Can machines ‘learn’ or ‘think’?

The marriage of computing power and data is finally bearing fruit in the field of cognitive computing, sometimes called machine learning or, more controversially, artificial intelligence.

In its most everyday form, we see it in tools such as Google Translate or Microsoft’s Bing Translate, which can translate phrases and documents effortlessly across multiple languages. More futuristically, the promise of self-driving vehicles, which can complete entire road journeys without driver intervention, is already being realised.

Yet the biggest revolution in work is happening at some of the most basic levels, such as reading and dissecting legal documents to extract meaning and useful information. The tedious slog of work can be transformed by computers which are able to read and parse legal phrases, and summarise them or enter relevant details into a database or spreadsheet.

Digital tech transforming the way we work

Are these thinking machines? The idea has fascinated philosophers and technocrats for ages. But with every advance that machines make into space normally thought of as “thinking”, the goal posts retreat. Until IBM’s Deep Blue defeated then world champion Garry Kasparov in 1997, chess had been thought of as a redoubt for human thinking.

Built to evolve

More recently, the British company DeepMind created a computer program which can learn to play 1980s arcade games, such as Space Invaders and Breakout, by trial and error, based on what it sees on the screen, but without being told any rules or given any objective except to maximise its score. It’s a classic conundrum: is the DeepMind system “thinking” or “learning”? Certainly, it improves its score, and discovers neat ways to play games better. Google acquired DeepMind for £400 million in 2014.

In ten years, there could be $9 trillion of cuts in employment costs as AI systems take over knowledge work

Yet the impressive feats of translation tools don’t indicate that the machines behind them can actually “think”, nor even understand what it is that they are translating. Instead, they rely on a huge resource of data, principally documents containing the same content, which have been translated simultaneously into multiple languages. Publications from the United Nations and the European Union are highly favoured, for example, which may explain why machine translations can sound so remarkably stilted and formal.

But to a company using such a translation service, it doesn’t matter whether the computer can “think”, what matters is whether it gets the job done as well or better than a human. And a growing number of studies suggest that more and more jobs are susceptible. A recent study by the Bank of America forecast that the market for robots and artificial intelligence (AI) solutions will be worth $153 billion by 2020, of which AI solutions will be worth $70 billion. In ten years, there could be $9 trillion of cuts in employment costs as AI systems take over knowledge work, as self-driving vehicles and drones make $1.9 trillion of efficiency savings compared with having the work done by people, and robots and AI could boost productivity by 30 per cent, while cutting manufacturing costs by between 18 and 33 per cent.

Are jobs at risk?

The broad wave of cognitive computing is thus ready to break over the world of employment. But it’s not a single, simple implementation. “The area splits into two fields,” explains Andrew Martin, who is studying for a PhD in cognitive computing at the Tungsten Centre for Intelligent Data Analytics at the University of London. “There are people trying to make more and more complex systems with more and more data, hoping against hope that the problem will solve itself through big complex systems. And the other group is sitting back and going to the philosophical drawing board trying to work out what intelligence actually is, and how it emerges.”

So which group is the Tungsten Centre in? “Sort of both. We’re making big systems, but aware of the limits of what computers can and can’t do,” says Mr Martin. “We have a view of the things that won’t be solvable.”

Factors transforming work

Some problems look as though they’re beyond solution by one approach, but that doesn’t mean it can’t be done. In self-driving cars, Mr Martin says, “you have a machine that has to act in very complex situations, but it will never have the full situational awareness that a human driver does”.

Yet this sounds like some of the arguments that used to be used about chess: a computer could never win at chess, some used to argue, because it wouldn’t be able to understand the nuances of certain moves or understand ideas such as control of the centre of the board. Those arguments went by the board when IBM’s Deep Blue defeated Kasparov. Being able to do lots of calculations very quickly turned out to be a sufficient substitute for a human’s full situational awareness of the chess board.

Indeed, Google’s cars have driven millions of miles in the United States and the only accidents have been the fault of other, human drivers. In fact, a police officer recently flagged down a Google car because its driving seemed over-cautious.

Mr Martin says that with cognitive computing, “some things are instantly solvable because they’re constrained – the problems have clearly defined limits – and some people might think that solving the quickest route to somewhere isn’t cognitive computing”. But that used to be the ambit of taxi drivers with huge experience; now it’s available to anyone with a smartphone.

Man vs machine

So which are the fields that will be most affected by advances in cognitive computing? Analysis of legal documents is a key one. London-based law firm Berwin Leighton Paisner recently made substantial time-savings by using such a system to analyse the content of hundreds of Land Registry documents automatically, rather than getting the same work done by interns and paralegals.

“The real value that you add as a lawyer is about anomalies,” says Wendy Miller, a partner at the firm. “If clients have a huge number of contracts and want to understand them, it’s useful to have these data extraction tools. It’s applicable to a surprising number of tasks and we’re working to put it to work in other areas of law.”

At the Tungsten Centre, Mr Martin says the areas of work which will be most affected are those which “don’t need much human inspiration”. The centre is already studying the world of finance.

He points to vehicle manufacture as one which could easily be done by such systems and more prosaically to supermarket self-service checkouts. “The road haulage industry is at the biggest threat of being seriously disrupted by AI,” he says, “because motorways and motorway driving are relatively constrained environments.”

The impact of cognitive computing

There have already been tests of self-driving trucks in the US, Germany, Holland and Japan by Daimler, Scania, Ford and others. The potential for employment disruption is huge, since there are 3.5 million professional truck drivers in the US alone, whose income generates support for millions more people, whether in their families or the truck stops they visit as part of their work.

The way to think of cognitive computing is that it gives us very fast and obedient, but extremely stupid, slaves

What then will they move on to? How will the world of work be affected? At its core, this is the same question as that faced by horse and stable owners at the end of the 19th century as motor cars arrived. The assumption is that grooms and bridlemakers all found new work. But what’s never clear is whether they found better-paid work or subsistence. That tends to be the concern around the march of the new world of AI, which can also be deployed far faster than the car factories of the early-20th century could ramp up production.

Global tech predictions

“The way to think of cognitive computing is that it gives us very fast and obedient, but extremely stupid, slaves,” says Mr Martin. “The parts of industries that will remain are those which require knowledge.”

But what parts are those? How do we define “knowledge” so that we can be sure it won’t be accessible to a machine-learning system in five or ten years? Mr Martin says it’s easier to think of the tasks that will be susceptible, “things that you can think of as mostly rule-following and rote behaviour, repetitive, with no creativity, or where there’s only a small amount of independent thought and a lot of people doing it”.

The contrast is with fields which require deep knowledge and experience, such as the law and medicine. Even though IBM’s Watson is being used to analyse scans and data from cancer patients in a number of hospitals in the US, the expectation is you will still need doctors and lawyers to deliver the final decisions on what to do and where to focus.

CASE STUDY: BERWIN LEIGHTON PAISNER

Case Study

London-based law firm Berwin Leighton Paisner had a very specific challenge: analyse more than 700 Land Registry documents for a client, to extract details about land ownership such as the name and address of the overall owner, and related interests such as outstanding mortgages and other debts tied to them, plus any third-party interests in the title. And the answers had to be 100 per cent accurate.

In the past, the only way to do that would be to assemble a team of interns and paralegals, give them the documents and leave them to slog through until they emerged with the answers. Together with training and necessary cross-checking to make sure that nobody had made any mistakes, this could consume huge amounts of time, as well as being boring.

“I once had to do legal disclosure checking on a huge dispute where I was put in a room with documents piled to the ceiling and told to get on with it,” recalls Wendy Miller, a partner at the firm and a litigator in commercial real estate disputes.

This time the law firm turned to cognitive computing, which has begun to revolutionise much of the tedious work in legal analysis. The firm had already been looking for ways to improve efficiency. “What we do is very personnel-heavy,” says Ms Miller. Also, the documents were likely to arrive in near-random groups, making resource planning difficult. “You don’t want a team sitting around doing nothing, but it’s tricky if you then find you need 200 documents analysed by tomorrow,” she says.

The company turned to a British company which specialises in cognitive computing systems for information-intensive businesses. It designed a system which could scan the PDFs generated from the Land Registry and generate a spreadsheet that could be queried by the law team.

Compared with the 45 minutes it would take a human to examine each document, the RAVN system has already saved more than 500 work-hours. “The great efficiency of artificial intelligence is that we have complete flexibility because it’s always there in the background,” says Ms Miller.

So are the people who would have done that work out of a job? “Extracting data from documents isn’t perceived as valuable, so we were using junior people on work that’s hard to charge for,” she says. “Instead, we’ve been able to use those people on later stages of the project which have more value.”