This article was taken from the May 2014 issue of Wired magazine. Be the first to read Wired's articles in print before they're posted online, and get your hands on loads of additional content by <span class="s1">subscribing online.
The relationship between employment and technology isn't a straightforward one.
Since the American psychology professor Paul E Meehl published his groundbreaking 1954 book, Clinical vs Statistical Prediction: A Theoretical Analysis and a Review of the Evidence, algorithms have consistently proven themselves capable of outperforming even expert opinion in a range of fields. In medicine, algorithms can prove more accurate than doctors when it comes to diagnosing illness. In law, artificial-intelligence "e-discovery" algorithms can perform many of the jobs that in previous years would have been carried out by (human) junior lawyers.
The same is true of manual jobs such as long-distance driving. A decade ago, MIT and Harvard economists Frank Levy and Richard Murnane published The New Division of Labor, in which they claimed that the job performed by drivers could never be carried out by algorithms because of the constant stream of visual, aural and tactile information drivers received from their environment. But a decade on, Google's algorithmically driven cars have completed more than 480,000km of road tests in a wide range of conditions, and causing no reported accidents. In other words, a person is safer in a car driven by an algorithm than they are in one driven by a human.
Even if your job is not -- yet -- in danger of being replaced by automation, that's not to say that your work life is free of the influence of algorithms. Recently, a number of schools have started using a program called CourseSmart, which uses e-book analytics to alert teachers if their students are studying the night before tests, rather than taking a long-haul approach to learning. In addition to test scores, the CourseSmart algorithm assigns each student an "engagement index" which can determine not just if a student is studying, but also if they're studying properly. In theory, a person could receive a "satisfactory" C grade in a particular class, only to fail on "engagement".
In Amazon's warehouses, meanwhile, product pickers (known as "fulfilment associates") are issued with handheld computers which transmit instructions to reveal where individual products can be picked up or dropped off. Because of the size of the warehouses in question, a routing algorithm is used to work out the shortest possible journey from point to point. That is not all the handheld computers do, though. They also collect a constant, real-time stream of data that monitors how fast employees walk and complete individual orders, thus quantifying their productivity. Like the bomb-laden bus in the movie Speed, workers must maintain a certain minimum speed, or else see their jobs go up in smoke. As with Henry Ford's assembly lines, machinery determines the pace of work. A warehouse manager at Amazon has been quoted as describing workers as "sort of like a robot, but in human form".
Although these precedents raise questions about our future employment and employability, not everyone thinks that automation has to lead to either techno-replacement or dehumanised humans.
A growing number of computer scientistsare seizing upon algorithms not as a way to replace human workers, but as a means by which to reimagine the workplace in a way that promotes inclusion and meritocracy, while smashing human bias and discrimination. One such company is Gild.
Founded in 2011, San Francisco-based Gild is a recruitment company which serves some of the tech industry's biggest and best-known players. Currently focused on automating the discovery of talented programmers, Gild has a mission statement: to automate the notoriously unreliable hiring process. To do this, the company uses algorithms to analyse individuals on tens of thousands (soon hundreds of thousands) of different metrics and datapoints -- mining them for insights in what Gild refers to as "broad predictive modelling".
The success stories the company trots out are impressive. A typical one tells of 26-year-old college dropout Jade Dominguez, who lived off an increasing line of credit-card debt in South Pasadena, California, while teaching himself computer programming.
After being "discovered" by Gild's algorithm, he now works as a programmer at the company which found him. His story isn't unique. "These are people whose CVs you wouldn't look twice at, but who our algorithm predicts would be perfect for the job," says Vivienne Ming, Gild's chief scientist. "For some of our customers, that is exactly what they're looking for. These are companies that are flooded with résumés. They don't need us to find people; they need us to find different people."
The first time I spoke to Ming, she was sitting in the back of a taxi on her way to San Francisco International Airport. A tall, striking woman with silver-blue eyes and strawberry-blonde hair, Ming is a theoretical neuroscientist with a Carnegie Mellon University background. She's effortlessly assured. Her geeky engineering side is evidenced by the fact that she wears a pre-release Google Glass headset. Her Twitter profile describes her as an "intrepid entrepreneur, undesirable superhero [and] very sleepy mother".
Ming is deeply invested in Gild's utopian vision of turning the workplace into the kind of meritocracy she believes it should be. "This is the way things ought to work, right?" she says, rhetorically. "The person making the hiring decisions really should have an accurate picture of who I am -- not just a snap judgement made because I look a certain way. But believe me, the way that people look is a huge influence on hiring."
If there is a reason why the idea of people being misjudged on first appraisal hits home particularly hard with Ming it may have something to do with her background. Born Evan Campbell Smith, Ming underwent gender-reassignment surgery in 2008, having "ghosted through life" up until that point. "I was not a classically good student," she explains. "I frequently failed my classes, I was constantly in trouble at school. I was not engaged, but I deeply cared about the learning experience. I loved learning."
To Ming, there are two main problems with classic hiring strategies. The first is that they are inherently biased. Although the majority of people appreciate the value of recruiting candidates with a different background from themselves, they are often not exposed to these individuals in social settings. Why, she asks, is a typical startup composed of similar-looking individuals of approximately the same age, with the same scruffy engineering look? Because they hired people they knew. The person who is good friends with a white, male, upper-middle-class, hard-working engineer is statistically more likely to be a hard-working engineer themselves. They are also likely to be white, male and upper-middle-class. As data-driven algorithmic culture has taken over, these casual assumptions have in many cases become codified.
To get a job at Facebook, one of the initial tests used in the weeding-out process is to find a person already working for Facebook who knows you. A similar process permeates LinkedIn, whose algorithms search for connections between an individual and the person they are trying to meet. Although the idea can be useful, it can also have the unfortunate effect of excluding a significant number of people from diverse regional, social and cultural backgrounds.
Ming also explains that legacy hiring strategies have proven inaccurate. In a place such as Silicon Valley, where the supposed objectivity of data-driven hiring is prized above all else, this is particularly unforgivable. Google, for example, employs what it calls the Lake Wobegon Strategy for hiring -- named after American humourist Garrison Keillor's claim that he grew up on the fictitious Lake Wobegon, where "all the women are strong, all the men are good looking, and all the children are above average."
According to Google's Lake Wobegon Strategy, to maintain a high level of skill in an organisation that is doubling in size each year, new employees should be above the mean skill level of current Googlers in order to be hired. To measure something as unquantifiable as "skill", Google traditionally placed a strong emphasis on academic results. A person's grade-point average and university were considered a strong predictor of workplace success, since they showed past evidence of rigour, commitment and the ability to meet deadlines. Someone who studied computer science at MIT might not be the best computer scientist in the world, but it is safe to assume that they are at least "good enough" to have got on to the course in the first place. Pick a random person who didn't go to MIT, on the other hand, and although there is still the chance that they will be brilliant, the likelihood that they will be terrible is higher. In a risk-averse industry where people are rarely given a bonus for betting on the long shot that pays off, but could very easily lose their job for hiring a person deemed unsuitable for a particular role, it is no wonder that many high-tech companies would choose to play it safe.
As data sets piled up, however, and algorithms began scouring that information for patterns, Google realised that the metrics it was using to predict job performance (including school grades, job experience and even interviews) offered very little in the way of accuracy when forecasting who was likely to excel in a particular position. In the 60s, US telecoms giant AT&T conducted IQ tests on low-level managers, and then followed them for 20 years to see how each employee progressed within the organisation. What was discovered was that IQ scores explained less than 15 per cent of the variance between managers in terms of career achievement. The rest was an unmeasurable combination of personality traits, emotional attributes, sociability and a number of other characteristics that can determine success.
Once Google realised the need to open up the parameters of what it looked for in a new employee, the cultural space was cleared for Gild. Instead of hiring from a group of tens of thousands each year, high-tech companies can now choose from a much wider talent pool. Gild assesses prospective employees by measuring all available data about them. Ming likens the difference between this approach and that of a human recruiter to a human chess player competing against Deep Blue. "The computer is plotting every possible move and choosing the optimal one," she says. "Chess experts are not. They have implicitly discarded the vast majority of possible moves and are only considering two, three, four possibilities. They just happen to be great ones." That's also true of people making hiring decisions, Ming suggests -- only that the metrics considered in this case don't happen to be so great. By looking at as many datapoints as possible about a person, anomalous factors such as whether a person being interviewed was having an off-day are bypassed. Gild also looks at where individuals spend time online, since this has been shown to be a strong predictor of workplace skills. "If you spend a lot of time blogging it suggests that you're not quite as good a programmer as someone who spends their time on Quora," Ming says. Even Twitter feeds are mined for their insights, using semantic and sentiment analysis. At the end, factors are combined to give a "Gild Score" out of 100. "The important takeaway is that what we end up with is truly independent dimensions for describing people," she says. "We're talking about algorithms whose purpose is to aggregate across your entire life to build a very accurate representation of who you are."
Gild isn't alone in looking to open up the number of metrics on which individuals are judged in the workplace. In 2012, three universities carried out a study as to whether or not Facebook profiles can be used to predict how successful a person is likely to be at their job. By analysing photos, wall posts, comments and profiles, researchers argued that questions such as "Is this person dependable?" and "How emotionally stable is this person?" can be answered with a high level of certainty. Favourable evaluations were given to those students who had travelled, had more friends and demonstrated a wide range of hobbies and interests. Partying photos didn't have to count as negative either, since people depicted as partiers were characterised as extroverted and friendly: both viewed as ideal qualities. Six months after making their initial predictions, the study's authors followed up with the employers of their 56 test subjects and found a strong correlation between job performance and the Facebook scores that had been awarded for traits such as conscientiousness, agreeability and intellectual curiosity. Their conclusion was that Facebook profiles are strong predictors, since candidates will have a harder time "faking... their personalities" on a social network than they would in a conventional job interview.
It is into a similar space that Silicon Valley startup Knack enters the picture. Founded by Israeli entrepreneur Guy Halfteck, it uses a combination of gaming technology, machine-learning algorithms and the latest findings from behavioural science to come up with universal measures for terms such as "quick-thinking", "perceptiveness", "empathy", "insightfulness", "spontaneity" and "creativity". Halfteck says that he hopes to trigger a "fundamental change in the human capital space" that will seek to unlock an individual's potential.
The basis for Knack's work is an insight that has been explored by psychologists for the last half-century: that the way we play games can be used to predict how we behave in the real world. "Even though to your eye, your behaviour in a game does not necessarily characterise your real-world behaviour, it is highly likely that the way you play a game and another person plays that same game would reveal differences about personality and the way that your brain works," Halfteck says. "Your working memory, your strategic thinking, your risk-taking -- these are all things which are manifested in how we game."
Knack's games currently include Wasabi Waiter and Balloon Brigade. Both are straightforward, pick-and-play affairs which nonetheless offer the player a number of different ways to compete.
In Wasabi Waiter, for example, players assume the role of a waiter and chef as they take customers' orders and then prepare the dish that matches his or her facial expression. This expression might be "happy", "sad", "angry", or "any mood" in the event that the player is unsure. When a customer finishes eating, the player brings their plate back to the sink and starts the process again with someone new.
This may appear simple, but beneath the surface this is anything but. In a game, literally everything is measurable: each action, message, item and rule is composed of raw data. For every millisecond of play, hundreds of data variables are gathered, processed and analysed based on the decisions players make, the speed at which they do things, and the degree to which their game-playing changes over time. What Halfteck perceives to be the accompanying behavioural qualities are then teased out using machine-learning tools and data-mining algorithms. "This is a very rich data stream we're collecting," Halfteck says. "It really allows us to get closer to the unique behavioural genome of a person."
There is, however, a danger in attempting to quantify the unquantifiable. When it comes to taking complex ideas and reducing these to measurable elements, the most famous critique came from evolutionary biologist Stephen Jay Gould. In his 1981 book The Mismeasure of Man Gould warned about the dangers of converting concepts such as "intelligence" into simplified measures such as IQ. Gould's concerns weren't just about the potential dangers of abstraction, but about the ways in which apparently objective truths can be used to back up human biases, rather than to expose genuine insights. In his own words, The Mismeasure of Man is an attack on "the abstraction of intelligence as a single entity, its location within the brain, its quantification as one number for each individual, and the use of these numbers to rank people in a single series of worthiness, invariably to find that oppressed and disadvantaged groups -- races, classes, or sexes -- are innately inferior and deserve their status." Measurement and reductionism do, of course, go hand in hand. Each time we begin to measure something, we lose whatever it is that the measurement tool is not designed to capture, or that the person measuring is not aware of the need to measure. Technological attempts to create simple quantifiable measures for ideas such as "creativity" and "love" meet with fierce opposition -- largely because the concepts are far from simple.
But Halfteck disagrees that the beauty of terms such as "empathy" and "insightfulness" lies in their abstract amorphousness. "We are talking about an ever-expanding universe of things that are being measured," he says. "It's not just about
'intelligence', but rather the sum total of the human condition.
It's far more nuanced than anything else." Surviving the employment market of the future won't just be a matter of picking those jobs that won't be able to be replaced by a well-designed bot, but also understanding the algorithms that will increasingly play a role in the hiring process.
The Formula: How Algorithms Solve All Our Problems... And Create More by Luke Dormehl is published on April 3 by WH Allen
This article was originally published by WIRED UK
