in 2008) and Uber (founded in 2009). These companies are each worth tens of billions of dollars, and they seem likely to be enduring, profitable businesses.

Still, Airbnb has only a modest share of the overall lodging industry. And in recent years, the quality of Uber’s service has deteriorated, with higher fees and longer wait times. Smartphone-based ride hailing is a marginal improvement over conventional taxis, but hasn’t been a revolution.

In his 2011 essay, Andreessen specifically mentions health care and education as industries ripe for disruption by software. But as far as I can see that hasn’t happened. Hospitals increasingly use computers for record-keeping and billing and software has been used to make new drugs and medical devices. Many people learn foreign languages using Duolingo or watch educational videos on YouTube. But people largely go to the same schools and hospitals they did 10 or 20 years ago.

The reason I’m relitigating this 12-year-old argument is that I hear echoes of it in contemporary discussions of AI. In the early 2010s, Silicon Valley thought leaders looked at the early success of companies like Airbnb and Uber, extrapolated wildly, and concluded that software was going to transform the entire economy. Today, AI thought leaders are looking at the early success of ChatGPT and Stable Diffusion, extrapolating wildly, and concluding that AI software is going to transform the economy and put tons of people out of work.

To be clear, I do think AI is going to be a big deal. I wouldn’t have started an AI newsletter otherwise. But as with the Internet, I expect the impact to be concentrated in information-focused industries and occupations. And most of the American economy is not information-focused: It’s focused on delivering physical goods and services like homes, cars, restaurant meals, and haircuts. It will be hard for AI to have a big impact on these industries for the same reasons that it’s been hard for Internet startups to do so.

A common line of thinking about AI and the labor market goes something like this: During the 20th century, we automated a lot of physical tasks. As a result, a worker’s value has increasingly been based on brain power rather than muscle power. But now it looks like AI will enable computers to surpass humans at many cognitive tasks as well. That could lead to a future where the average human worker isn’t better than machines at anything, and as a result they can’t find a job.

One thing this argument misses is that there are still lots and lots of tasks that human beings can perform better than any machine.

Take plumbers, for example. They need to get in and out of their cars, climb stairs and ladders, and carry heavy objects. The job also requires fine motor skills.

Today’s robots can’t perform these tasks at anywhere close to a human level. So even if you had human-level AI that perfectly understood how to be a plumber, it’s not obvious we could build a robot body versatile enough to do the job.

Back in 2015, the Defense Advanced Research Projects Agency held a contest where teams competed to build humanoid robots that could perform simple tasks like driving a vehicle, opening a door, turning a valve, and cutting a hole in a wall using a drill. A lot of the robots fell over during the competition. The successful ones performed the tasks far slower than a human being.

And these were research prototypes, not commercial products. Each robot in the DARPA competition was supported by a team of researchers that performed extensive maintenance between trials. So it would take far more human labor to prepare one of these robots to do a plumbing job than to just have a human plumber do the job.

Of course, that was eight years ago. Have robots improved since then? Earlier this year, Boston Dynamics, a leading robotics company, released a video showing that its humanoid Atlas robot (also still a research prototype) now has claw-like hands and can pick up heavy objects. While this represents impressive progress (previous versions had no hands at all), the company still has a long way to go.

“The simple claw grippers mean Atlas crushes everything it picks up,” Ars Technica’s Ron Amadeo wrote in his writeup of the video. You wouldn’t want this robot to fixing your leaky faucet no matter how intelligent its software was.

Some companies, including Tesla and Figure, claim they’re developing humanoid robots that will be capable of safely performing a range of domestic tasks. As far as I know, none of these companies have shipped these products yet, and I’m skeptical they’ll be able to live up to their own hype. Even if they do, it’ll take many years to manufacture enough humanoid robots to have a major impact on the labor market. And in the short-to-medium term, we’d get a lot of new jobs in the robotics sector.

But let’s suppose robotics technology progresses rapidly and in a couple of decades there are millions of humanoid robots with the physical capacity to perform most jobs people can do. Human workers will still retain another big advantage in the marketplace: the fact that other people like to interact with them.

This is most obvious in the caring professions. Even if someone invented a robot that was proven to be as safe as a human babysitter, I wouldn’t want it taking care of my 2-year-old. I bet you wouldn’t either. By the same token, elder care workers, physical therapists, psychologists, nurses, and workers in similar jobs should be insulated from AI-related disruption because most people are going to prefer human caregivers over robots.

But this point applies far beyond the caring professions. Think about coffee shops, for example. Many Starbucks locations already have high-end coffee machines that grind the beans and produce coffee that’s indistinguishable from what a human barista would make. It wouldn’t be hard to fully automate Starbucks locations so that a robot arm hands you your coffee when it’s ready.

But Starbucks isn’t going to deploy a robot like that because Starbucks customers aren’t just buying cups of coffee—they’re buying the relative luxury experience of having another human being prepare coffee for them.

In 2012, the Wall Street Journal reported that Starbucks had ordered its baristas to “stop making multiple drinks at a time” because customers were complaining that harried baristas had “reduced the fine art of coffee making to a mechanized process with all the romance of an assembly line.” You know what’s even less romantic than your barista making multiple cups of coffee at the same time? Having a robot make your coffee.

Lots of other industries work this way:

    Despite the low cost of workout apps (and before that workout videos) lots of people pay a premium to attend fitness classes with human instructors.

    Recorded music is cheap and ubiquitous, but fans pay a premium to see their favorite artists perform live.

    Downscale restaurants take orders at the counter (or on a touchscreen) and make customers bus their own trays, while fancier restaurants have a small army of waiters ready to refill your water glass and sweep up crumbs.

This dynamic will become even more important in the AI era. For example, in the future there will probably be low-cost virtual schools where students interact with chatbot tutors rather than human teachers. Undoubtedly some students will find them useful.

But many others will hate it for the same reasons they hated Zoom school during the pandemic. Lots of people like listening to in-person lectures. It’s easier to stay on track with your studies if you know a human professor (or teaching assistant) will be disappointed in you for turning in an assignment late or getting a bad score on a test. And going to a brick-and-mortar school lets you make friends with your classmates and participate in extracurricular activities. That’s going to continue drawing students to conventional schools no matter how good AI-powered virtual schools get.

Similarly, you might be able to get a medical diagnosis from a chatbot, but many people will be willing to pay extra for advice from a human doctor or nurse—especially if they need a physical exam. People may feel more comfortable sharing intimate details with a human doctor, not only because they feel an emotional connection but also because they may have more confidence that the information they share won’t be misused.

There’s a risk we could wind up with a two-tier health care system where affluent patients can talk to human doctors on demand while lower-income patients are required to interact with a chatbot first. But it’s hard to imagine AI eliminating demand for doctors altogether.

AI very likely will lead to lower employment in some occupations. A few categories of workers may be completely replaced by AI software. But there will be many more occupations where AI automates some parts of the job, allowing human workers to focus on other tasks and become more productive overall.

Even when machines automate a substantial portion of an occupation, that won’t necessarily lead to big job losses. Take automatic teller machines. Banks installed hundreds of thousands of ATMs in the 1980s and 1990s. This made it cheaper to operate a bank branch, so banks opened more branches. And each branch still had some tellers to handle tasks too complex for the ATM. As a result, the total number of bank tellers in the U.S. changed little between 1980 and 2010.

I expect AI to have a similar impact on many occupations. For example, Github Co-pilot is AI software that is already helping programmers increase their productivity by 20 to 50 percent. So will companies take this as an opportunity to cut their engineering workforces? Some might, but I bet most will instead take the opportunity to produce more software.

And in cases where AI does reduce employment in an occupation, workers will move to other occupations where demand is strong. And as we’ve seen, there are plenty of jobs, from plumbers to baristas to doctors, that aren’t going to be replaced by AI any time soon.

A recent paper by computer scientist Ed Felten, business professor Manav Raj, and economist Robert Seamans looked at a database of occupations and tried to define how “exposed” each job was to large language models like ChatGPT. They did this by breaking each job down into 52 human abilities (things like “oral comprehension, oral expression, inductive reasoning, arm-hand steadiness”) and then estimating how well AI software could perform each of these tasks.

Telemarketers topped the list of most exposed occupations, followed by several categories of college professors. Judges, arbitrators, and clinical psychologists also made the list. This seems like a reasonable first step for estimating the impact of AI on the labor market, but I also think the list illustrates the limitations of the methodology.

For example, it’s very hard to imagine we are ever going to replace human judges with AI. No matter how technically competent AI might get at interpreting and applying the law, people are going to want to keep that function in human hands. People will be more likely to trust rulings made by a human being whose reasoning they can understand and who seems to share their values.

I’m also skeptical that we’ll see falling demand for clinical psychologists. While some people will undoubtedly enjoy talking to chatbots about their mental health problems, many others will prefer to talk face-to-face with a human being. And as I explained above, it’s not obvious that AI will put many college professors out of work.

Interestingly, the researchers found a strong positive correlation between language model exposure and income: higher-wage occupations are more likely to be impacted by the latest AI technology. So it’s very possible that AI technology will narrow the large wage premium that opened up between college graduates and less educated workers in the late 20th century.

Finally, it’s important to remember that the level of employment across the economy is ultimately driven by macroeconomic factors: If consumers spend more money, then businesses will respond by hiring more workers. The last three years have illustrated how powerful this can be: In the wake of the pandemic, Congress and the Fed worked a little too hard to boost the economy, producing a super-tight labor market and rising inflation.

If AI starts replacing workers in the coming years, that will put downward pressure on wages and prices while growing the economic pie. That will give the Fed more leeway to cut interest rates and give Congress more room to raise spending or cut taxes. As long as Congress and the Fed are doing their jobs, there’s no reason for the total number of jobs, economy-wide, to decrease.

The Internet’s biggest impact on the world may turn out to be cultural rather than economic. Music, television, and journalism have been transformed by platforms like Spotify, YouTube, Netflix, and Twitter. Social media has transformed politics, increasing polarization and powering the rise of populist movements around the world. And many white-collar workers have had to master new tools to remain at the forefront of their careers.

But in material terms, our lives aren’t much different than they were 30 years ago; our homes, grocery stores, neighborhoods, schools, and hospitals haven’t changed very much. Economic growth in the United States actually slowed down during the Internet era: inflation-adjusted GDP per capita rose more between 1962 and 1992 than it did between 1992 and 2022.

I expect a similar story with AI. Stable Diffusion has already revolutionized how we make digital images, and audio and video content won’t be far behind. The ability to generate customized (and possibly fake) text, images, and video on a large scale could have dramatic and unpredictable effects on our politics. Many white-collar workers will need to master new AI tools to remain at the forefront of their careers. Some might be forced to change careers altogether.

And if (like me) you’re a white-collar worker who sits in front of a computer all day, that will probably feel like a revolution, just as the Internet felt like a revolution to us 20 years ago. But most of life—and most jobs—are not online. The impact of AI on most people’s day-to-day lives is likely to be correspondingly modest.

I’ve got a favor to ask. To help me keep up with the rapid pace of change in AI, I’m looking to talk to readers who work with AI—whether that’s building it, using it, or studying it.

I’ve opened up a number of 30-minute slots on my calendar tomorrow and Wednesday. If you’d be willing to talk to me, please click here. All conversations will be strictly off the record. Thank you!