The Countdown to Artificial Intelligence

Globally, crime committed online costs economies and individuals over 3 trillion Dollars per year at the moment, and that’s predicted to more than double within the next 5 years. So can Artificial Intelligence help to combat the threat, or will it help to fuel the fire? Georgia Mills spoke to Dave Palmer, who works with the Cambridge based cybersecurity company Darktrace...

Dave - Darktrace’s interest is can we use advanced mathematics in AI to really replicate the idea of an immune system where we know the normal self of everyone and everything inside of a business and how they all relate to each other? If someone or something's behaving really strangely then we can start to deal with that problem before it gets to the point where millions of credit card details are lost or medical records, or a manufacturing plant gets shut down.

Georgia - How would that work in practice?

Dave - We’re very predictable in how we behave using our smartphones and out laptops and all the different technology that exists within particularly businesses. So by understanding what it means to be me, what it means to be Dave and how I use my email and all the pieces of technology inside of my business, then we can tell if perhaps my laptop's been infected and is starting to hoard data, or communicate with the outside world in a way that suggests it might be under someone else's control, and then we can start to do something about it. Now that could either be telling a human being hey, here’s a problem and you should go a check it out. Or, increasingly, how the cyber security industry's going to be moving into autonomous response, having the machines on our behalf start to deal with problems and slow then down, or even potentially clean them up in the longer term.

Georgia - What’s the machine learning aspect of this technology?

Dave - Imagine a modern business, or even look around wherever you are now, you’ll start to see technology everywhere whether it’s digital phones, smart TVs, video conferencing units and, of course, all the things we take for granted like laptops, and smartphones and data centres and the cloud. There’s an enormous amount of complexity there. It’s not unusual to find in an organisation of 10,000 people that there are probably at least 50,000 pieces of technology as a rule of thumb.

So using the AI techniques to be able to learn what’s normal and really truly understand the relationships between all those technologies and all those people, instead of asking the humans to do it is really very useful indeed. Then the humans can just be told about the things that are interesting instead of having to try and wade through all of that complexity and guess everything that might go wrong.

Georgia - You mentioned Darktrace spots unusual activities straight away if something’s not quite right, but is there a way to sort of block the holes before they’re entered in the first place? Is it possible to use machine learning to have a hackable-proof system?

Dave - I’m very cautious about saying yes to that given where we are in a society under considerable digital attack at the moment. I think the things that’s really hard about cyber security is there isn’t a perfect answer on what secure looks like. Every part of our digital life is based on millions, if not billions, of lines of code written by different people from all over the world, and different companies, and supply chains that are very deep indeed. So the idea of using machines to go through and evaluate the riskiness of every single line of code and piece of software that we kind of take for granted in the interactions that we have on a daily basis, I think it’s quite far away. I think we need to have made an awful lot of progress on AI before it’s smart enough to do that. Getting much closer to artificial general intelligence than the artificial narrow intelligences we have today.

But that said, I think AI will start changing everything in the cyber security sector. It think there will be replacements for the antivirus that we all run on our laptops with something that’s AI enhanced and better at stopping bad stuff.

Georgia - Does the technology make it easier or more difficult for people like you to protect our data?

Dave - We can definitely expect AI to start making spam much more effective than it’s been in the past. There’s a really quick example: imagine my laptop got hacked and a piece of AI software on my laptop was able to train itself on all of my emails, my calendar, my imessages, my whatsapp, it would then be able to individually communicate with all the people in my life replicating my communications style to spread itself. So you and I have a shared diary appointment to talk today, perhaps it sends a little note to you saying oh, I have some questions, could you have a look at this attachment and let me know what you think. I think you’d probably open that email because it’s going to sound like it's from me and it’s going to be contextually relevant.

Are we at risk from being put out of work from an artificially intelligent machine? Chris Smith was fairly confident that podcast presenters were safe, but Simon Beard from the Centre for study of Existential Risk had other ideas.

Simon - You say that it won’t take your job any time soon. Actually, these things are very hard to predict because of the non-linearity. Yes, you couldn’t plug Alexa into the decks and she’d run the programme but it’s not impossible that the next algorithm that comes out will be a whole step change better and you will get artificial radio presenters way before we know it.

It’s very hard to know what’s going to happen in the short term. It’s not so hard to know what’s going to happen in the long term just because we’ve got a lot of years of data here about how machines, and computers, and artificial intelligence have worked and you do get quite predictable long term patterns of improvement.

In the long term, the answer to this question on one level does seem like it’s yes. It’s hard to imagine a job which you could not get an algorithm to do at least as well as a person and more cheaply, or probably much better than a person could and more cheaply as well.

However, that’s only the first level. That’s the kind of rational economics view of employment and we know that it doesn’t work like that. Actually, people spend a lot of money on things that aren’t about doing a job better, or doing it more cheaply. People’s habits are based on ethical values; they’re based on social interactions; they’ve based on status, on appearance, on how things look to other people. There are very good reasons to think that human beings will have an edge from many, many of those things for a long time to come.

People care about people; we like to interact with people even if it’s not so good sometimes as interacting with a machine - it’s not so efficient. At the moment, we’re more used to it being frustrating to interact with machines but as I say, in the long run, that is likely to change. But we still like people; we still like to have handmade goods even if they’re not so well made we pay more for them.

So, on that basis, I find it hard to see a situation in which people will be written out of the job market, in which there will be no jobs for people, but the reason why people employ other people will be different. It won’t just be about getting the job done as cheaply as possible; we won’t be employing people in sweatshops or on the minimum wage. There will be no reason to do that. The reason for employing people are going to be social; they’re going to be value based; they’re going to be status based.

Chris - Do you not think there’s going to be a problem then? Because, if what you’re saying is true, and there are very few jobs that would not be predated by artificial intelligence, not necessarily tomorrow but in the future, does that not add up to a recipe for a kind of mass panic and people being out of jobs, and a meltdown in the fabric of society which, for thousands of years, has been based around being paid for industry?

Peter - Well so we have luckily seen ourselves go through a variety of this kind of step change where the reason people get employed to do things, and the way they get employed to do things changes dramatically. It changed during the industrial revolution, it changed during the first wave of automation, and the ending of the industrial revolution in developed countries, and these are definitely stressful, difficult transitions. They do link to increases in violence and political dissatisfaction but, on the other hand, we are an amazingly adaptive species. They don’t produce social meltdown. Sometimes they produce very local revolution but, actually, more often they don’t.

Chris - Do we need some legislation in place though? Because are we not sleepwalking a bit into a situation where suddenly we might find things changing very, very quickly and we have not got any laws in place, or anything to make sure that companies do do right by their workers?

Peter - Absolutely we do. But the legislation we need needs to be forward looking, not past looking. One of the examples that I use is that if you consider the industrial revolution, it was the same set of technological changes that produced very egalitarian societies like Sweden and Japan, and also very unequal societies like the USA and China.

Government regulation has an awful lot to do with that. But, a big part of it is we can’t legislate on the basis of just keeping hold of what we’ve got at the moment. Not only because that will be ineffective, but because the wrong sort of legislation is likely to lead to bad outcomes. We have to look at what is coming and what we want to get out of that. So that’s going to mean keeping hold of working protection but refocusing it. We’re not so worried about working protection in terms of industrial accident, we should be now much more worried about worker’s rights in terms of social and emotional strain because that’s what people are going to be employed to do. It’s going to be providing these social and emotional services so the kind of exploitation you’re going to see is going to be exploitation of the people’s emotional resilience, and the social connectedness, and so on. That’s the sort of thing we need to protect against going forward.

Georgia Mills put this to philosopher of mind Henry Shevlin...

Henry - Consciousness is, on the one hand, one of the greatest mysteries that we still face as a species, but it is something where we are getting greater understanding from neuroscience and cognitive science more broadly. Still though, I think a lot of scientists would really be happy if we stopped talking about consciousness, but I don’t think we can. Consciousness is deeply bound up with our ideas about morality and value. Just to give a simple example, if we’re thinking about the ethics of boiling a lobster alive, which is often how lobsters are cooked, I think the key question we’re going to ask in thinking about whether that’s humane is does the lobster feel pain? Does it have conscious experience of it’s pain? So, I think, we can’t ignore problems about consciousness. Broadly speaking, there’s no reason to think we couldn’t build a conscious AI, but it may not be very obvious when we’ve done so. There’s still no clear consensus about what consciousness is for and what kind of functions in the human mind it’s associated with. So, even if we do build a conscious AI, we may not know it right away.

Georgia - Alexa; are you conscious?

Alexa - I know who I am - let’s put it that way.

Georgia - Alexa could have said anything there so I guess this takes us to the point how would we ever know if something was conscious or it wasn’t?

Henry - The gold standard for thinking about tests of consciousness is a thought experiment dating back to the amazing British polymath Alan Turing. In 1950 he posed this test, which is now known as the Turing test, where basically the idea is if you have a computer system in one room and a human in another and you’re talking to them both via a terminal so you can’t see which one is which. And you’re at a chance when you’re forced to guess as to which one is the computer and which one is the human. In other words, the computer can completely fool you into thinking it’s a human being. At that point, Turing says, we’ve got no business denying consciousness and intelligence to the computer system because it’s passing itself off completely as a human being.

Now that I think is quite a popular test for consciousness and it’s been very influential, but it did face some push back in the 80s by a classic followup thought experiment by American philosopher John Searle who came up with this idea of what he called The Chinese Room. And basically, the point of this is to show that you can get stuff that looks like real intelligent understanding which actually is very simple. The way the Chinese Room works is if you imagine you’ve got someone who doesn’t speak and Chinese at all and they’re sitting in a little booth surrounded by index cards. There’s a slot in the booth and you can post a question in Chinese script in through this slot. What the person inside the booth does - they don’t understand Chinese - but they can look in the index cards, each of which tell you what to reply in Chinese.

So the index card gets posted in, they find the corresponding answer card, write out these symbols that they don’t understand - they’re just copying them, and they post it out. The idea is if you had a big enough library of index cards so that all possible questions and answers were covered, this person could do a brilliant job of simulating that they understand Chinese whilst, in fact, all that they’d be doing is following a simple lookup table of instructions.

Now extending this to human beings, the idea is maybe we could have a computer that passed the Turing test by doing basically that. It’s got a huge, massive database of possible questions and answer responses it can give, but it doesn’t understand the words. So that’s put some pressure on the Turing test as this measure of consciousness. Maybe a computer could trick us into thinking it’s conscious by actually doing something pretty dumb.

Chris - Does it also matter, Henry, if a machine is conscious? Does that really matter?

Henry - I think it matters in two ways. The first is we actually start to worry about our ethical treatment of machines if there is a reason to think they’re conscious. This is something that’s explored in shows like West World or Humans where you have these tools basically that people use for their basest instincts. Maybe if you’re just dealing with things that are basically human shaped robotic vacuum cleaners, that doesn't make a difference. But, if you dealing with things with real emotions or real cognitive capabilities, then you might start to think we need to regulate that behaviour.

The other reason you think it might matter is if people start getting relationships, friendships,or even romantic relationships with AIs, as we’ve seen in movies like Her, then it might matter to the people involved. They want to know that not just that their AI can simulate love or emotions but it’s actually feeling and reciprocating these things.

Chris - Isn’t one of the attractions of things like AI that actually it’s not biased by many of the what we regard as human flaws like innate biases, or prejudices, or emotions that get in the way of a decision where your heart rules your head? And if we end up with machines that become a bit conscious and they start doing things very well like we do, they’ll end up like us, so they’ll think like us and be flawed like us, won’t they?

Henry - We can decide as we develop these machines what kind of architectures, what structures we want to put in place in terms of how their cognitive systems develop. So we might be able to make some conscious choices to avoid the biases along the way. I don’t think consciousness necessarily means the full emotional complexity of human beings, it could just be something as simple as feeling pain as a case of consciousness.

Nonetheless, I think there are reasons to worry about importing human biases over into machines. Particularly given that a lot of AI progress in recent years has been driven by mass data revolution. Taking data out there on the internet, plugging it into computational systems. There’s a risk in doing that in things like stereotypes about gender or race get imported into the AI just because they’re being exposed to the internet and all the different biases people have.

Georgia - Yes. In fact, wasn’t there an episode earlier in the year, Microsoft created a Twitter bot to learn from the rest of Twitter and I think within 24 hours it had become racist?

Henry - Yeah, pretty unsurprising, and you find things like gender stereotypes about profession. Just as Google or autocomplete faces worries about the kind of things people most frequently search for often carry hints of gender bias or racial bias. If that’s just the data we’re plugging into the machines, then it’s a bit much to expect them to be better than us.

Georgia - Could this be a real problem in society if people like the police end up using AI to do their jobs, could we end up with something that we assume is neutral, then not being, and this could cause problems?

Henry - Yeah. So this is a real motivation for a desire for transparency in AI. It’s particularly complicated these days because a lot of the algorithms that are being used are basically blind. We don’t know how they’re solving the problems. We can sort of reverse engineer them, but because there’s a lot of self-learning going on, it’s not always immediately clear why a system is making the decisions it’s making. If a system singles out a person for a stop and search, or someone going through airport customs gets singled out, we may not simply be able to say okay, what were the criteria used it it’s been subject to this kind of elaborate self-learning system, so we need to think about transparency too.

Georgia - Peter; as someone works on this, what would you say to this question of consciousness?

Peter - I think it’s very interesting. I still having thought quite a lot about it don’t really know what consciousness actually is. I think there’s that sort of sense of you being here and now and living in that moment and feeling, if that’s the definition of consciousness then I think that probably extends across the vast amount of the animal world. And I think if we’re going to worry about consciousness in artificial intelligence, we probably need to spend more time worrying about consciousnesses within nature.

Chris - Simon; a quick thought from you?

Simon - The definition of consciousness is the key thing here. Consciousness is an objective experience and we all feel we know what it is, but there have been various attempts to define consciousness in ways that allow us to make more progress in identifying it in animals or machines. So defining it as the property of informational systems by various axioms of conscious systems have so they're indivisible, they model the universe, they model themselves in relation to the universe, and so on.

I think that these kind of theories give us a lot of hope for understanding consciousness. My only problem is I think when a lot of people talk about consciousness they want it to be magic. They want it to be unexplainable and so if we come up with a system like this that does allow us to say, for instance, is Alexa conscious? Then people just won’t buy it because we feel that our consciousness is the last domain of our uniqueness as a species and we’re quite strongly set up to defend that. But, in defending it we tend to come up with phrases and questions that are just unanswerable, undefinable, and that is really getting in the way of clearly thinking through the ethics of AI.

Georgia - I think is one we could debate till the cows come home. Maybe something to take to the pub after the show.