Will robots steal our jobs?

Nicola appears to be the perfect employee. She’s quiet, conscientious and works well with her team at AstraZeneca’s Cambridge research centre. Despite being new to the lab, she is happy to toil day and night to identify compounds that could produce life-saving drugs.

Socially, she’s rather gauche – she never comes to the pub after work or takes her turn making the coffee – but there’s only so much you can expect from a robot.

“We call it Nicola! It has no emotional expression, but nevertheless the team has anthropomorphised it,” said Roeland van der Heiden, Astra Zeneca’s digital director, speaking at a recent conference at the Leverhulme Centre for the Future of Intelligence (CFI) in Cambridge.

“Work that used to take months now takes one or two weeks. We can use it remotely too,” he told me.

NiCoLA-B is one of a growing number of intelligent autonomous systems that threaten to steal our jobs or at least parts of them. Self-driving vehicles, health rehabilitation robots, fruit-picking robots, bricklaying robots and countless AI-based information processing systems are all part of the automation invasion.

‘Harnessing automation for a future that works’, a report by the think-tank McKinsey Global Institute, says that around half of all today’s work activities could be at risk by 2055, give or take 20 years. That’s one billion or so humans pushed over the cliff-edge of unemployment by NiCoLA-B and her kind. Price Waterhouse Coopers (PwC) has also been looking into this and predicts that 30 per cent of UK jobs, 38 per cent in the US and 35 per cent in Germany will be under threat by 2030.

Sales of intelligent hardware robots over the last year include five million domestic robots (mainly vacuum cleaners), 1.9 million drones, 30,000 industrial robots, 20,000 logistics robots, 11,200 military robots, 6,500 field robots, 6,000 milking robots and 1,500 medical robots, according to research by Sabine Hauart, lecturer in robotics at Bristol Robotics Lab.

Powered by learning algorithms, these autonomous systems promise to speed up a host of routine tasks that used to rely on human knowledge, reasoning, perception and manual dexterity. They can learn from experience and don’t need holidays or a living wage. They are also unlikely to organise themselves into unions.

Historically, technology has “enriched labour, not immiserated it,” Andy Haldane, chief economist of the Bank of England, told the UK Trades Union Congress in 2015.

As Martin Goodson, founder of the UK start-up Evolution AI, points out: “Between 1910 and 2000, the proportion of people employed in white-collar work went up by a factor of five: a time that saw the invention of the desktop and handheld mechanical calculator, the electronic calculator, and then the desktop computer.”

Each automation phase destroys jobs and livelihoods and then results in a “growing tree of rising skills, wages and productivity,” says Haldane. Yet this growth phase is also associated with a “hollowing out of jobs”.

Hollowing out means that mid-skill, well-paid jobs disappear and the workforce is polarised into high-skill, high-income and low-skill, low-income employment e.g. elite robot designers at one end and warehouse pickers (assisting the robots) at the other.

“Technology increases the number of jobs that require high technical skills, high organisational skills, high interpersonal skills or all three,” explained David Alan Grier, professor of international science and technology policy and international affairs, George Washington University, who spoke at the CFI conference. “These jobs tend to be closer to capital, i.e. jobs that influence the decisions about how an organisation should allocate its resources.”

Amazon is one of the largest users of robots. Its warehouses are populated by scores of squat orange logistics bots that whizz around carrying the heavy stuff, while humans package and stow goods.

Amazon has tripled the number of logistics robots in use from 15,000 to 45,000 over the last three years, while increasing jobs for people by 50 per cent.

A robot-equipped fulfilment centre can hold 50 per cent more inventory, and can complete a customer’s order within minutes rather than hours. Annually, a logistics robot costs only a little more than a human warehouse worker (judging from reported prices of the LocusBot from Locus Robotic and the MiR100 bot from Mobile Industrial Robots). These robots quickly pay for themselves.

At Amazon’s million-square-foot centre in Dupont, Washington, there are around 1,000 full-time employees to around 800 or so robots. It is only a matter of time before the robot-to-human ratio rises as robot dexterity improves.  In the UK, online grocer Ocado has demonstrated a robotic hand that can pick up fruit and vegetables in its warehouses (part of a five-year EU-funded collaboration called SOMA between five European universities and Disney).

The 35,000 robotic milking systems now on dairy farms around the world are another success story. Made by companies including DeLaval, Lely, Fullwood and GEA, these machines cost $100,000 or so each and can milk around 60 cows. The cows go voluntarily to be milked with the bribe of being able to eat treats called cow nuts. The machines use laser beams to guide the placement of individual pumps on the udders.

For farmers, these machines improve the quality of life, getting the tedious job of milking done in an undemanding time frame, Doug Stensland, a farmer in Iowa told magazine ‘Business Insider’. “Once we got these robots in, we were able to lighten the load on the family, plus actually we were able to eliminate a couple [of] jobs,” he said.

Robots are moving into other areas on the farm too. US start-up Abundant Robotics has developed one that picks apples off trees. Thorvald, a robot built by Pal Johan From, a professor at the University of Lincoln, can carry trays of strawberry plants to humans during the day and at night shine UV light on plants to kill mildew. Professor From thinks such machines could plug the agricultural labour shortage expected after Brexit.

Software-based AI systems for information processing are less visible, but they will have a more profound impact on the way we work, from clerical tasks to scientific research.

For example, take two deep-learning, natural language-processing applications built by Evolution AI. One system, a collaborative project with University College London and the NHS, has been trained (by pre-reading large numbers of example records) to pick out blood test data from 200,000 anonymised psychiatric records. It will enable one of the largest studies into the role of inflammation (shown by blood test markers such as CRP) in mental illness such as bipolar disorder, schizophrenia and depression.

Psychiatric records run to pages of description of patients’ symptoms over episodes of illness, with blood test results buried in the text. As they are long and unstructured, rarely more than a few hundred records – i.e. numbers that a small team of humans can read – have ever been studied in this way. A link with inflammation could mean anti-inflammatory drugs could be used in psychiatric treatment.

The second deep-learning system, developed for the commercial data provider Dun & Bradstreet, autonomously researches the internet or specific databases to learn the ‘language’ of different industries such as accountancy or civil engineering. By matching these ‘language fingerprints’ against information Dun & Bradstreet already holds on each of the five million companies in its UK database, it checks these firms are correctly listed into around 1,000 industry categories. Previously, a large team of phone researchers took a year to do the work. The system has saved them about 10,000 hours.

A £622,000 grant that Google has awarded the Press Association (PA) will produce a similar kind of deep-learning system for generating thousands of news stories a month. Working with UK-based news start-up Urbs Media and five journalists, PA will develop software called Radar (Reporters and Data and Robots) to automate AI-based news-generation for the UK’s struggling regional newspapers. Radar will analyse information from public databases (government agencies or local law enforcement) and write stories using Natural Language Generation. Journalists will identify datasets, edit copy and check facts.

With media outlets expected to deliver news 24 hours a day, all news journalists have become hard-pressed to keep up with the flow of press-released stories and Twitter feed news, let alone to dig any deeper. Radar could give journalists more time to do a proper job in a climate of stretched resources.

Frazzled British GPs with only minutes to diagnose patients’ strange symptoms might also welcome AI assistance, but there is a threat to their expertise. Recently IBM’s Watson supercomputer (powered with its DeepQA machine-learning, natural-language-processing software) correctly diagnosed – in 10 minutes – a patient suffering from leukaemia, who had been baffling doctors in Japan. Watson did this by studying the patient’s medical information and then cross-referencing her condition against 20 million oncological records, uploaded by doctors from the University of Tokyo’s Institute of Medical Science.

A study led by Andrew Beck, director of bioinformatics at the Cancer Research Institute at Beth Israel and an associate professor at Harvard Medical School, points to human-AI teamwork as a way forward in medical diagnosis. He compared how skilled deep-learning software and pathologists were at spotting metastatic breast cancer from images of lymph nodes. Pathologists were better than software (3.5 per cent error rate compared to 7.5 per cent). Yet working together makes combined error less than 1 per cent because humans and AI software make different kinds of mistakes.

Machine/human partnerships may show the way ahead for scientific research. However, what happens to people like America’s 2.5 million truck drivers, replaced by self-driving trucks with an estimated labour-cost saving of $70bn (2013 Morgan Stanley Research)? Who gets the benefits of those productivity gains?

“What technology companies need to remember is that without publicly-funded research that underpins their tech, they wouldn’t exist. They couldn’t thrive without the societies they exist within. Companies who make money out of automation need to respect this ‘social contract’,” said Alan Winfield, professor of robot ethics at the University of West of England, Bristol, talking to me at the CFI conference.

Certainly, potential destruction of employment needs a transition plan to retrain people, and some creative thinking about a future economy. Stuart Russell, from Berkeley Centre for Human-Compatible AI, proposes we look to movies and books for inspiration. King Midas, he told the CFI audience, is a ‘warning’ story. “He asked to turn everything to gold, but forgot that that would include his food, his family and friends. The film ‘Wall-E’ is useful as it shows a human race enfeebled by computers,” said Russell. “Economists don’t invent new economies. Let’s put economists and writers together and brain storm this.”

One idea supported by Microsoft’s Bill Gates and SpaceX’s Elon Musk (architect of Tesla) is to pay everyone a universal basic income, perhaps from a robot tax. “If a human worker does $50,000 of work in a factory, that income is taxed,” Gates said in a recent interview with Quartz, a business news site. “If a robot comes in to do the same thing, you’d think we’d tax the robot at a similar level.”

Over the next two decades, humans will have the edge in occupations involving complex perception and manipulation tasks, creative intelligence tasks, and social intelligence tasks. That’s the conclusion of Dr Michael Osborne and Dr Carl Benedikt Frey of Oxford University in their 2013 paper ‘The Future of Employment: How Susceptible Are Jobs to Computerisation?’. Actors and engineers are among those with jobs least at risk, along with those who care for others.

While Pepper, a humanoid robot created by SoftBank and Alderbaran Robotics (see box ‘In Love with Pepper’), can recognise some human emotions, it is at a trivial level. Getting a robot to handle complex caring tasks such as looking after an elderly person with dementia will be very hard, agrees Hauart. “Even fetching a glass of water requires it to understand human commands, navigate a home, open cupboards, find a glass, manipulate the glass, put it under a tap, fill it just enough, and bring it to the person in a non-threatening way.”

Neil Lawrence, senior principal scientist at Amazon Research in Cambridge, says the big difference between robot and human intelligence is the ‘embodiment factor’. He describes this as the ratio between the ability to compute and to communicate. Humans are like a 1,000-teraflop computer with a communication channel of 100bits per second i.e. a very high embodiment factor. A computer might combine 10Gflops with one gigabit per second communication, which is a very low embodiment factor.

 “Because of that restricted communication channel, humans do an immense amount of modelling of each other, based on what we know, and how they talk and what they look like, and host of other signals,” he explained at the CFI conference.

“Conversations between humans require that you model the other person in your head and model yourself and model what the other person thinks of you. And so on. It’s a ‘Russian doll’ effect. Computers will really struggle with that.”

This influences human conversation, emotional trust and creativity. To you and me, the famous line “For sale: baby shoes, never worn” is loaded with meaning and involves a huge shared understanding of the human condition, says Lawrence. Perhaps machines, however clever, will always struggle to compete on that level.