Will an artificially intelligent robot steal your job?

Professor Michael Obsborne published a paper 2 years ago which suggested 47% of jobs are at high risk of being taken over by robots. 47% is a staggeringly high number but what jobs are safe from machines? And how will the take over happen? Graihagh Jackson put these questions to Michael...

Mike - As much as 47% of current US employment might be at high risk of automation over the next 20 years.

Graihagh - This was published a couple of years ago and there's a link to it on our website - naked scientists dot com and you know what? Michael is so well placed to talk about this, as he's not only an associate professor of machine learning at the University of Oxford but a co-director of the Oxford mountain programme of technology and employment... In other words he knows a lot about intelligent about machines and how they impact society.. 

Mike - So I have these two strands - these two hats that I wear.  One is an academic researcher in machine learning itself, developing novel intelligent algorithms capable of processing data in an intelligent, sophisticated way.   But the other hat, the second title that I mentioned before is investigating the societal impact of the development of such algorithms. So our second programme looks at the employment impact that intelligent algorithms are likely to have over the next 20 years.

Graihagh - And as you heard before, 47 % of US jobs are at HIGH risk of being taken over by robots. 47%...

Mike - It's a large figure - I should put some caveats on it, so that figure does not take into account the fact that firstly, what is actually automated would depend on regulations, societal acceptance, a whole host of other factors.  It would depend on relative costs of the machines and the humans involved in those tasks and crucially we're also not taking into account in that figure the emergence of new jobs.

Graihagh - Okay then.  So how did you work this number out then - this 47%?

Mike - We're drawing on data from the US from an agency called O-NET which provides for 700 different occupations, through to those occupations, a fairly long list of quantitative measures of skill requirement.  So these measures would be things like a number between 0 and 100 saying just how much persuasion this required, or finger dexterity, or originality...

Graihagh - How do you determine that sort of thing?

Mike - It's an excellent question and you'd have to ask O-NET, but drawing upon that data we thought that those skill requirements might be an interesting way of determining what distinguishes an automatable job from a non-automatable job. So we cross-linked the skill requirements against the jobs that we had seen automated and against another list of jobs that we were fairly confident were not going to be automated within our horizon of 20 years. Using those two different categories and relating it to those different skill requirements, we actually taught a machine learning algorithm exactly the same kind of algorithm we're expecting might have impact on employment, the difference between automatable or non-automatable jobs and, on that basis, it was able to come back and tell us just how much of it might be susceptible to automation.

Graihagh - What makes a job susceptible to automation?  There you talked about un-automatable jobs and automatable jobs.

Mike - So that was another conclusion from our study, the characteristics of jobs that are not automatable and we lumped these together into what we call bottlenecks to automation.  The first of those bottlenecks was originality.  Simply put, the more original your job is the less susceptible it is to automation.

Graihagh - So painters are safe?

Mike - Painters are relatively safe, yes.

Graihagh - What other skill sets did you decide were also un-automatable?

Mike - The second of our bottlenecks was social intelligence.  So here we're thinking of skills like negotiation or persuasion.  The kinds of high level social functions that come relatively natural to us but are relatively difficult to make explicit in such a way that those tasks could be reproduced by code.

Graihagh - And the third bottleneck...?

Mike - The third bottleneck was that of autonomous perception manipulation and this one's a little bit more subtle perhaps.  It certainly is possible to get a robot to interact with physical objects in the world around it but it's important to distinguish that kind of manipulation from the manipulation that we perform in our day to day lives.  So again, to give a concrete example, I'm sitting here in the studio picking up a glass of water and in doing so, I was required to distinguish the glass from the table that sits beneath it.   Despite the fact the glass is transparent, I needed to, before I even picked up the glass, have some idea of how much it weighed so I could grip it with sufficient force but also have some expectation for its material characteristics so that I didn't shatter it when I picked it up by applying too much force.  So you can see that, even in that relative intuitive action, I had to apply a whole host of subtle tacit knowledge about my environment that again, is very difficult to reproduce in an algorithm.

Graihagh - What would that be in terms of jobs though, because I'm not lucky enough to have someone pick up my glass of water every day and feed it to me, so where would I see that day to day?

Mike - So the kinds of jobs that might be non-automatable as a result would include hairdressing, for example, it might include gardening.

Graihagh - The bottlenecks you've been describing to me in some ways are things that are very natural to us.  They don't seem to have a defined set of rules that at least I could very easily distinguish and is that, in part, why they're very hard to code for and create an algorithm for?

Mike - Exactly.  We, as humans, draw upon these deep reservoirs of tacit knowledge about our society, our companions, our environment; all stuff that's very difficult to unpick and, as I say, write out explicitly in code.

Graihagh - We've talked about the jobs that are going to be relatively safe.  Who is then at risk because, surely, a little bit of these skills are involved in everyone's jobs?

Mike - One particular category of jobs that's likely at high risk are those jobs that rely almost entirely upon storing, accessing and perhaps doing some simple processing of data.  So I'm thinking here of examples including paralegals, whose job in large part is digging through case files.  I'm thinking about people like auditors who might be required to inspect large amounts of financial data in a company.  These are the kinds of tasks, to be honest, right now might be better performed by an algorithm which is able to scale to processing much larger volumes of data, is perhaps more vigilant, they're going to be just as attentive to the millionth bit of data as the first.  They're not going to, for example, be influenced by how long it's been since they had lunch.

Graihagh - I suppose they don't need a sick day or anything like that?

Mike - Exactly.

Graihagh - How about radio producers and presenters?

Mike - Right.  So, let me say that in this kind of social interaction that we're having, for example, there's quite likely a large amount of social intelligence that's deployed that's going to be difficult to reproduce in an algorithm so...

Graihagh - Phew.  I'm safe!

Mike - Congratulations.

Graihagh - And if you want to see how safe you are, NPR have designed a nifty web tool based on Michael's work that means you can select your profession and see if you're safe or not so safe. It's on our website - naked scientists dot com or just search for will your job be done by a machine.

Mike - Firstly, a lot is going to depend on how cheap the automated resolution is relative to the human labour.  Now the scary this is, of course, software has next to zero marginal cost of reproduction so once someone has developed and intelligent algorithm capable of doing a task that was previously performed by humans that software, at least in principle, can be deployed across the entire world at next to zero cost.  So, in many cases, we might see relatively rapid transition to automated software solutions once that technology is developed.

Graihagh - Are you surprised or, indeed, do you think things should become increasingly automated?

Mike - On this shrewd question, of course, we shouldn't lose sight of the fact that these automated solutions are delivering products to us at increasingly low cost.  On the other side of things, things are not so rosy; we shouldn't lose sight of the fact that while these jobs are replaced by an algorithm, they might not be jobs that any human fundamentally wants to do.  There's an argument to be made that if a job is able to replaced by an algorithm, it's kind of below the dignity of a human being and...

Graihagh - Wow, that's pretty...

Mike - It's a strong claim I have to agree.

Graihagh - Yes...

Mike - To me, those bottlenecks that I identified before; the creativity and social intelligence are really the hallmarks of what we as human beings find satisfying and pleasurable.  Consider the tasks we do in our spare time (our hobbies), they are almost universally things that involve some degree of interacting with other humans or being creative.  So, to me, the kinds of jobs that will remain after automation are going to be increasingly satisfying and enjoyable  

If a robot is going to take over the world, we've got to get to a more general artificial intelligence where a machine can think for itself. So, how do we do that? There are two ways and both are inspired by nature itself. One is by reverse engineering the brain and the other is by copying evolution. Graihagh Jackson sat down with Georgia Mills to see how this would work in practice...

Georgia - We're going to use evolution to make the next generation of super robots.

Graihagh - So if we take us humans as examples, how did we get clever?

Georgia - Well, there are a number of different theories on exactly when and how humans and our ancestors became intelligent and, even though we don't know the exact reason we became smart,

we do know the mechanism and that's evolution. So, the individuals who weren't so smart gradually got booted from the gene pool, essentially...

Graihagh - To put it nicely...

Georgia - Yes. Booted - dead. Meaning that as a species we got smarter and smarter and then, eventually, we get to where we are today where we are able to cure diseases, sending rockets into space...

Graihagh - I mean this is a really complicated process; it involves genes and environmental pressures. So how do you mimic this?

Georgia - Well, if we think about evolution it involves a code, so that's DNA, and when a living thing has offspring, there are tiny changes in this code which is maybe caused by mixing the DNA with another organism or by errors in the code, so that's a mutation and this causes variation. And if some things do better than others in this variation, we have what we call natural selection which is kind of the bare bones of evolution. So if you take this idea and change it from a biological code into computer code...

Graihagh - Simple!

Georgia - Simple! You can create a system where these little iterations crop up in each generation of code and then you can have a system in place, which acts as the natural selection, testing the code for the output you want. It selects the ones that are most successful and then the idea is, as each generation goes on, you get better, and better, and better at doing what you want to do.

Graihagh - All I'm imagining is robots having sex and I'm assuming that's not quite how it's going to work in practice...

Georgia - Maybe one day!

Graihagh - I'm not sure I'd want to witness that though.

Georgia - I would.

Graihagh - That is recording Georgia. You do realise everything I record is licensed to go in?

Georgia - You're not going to use that... So at the moment it's more like the idea of an algorithm that makes this little piece of software run around on a computer screen - that kind of very basic thing. But there is an example here in Cambridge where someone has built a robot that can make these little babies, test them and then make more and more babies that can go on and do better and better. I actually went to see this robot at a conference and I was introduced to it by a Louis Osprobreck. It looks to me like one of the stereotypical factory robots - a sort of overreaching arm. What is it actually trying to do?

Louis - It is a factory robot. It has a gripper and like a hot glue gun in front (an industrial one). So it only has two types of components; it has wooden cubes and he has those cubes with the motors in there and it glues them together to build the robot. We are trying to automate a design for a robot so it's randomly coming up for a robot that walks and we are iterating this design by testing it. So we built ten robots and we test how it works and then take the best ones and use like evolution to generate a next generation of robots. And then we built them again, test them again, and so on. And eventually the small robots can move forward when we turn on its motors. Over there, that's one of the working robots.

Georgia - Can I see it walk?

Louis - Sure...

Georgia - I don't know quite how to explain what just happened. The squares sort of had a spasm

and then it moved around. What's going on here?

Louis - As I say, we glue these cubes together and two of them have a motor in there which can move one side of the cube back and forth. We don't know in advance whether it's going to work and how it's going to work, but somehow it moves away.

Graihagh - They were doing what! They were crawling?

Georgia - Crawling is a kind way of putting it. Imagine two cubes just stuck together and sort of swizzling about and then, somehow, this kind of cube abomination sort of shuffles slowly across a table.

Graihagh - Sounds great, but what is this demonstrating.

Georgia - So this cube whizzing along is, actually, twice as good as the first cube this robot created but what was interesting, there was no input from a coder or an engineer to make this new one better - it was all left to the robot. So what happened is it made ten little babies and in these baby cubes there are these variations in quite how they're glued together, and then these ten babies are tested against each other and the fastest one of these cubes that twists along it's sort of said, you're the winner, you've survived, you're in the gene pool. The computer takes the design from that cube baby and says, this is the winner. Makes ten slight iterations from that design, like the mutations I mentioned earlier, and then does the whole process again. So the idea being each generation the designs get slightly, slightly, better and I think after ten generations they got twice as fast at crawling.

Graihagh - I imagine you'd be able to scale this up beyond just crawling into other skills that we might actually need to form something that resembles artificial intelligence then?

Georgia - So there are examples of people using this kind of mimicking of evolution to design the best kind of antennae or to design mirrors that can funnel light for solar panels, and things like that and there's no reason, in principle, that it can't be used to try and make intelligent computers.

Graihagh - In a way then these computers are thinking outside the box. They're thinking in ways that humans wouldn't be able to think unless thinking of ways that they can change and evolve to be much better in their design than perhaps we even could design them to be.

Georgia - Well, I'd hold you back on the word thinking there. There is no thinking going on. It is just making random changes in this code and then, by just simply testing and changing, and testing and changing again, the actual effect we get is something that looks like this robot has been really creative and has solved the problem but really, it's just using this random process of evolution.

Graihagh - By having this process being random, I'm assuming that's not perfect. There aren't environmental pressures which say, do this, it's completely random, so surely there's got to be a catch here.

Georgia - You've got your finger on it right there. So the thing about evolution; it doesn't say let's go and have great big brains, that's exactly what we need. There are a whole host of things that might be selected for. There's also the problem that evolution, unfortunately, takes millions of years so we want to speed the process up. And the other problem is mutations - I mentioned them earlier -they're often bad. If you think about mutations, they're associated with things like cancer, but we have the advantage here so we can be a bit sneaky. We can say only good mutations in the code from now on, we can speed up generation times and things like that, and we can say we only want it to be selected for intelligence. So we have these advantages here over evolution, but it's still not clear if it will actually be enough to get these super-intelligent computers.

There may be a long way to go when it comes to figuring out the details of creating super intelligent, humanoid robots - in principle though, it's possible so let's say, hypothetically, we were able to make an a super intelligent algorithm.  How would we get from something that operates from a hard drive to the likes of Skynet? Kat Arney spoke to Martin Robbins to find out whether something like this is even possible...

Martin - No!... Or if you want the long answer... No! 

Graihagh - Well, that's shut me up hasn't it...

Martin -   My name's Martin Robbins.  I've worked in A.I. in industry for the last eight or nine years and I write the Guardian's "Raising Hal" blog on artificial intelligence and data science.

Graihagh - Martin and our Kat Arney sat down for a cuppa to mull over why the terminator isn't going to become a reality any time soon...

Martin - The idea of the singularity has been around in one form or another since the 1950's and it's become more popular in the last 20 or 30 years.  The idea behind it is that at a certain point of development, an A.I. would become better at producing successive versions of itself than a human would. Once that happens, then it's evolution would become, essentially, unpredictable and, essentially, the singularity is the point at which we have no idea what would happen.

Kat - Do you actually see that happening in the future?

Martin - I think it's a question that's almost impossible to answer.  In a way it's very similar to the Victorians speaking what constitutes life and how do we create life.  So, if you go back 120 years, people like Mary Shelley were obsessed with the idea that you could reanimate a corpse and somehow take a inanimate object and imbue in it with the qualities of life.  The problem with that is you immediately start falling into all these semantic questions and practical questions about what do you mean by life.  What does it mean to give life to an object and you hit the exact same problems when you start talking about sentience.  Until you've got a meaningful working definition of what it means to be sentient that comes beyond just philosophy and, actually, has some kind of practical, real world equivalents.  There's just no real way to answer the question.

Kat - I guess one of the problems is when we see depictions of artificial intelligence, robots for want of a better word, in films, there's this idea that they have feelings, thoughts, they have interactions with humans.  Is this just us projecting onto A.I.s what we think that they might be -  building them in our own image in our own minds?

Martin - Well that's part of the problem.  Every time we talk about something like sentients, we immediately start to bring to that word all the baggage that we have from our own experiences of being a sentient creature.  So we talk about artificial intelligence and remember that artificial intelligence at its root is purely about trying to recreate mechanisms of intelligence and problem solving that humans use.  As soon as you start talking about motivation, or emotions, or feelings, or goals, or objectives, all of these other things, you're going beyond the original remit and you're starting to link intelligence to lots of things that are basically part of the human experience and this is one of the problems really.  If you look at the way we relate to artificial intelligence, it's incredibly ego driven and it becomes less about what would an intelligent system be like and it becomes more about how we create something that we can relate to and that sort of reproduces us in an immortal or god-like form.

Kat - Is there any inkling that an artificial intelligence would like humans, hate humans, want to destroy us?

Martin - You've literally no idea what direction this could go in.  It might want global conquest or it might be happy like trolling chess programmes.  There's no reason to think it's going to do one thing or the other.

Kat - Particularly in films or in popular culture, there's this idea that intelligent robots will rise.  The idea of a clone army or robotic artificial intelligences taking over but I've seen videos robots trying to walk, trying to do stuff.  I think there's quite a long way to go here - am I right?

Martin - Yes, I'd say so.  There's obviously been a lot of interest recently in companies like Boston Dynamics who've been experimenting with robots that walk on legs.  At the moment, the state of walking robots is pretty poor.  They can go a short distance but they need a lot of supervision.  There's no way that you take a robot and send it off across several miles of open terrain and expect it to get back again in one piece unless it was on wheels and the terrain was reasonably flat.

Kat - So no marauding armies of robots anytime soon?

Martin - No - not soon anyway.