Machines put half of U.S. jobs at risk: study
SAN FRANCISCO — Who needs an army of lawyers when you have a computer?
When Minneapolis lawyer William Greene faced the task of combing through 1.3 million electronic documents in a recent case, he turned to a so-called smart computer program. Three associates selected relevant documents from a smaller sample, “teaching” their reasoning to the computer. The software’s algorithms then sorted the remaining material by importance.
“We were able to get the information we needed after reviewing only 2.3 per cent of the documents,” said Greene, a Minneapolis-based partner at law firm Stinson Leonard Street.
Artificial intelligence has arrived in the American workplace, spawning tools that replicate human judgments that were too complicated and subtle to distil into instructions for a computer. Algorithms that “learn” from past examples relieve engineers of the need to write out every command.
The advances, coupled with mobile robots wired with this intelligence, make it likely that occupations employing almost half of today’s U.S. workers — from loan officers to cab drivers and real estate agents — become possible to automate in the next decade or two, says a study done at the University of Oxford in the U.K.
“These transitions have happened before,” said Carl Benedikt Frey, co-author of the study and a research fellow at the Oxford Martin Program on the Impacts of Future Technology.
It’s a transition on the heels of an informationtechnology revolution that’s already left a profound imprint on employment across the globe. For both physical and mental labour, computers and robots replaced tasks that could be specified in step-by-step instructions — jobs that involved routine responsibilities that were fully understood.
That eliminated work for typists, travel agents and a whole array of middle-class earners over a single generation. Yet even increasingly powerful computers faced a mammoth obstacle: They could execute only what they’re explicitly told. It was a nightmare for engineers trying to anticipate every command necessary to get software to operate vehicles or accurately recognize speech. That kept many jobs in the exclusive province of human labour — until recently.
Oxford’s Frey is convinced of the broader reach of technology now because of advances in machine learning, a branch of artificial intelligence that has software “learn” how to make decisions by detecting patterns in those humans have made.
The approach has powered leapfrog improvements in making self-driving cars and voice search a reality in the past few years. To estimate the impact that will have on 702 U.S. occupations, Frey and colleague Michael Osborne applied some of their own machine learning.
They first looked at detailed descriptions for 70 of those jobs and classified them as either possible or impossible to computerize. Frey and Osborne then fed that data to an algorithm that analyzed what kind of jobs lend themselves to automation and predicted probabilities for the remaining 632 professions.
The higher that percentage, the sooner computers and robots will be capable of stepping in for human workers. Occupations that employed about 47 per cent of Americans in 2010 scored high enough to rank in the risky category, meaning they could be possible to automate, “perhaps over the next decade or two,” their analysis, released in September, showed.
“My initial reaction was, wow, can this really be accurate?” said Frey, who’s a Ph.D. economist. “Some of these occupations that used to be safe havens for human labour are disappearing one by one.”
Loan officers are among the most susceptible professions, at a 98-per-cent probability, according to Frey’s estimates. Inroads are already being made by Daric Inc., an online peer-to-peer lender partially funded by former Wells Fargo & Co. chairman Richard Kovacevich. Begun in November, it doesn’t employ a single loan officer. It probably never will.
The startup’s weapon: an algorithm that not only learned what kind of person made for a safe borrower in the past, but is also constantly updating its understanding of who is creditworthy as more customers repay or default on their debt.
It’s this computerized “experience,” not a loan officer or a committee, that calls the shots, dictating which small businesses and individuals get financing and at what interest rate. It doesn’t need teams of analysts devising hypotheses and running calculations because the software does that on massive streams of data on its own.
The result: an interest rate that’s typically 8.8 percentage points lower than from a credit card, according to Daric. “The algorithm is the loan officer,” said Greg Ryan, the 29-year-old chief executive of the Redwood City, Calif., company that consists of him and five programmers.
Similar technology is transforming what is often the most expensive part of litigation, during which attorneys pore over emails, spreadsheets, social media posts and other records to build their arguments.
Each lawsuit was too nuanced for a standard set of sorting rules, and the string of keywords lawyers suggested before every case still missed too many smoking guns. The reading got so costly that many law firms farmed out the initial sorting to lower-paid contractors.
The key to automate some of this was the adage to show not tell — to have trained lawyers illustrate to the software the kind of documents that make for gold. Programs developed by companies such as San Francisco-based Recommind Inc. then run massive statistics to predict which files expensive lawyers shouldn’t waste their time reading. It took Greene’s team of lawyers 600 hours to get through the 1.3 million documents with the help of Recommind’s software. That task, assuming a speed of 100 documents per hour, could take 13,000 hours if humans had to read all of them.
“It doesn’t mean you need zero people, but it’s fewer people than you used to need,” said Daniel Martin Katz, a professor at Michigan State University’s College of Law in East Lansing who teaches legal analytics.