Categories
Antiwork

Mass Automation and The Future of Work Part 2

WHITE-COLLAR JOBS WILL DISAPPEAR, TOO Here’s an article written in 2017 about an earnings report for a jam company, J.M. Smucker: EPS ESTIMATES DOWN FOR J.M. SMUCKER IN PAST MONTH Over the past three months, the consensus estimate has sagged from $1.25. For the fiscal year, analysts are expecting earnings of $5.75 per share. A year after being $1.37 billion, analysts expect revenue to fall 1% year-over-year to $1.35 billion for the quarter. For the year, revenue is expected to come in at $5.93 billion. A year-over-year drop in revenue in the fourth quarter broke a three-quarter streak of revenue increases. The company has been profitable for the last eight quarters, and for the last four, profit has risen year-over-year by an average of 16%. The biggest boost for the company came in the third quarter, when profit jumped by 32%. ​ Notice anything off about the piece? The prose…


WHITE-COLLAR JOBS WILL DISAPPEAR, TOO

Here’s an article written in 2017 about an earnings report for a jam company,

J.M. Smucker: EPS ESTIMATES DOWN FOR J.M. SMUCKER IN PAST MONTH Over the past three months, the consensus estimate has sagged from $1.25.

For the fiscal year, analysts are expecting earnings of $5.75 per share.

A year after being $1.37 billion, analysts expect revenue to fall 1% year-over-year to $1.35 billion for the quarter. For the year, revenue is expected to come in at $5.93 billion. A year-over-year drop in revenue in the fourth quarter broke a three-quarter streak of revenue increases. The company has been profitable for the last eight quarters, and for the last four, profit has risen year-over-year by an average of 16%.

The biggest boost for the company came in the third quarter, when profit jumped by 32%.

Notice anything off about the piece? The prose isn’t going to win any awards. But it’s perfectly understandable. As it turns out, the article was written by AI. A company called Narrative Science produces thousands of earnings previews and stock updates for Forbes and recaps of sports stories for fantasy sports sites in real time.

The company’s bots won’t be winning any Pulitzers in investigative reporting, but in the coming years, the quality of AI-produced writing will go from acceptable to very good—and those journalists who write routine stories like this will find their jobs increasingly at risk. We tend to think of automation as displacing blue-collar workers with jobs that involve basic, repetitive skills. The truth is a little bit more complicated than that. The important categories are not white collar versus blue collar or even cognitive skills versus manual skills.

The real distinction is routine vs. nonroutine. Routine jobs of all stripes are those most under threat from AI and automation, and in time more categories of jobs will be affected. Doctors, lawyers, accountants, wealth advisors, traders, journalists, and even artists and psychologists who perform routine activities will be threatened by automation technologies. Some of the jobs requiring the most education are actually among the most likely to become obsolete. Some of these threatened workers, like investment advisors, may find themselves surprised to be on the chopping block after supporting the profit-growing potential of automated technologies.

A friend of mine is a radiologist at Columbia University. He told me a story about how the chair of his department was recently invited to General Electric to take part in a demonstration where humans would compete with computers to read patient films. GE invited doctors with decades of experience, the tops in their field, to see whether the doctors could more effectively diagnose tumors based on radiology films than a computer.

Guess who won? The computer won quite easily. It turns out a software program can “see” a shade of gray on a film that is invisible to the human eye. The computer can also draw on millions of films to compare it with, a much larger reference set than even the most experienced doctor.

We are entering an age of super-intelligent computers that can take any complex data set—every legal precedent, radiology film, asset price, financial transaction, actuarial table, Facebook like, customer review, résumé bullet, facial expression, and so on—synthesize it, and then perform tasks and make decisions in ways that are as good as or better than the smartest human in the vast majority of cases. To think that this will not dramatically change the way organizations perform work and the employment of people is to ignore the way companies operate. Companies are paid to perform certain tasks, not employ lots of people. Increasingly, employing lots of people will mean that you’re behind the times.

During my brief tenure as a corporate attorney when I started my career back in 1999, I practiced at Davis Polk and Wardwell, one of the top firms in the world. When we were assigned a deal, the first thing we would do was look for whatever deal precedent we had in the system that was most similar. We used to joke about how much of what we did was “finding and replacing” terms in a contract. There is a lot of repetitive functioning in what we consider high-end professional jobs—what I call intellectual manual labor.

A doctor, lawyer, accountant, dentist, or pharmacist will go through years of training and then do the same thing over and over again in slightly different variations. Much of the training is to socialize us into people who can sit still for long periods and behave and operate consistently and reliably. We wear uniforms—either white coats or business suits. We are highly rewarded by the market—paid a lot—and treated with respect and deference for accruing our expertise and practice. Basically, we are trained and prepped to become more like machines. But we’ll never be as good as the real thing.

The Federal Reserve categorizes about 62 million jobs as routine—or approximately 44 percent of total jobs. The Fed calls the disappearance of these middle-skill jobs “job polarization,” meaning we will be left with low-end service jobs and high-end cognitive jobs and very little in between. This trend goes hand-in-hand with the disappearance of the American middle class and the startlingly high income inequality in the United States.

The vanishing jobs are due in part to the incredible development of both computing power and artificial intelligence. You might have heard of Moore’s Law, which states that computing power grows exponentially, doubling every 18 months. It’s hard to understand what exponential growth means over time. Take the example of a 1971 Volkswagen Beetle’s efficiency.

If it had advanced according to Moore’s Law, the vehicle, in 2015, would be able to go 300,000 miles per hour and get two million miles per gallon of gas. That’s what’s happening with computers. People didn’t think Moore’s Law could hold for the past 50 years, but it has, and computers continue to get smarter. Intel, Microsoft, Google, and IBM are investing in quantum computers—computers that store information on subatomic particles—that would extend Moore’s Law for years to come.

We are just now hitting the rapid ascent of computers that are unfathomably fast and powerful. When the IBM computer Deep Blue defeated the world’s foremost chess master in 1996, people were impressed but not that impressed. Chess is a game where there is a very large but finite number of moves and possibilities, and if you have enough computing power you can project out all of the next possible steps.

Go is another story.

Go is a 3,000-year-old Chinese game with theoretically infinite moves. In order to beat the world’s best go players, an AI would need to use something resembling judgment and creativity in addition to pure computation. In 2015 Google’s DeepMind beat the world’s best go player and then did it again in 2017 against other world champions. Go champions looked at the DeepMind strategies and said that it used moves and tactics no one had ever seen before. New kinds of AI are emerging that can do much of what we now consider intelligent and creative. You might have heard the term “machine learning,” which is an application of AI in which you give machines access to data and let them learn for themselves what the best methods are.

Machine learning is particularly powerful because you don’t have to prescribe the exact actions and routes. You set guidelines, and then the AI starts synthesizing data and making choices and recommendations. Some of the early applications of machine learning include tagging images, spam filtering, finding keywords in documents, detecting outliers for credit card fraud, recommending stock trades, and other rules-based tasks. Machine learning is often used in conjunction with another term you’ve heard: big data.

Because of the digital revolution, we now have access to much more information than at any point in history, and the rate of new information is growing exponentially. One estimate is that more data has been created in the past two years than in the entire history of the human race. For example, we perform 40,000 search queries every second just on Google, which adds up to 1.2 trillion searches per year, each of which represents a new piece of information.

By 2020 about 1.7 megabytes of information will be created every second for every human being on the planet. Much of this information is mundane—a catalogue of people clicking on friends’ photos on Instagram and the like. But the point is that in this flood of new data, there will be very useful pieces of actionable information. The author Yuval Harari postulates a world where, based on analyzing your online data, an AI could tell you which person you should choose to marry.

There is now big money pouring into trying to process all of this information—one estimate is that a typical Fortune 1000 company could make another $65 million a year by increasing its use of data by 10 percent, and that only 0.5 percent of available data is presently analyzed and used. Another estimate is that the health care system could save $300 billion per year—or $1,000 per citizen per year—with improved use of data. Industries that utilize large amounts of data—like financial services—are already being transformed to take advantage of new capabilities.

The finance industry is in many ways a natural home for automation; the tasks are highly repetitive and logical, the institutions are rich and efficiency-minded, and the culture is hypercompetitive. Founded in 2008, Betterment is an automated investment service that by 2017 had more than $9 billion under management. With lower fees and automated investment decisions, Betterment and its competitor Wealthfront largely replace the traditional financial advisor.

Said the Financial Times, “Younger clients don’t want, and can’t afford, an annual meeting with an advisor talking about the relative pros and cons of emerging markets, bonds or structured products. They want simple guidance and 24-hour access… they don’t want advice delivered in an office, they want an app.”

By 2020, global assets under management of robo-advisors are projected to skyrocket to $8.1 trillion, and 72 percent of investors under 40 said they would be comfortable working with a virtual advisor. Fifty-five hundred floor traders once roamed the trading floor of the New York Stock Exchange. Now there are fewer than 400, as most trading jobs have been taken over by servers running trading algorithms. Those scenes you see on CNBC are not of the New York Stock Exchange but of the Chicago Mercantile Exchange, where they still have enough humans to make a good backdrop.

Goldman Sachs went from 600 NYSE traders in 2000 to just two in 2017 supported by 200 computer engineers. In 2016 the president of the financial services firm State Street predicted that 20 percent of his 32,000 employees would be automated out of jobs in the next four years. A new AI for investors platform called Kensho has been adopted by the major investment banks that does the work that used to be done by investment banking analysts to write detailed reports based on global events and company data—Kensho is valued at $500 million after less than four years in business.

With Kensho, a report that would have taken 40 hours for a highly educated human being paid $250,000 per year can now be done in minutes. Accordingly, Bloomberg reported that Wall Street reached “peak human” in 2016 and will now shed jobs progressively, which has been borne out by layoffs this year at most of the major banks. The insurance industry, which employs 2.5 million Americans, revolves around processing information, which also makes it particularly ripe for automation.

McKinsey predicts a massive diminution in insurance staffing across the board, particularly in their operations and sales agent departments, projecting a 25 percent total decrease in employment by 2025.

That will mean hundreds of thousands fewer white-collar workers in cities around the country. Accountants and bookkeepers are vulnerable, too. One accountant described switching from billing per hour to monthly retainers because cloud accounting software was automatically doing the bookkeeping and he suddenly wasn’t spending any time on it.

There are 1.7 million bookkeeping, accounting and auditing clerks in the United States and an additional 1.2 million accountants and auditors. Bookkeepers and clerks are already starting to disappear. Accountants talk bravely about shifting their time to advise clients on financial strategy. I’ve employed half a dozen accountants in my life, and most of the time you just want to get your taxes done and filed. Even occupations that revolve more around words than numbers are at risk.

A Deloitte report in 2016 projected that 39 percent of jobs in the legal sector will be automated and that the industry should expect “profound reforms” in the next 10 years. In particular, paralegals and legal secretaries are expected to be replaced, and overall employment in the sector is expected to shrink as many law firms will contract or consolidate.

When I went to law school in the late 1990s, people regarded it as a safe career move. Today, law schools churn out many more graduates than the market requires, and the market for their services is shrinking. A friend of mine runs an AI company that is automating basic litigation tasks—routine responses, filings, and document review—for large companies, who won’t need to hire as many freshly minted lawyers as a result.

I met with Cliff Dutton, the chief innovation officer of a global legal processing company, who described how human attorneys have about a 60 percent precision rate reviewing boxes of legal documents. I remember performing document review as a young associate—my eyes glazed over after a couple of hours even when I was trying hard to focus.

The comparable rate for AI-enabled software is already closer to 85 percent accuracy, and it’s a lot faster than a team of lawyers could ever be. Even more than lawyers, doctors have built up their expertise, wisdom, and decision-making ability through many painstaking years of both training and practice. Yet I asked a high-end doctor friend who attended MIT and Harvard how much of medical practice he thought could be performed via automation. He said, “At least 80 percent of it is ‘cookbook.’ You just do what you know you’re supposed to do.

There’s not much imagination or creativity to most of medicine.” I sat with a technologist to project which aspects of medicine were most ripe for automation. His responses were radiology (as discussed above), pathology (very similar), family medicine (a nurse practitioner or even layperson could handle most issues with the assistance of AI), dermatology (similar), and a couple other specialties.

He also talked about how surgeons he knows enjoy the robot-assisted operating theatre, because it greatly enhanced their vision and ability to see things and the robot tools automatically accounted for unwanted movements and motions like a trembling hand. Also, students who were meant to train could see everything without being in the room, and the surgeon could review his procedure after the fact. I asked if doctors could potentially perform surgeries from remote locations. He responded, “Eventually. Right now doctors want to be

nearby, and the latency of long-distance data transmission still could cause delays or lags.” Still, he agreed that robo-assisted surgery will soon open up the ability for a top surgeon to perform surgeries around the world. It also means that you can record surgeries and all of the micro decisions that surgeons make. With that data, eventually AI could analyze thousands of surgeries and know what to do in every situation.

The first robot dental implantation—with no human intervention—just took place in China in September 2017. The robot went in and installed two new implants that had been printed by a 3D printer. Robot super surgeons might be one generation away. Most people assume that humans will always have the advantage over AI when it comes to work that requires creativity, like painting or music, and jobs that require nuanced, sensitive human interaction, like therapy.

In fact, Google’s neural network, a computer system modeled to “think” like a human, has produced art that could easily be confused for a human being’s, like the work on the next page. You can also check out a symphony online that was composed by a software program, Iamus, which many listeners found indiscernible from a human composition when it was performed. Google “Adsum” by Iamus and take a listen.

You might have figured that therapy would be the last province of automation. If so, you were wrong.

USC researchers funded by the Department of Defense in 2016 created an AI therapist named Ellie to treat veterans for post-traumatic stress disorder (PTSD).Ellie appears as a video avatar and provides soothing questions and responses. Ellie measures voice tone and facial expressions to try to identify whether a soldier needs to seek additional treatment with a human counselor. Early research is promising and indicates that soldiers often feel more comfortable confiding in a clearly artificial therapist than an actual human being. Ellie is meant to be a complement to human therapists—but one can easily imagine her checking in with patients in between appointments and taking on more over time.

When I was 13, I had to have four teeth pulled in preparation for wearing braces. I was actually kind of excited about it because I saw my dad’s teeth and was like, “whatever it takes, let’s not have those.”I remember going to the dentist and wondering what kind of magic he would employ to pull the teeth. Not much—Dr. Goodman just put some pliers on the first tooth and yanked and jerked until it came out.The second one was stubborn and he had to shift positions a few times—I remember him putting his foot on my chest and yanking away. I walked away thinking, “Wow, dentists have to be kind of strong to do what they do.” Also, my jaw hurt. I tell this story because often the boundary between what we consider intellectual and manual work will be unclear.

Surgeons are among the highest-trained, most highly compensated doctors because cutting people open is a big deal. Yet their highest-value work is, for the most part, manual and mechanical. My surgeon friends often swear off activities like basketball because they are worried they’ll hurt their fingers or hands. Some jobs might not go away the instant new technology arrives that could replace them.

Much of how automation unfolds in medicine is dependent upon regulations and licensing. It is presently illegal to do many things without a doctor or pharmacists’ license. This is very likely to be a field where technical innovation far outstrips implementation because doctors will fight the steps, and they have a very powerful lobby. They will argue that no one is as good for a patient as a highly trained human doctor, even in the face of dramatic evidence to the contrary as AI improves.

Some patients also might prefer seeing a human doctor, though I suspect this preference will fade over time. There are many obstacles to AI truly becoming broadly intelligent—one neuroscientist described most systems today as being better than a human could ever be at one specific task and dumber than a two-year-old at anything else. Still, our conception of what is beyond the capacity of a computer is about to change.

There is a lot of white-collar and creative work that can be automated. In startups, we have a saying of what to do when you’re not sure what the answer is: “Throw money at the problem.” Soon, the answer to everything will be “Throw AI at the problem.” If you think your job is safe from computers, you’ll probably be wrong eventually.

The purpose and nature of work is going to change a lot in the next 10 years. The question is what will drive this change aside from the fact that fewer of us will have jobs to go to.

Andrew Yang, The War on Normal People.

Leave a Reply

Your email address will not be published.