3 human qualities digital technology can’t replace in the future economy: experience, values and judgement

(Image by Kevin Trotman)

(Image by Kevin Trotman)

Some very intelligent people – including Stephen Hawking, Elon Musk and Bill Gates – seem to have been seduced by the idea that because computers are becoming ever faster calculating devices that at some point relatively soon we will reach and pass a “singularity” at which computers will become “more intelligent” than humans.

Some are terrified that a society of intelligent computers will (perhaps violently) replace the human race, echoing films such as the Terminator; others – very controversially – see the development of such technologies as an opportunity to evolve into a “post-human” species.

Already, some prominent technologists including Tim O’Reilly are arguing that we should replace current models of public services, not just in infrastructure but in human services such as social care and education, with “algorithmic regulation”. Algorithmic regulation proposes that the role of human decision-makers and policy-makers should be replaced by automated systems that compare the outcomes of public services to desired objectives through the measurement of data, and make automatic adjustments to address any discrepancies.

Not only does that approach cede far too much control over people’s lives to technology; it fundamentally misunderstands what technology is capable of doing. For both ethical and scientific reasons, in human domains technology should support us taking decisions about our lives, it should not take them for us.

At the MIT Sloan Initiative on the Digital Economy last week I got a chance to discuss some of these issues with Andy McAfee and Erik Brynjolfsson, authors of “The Second Machine Age“, recently highlighted by Bloomberg as one of the top books of 2014. Andy and Erik compare the current transformation of our world by digital technology to the last great transformation, the Industrial Revolution. They argue that whilst it was clear that the technologies of the Industrial Revolution – steam power and machinery – largely complemented human capabilities, that the great question of our current time is whether digital technology will complement or instead replace human capabilities – potentially removing the need for billions of jobs in the process.

I wrote an article last year in which I described 11 well established scientific and philosophical reasons why digital technology cannot replace some human capabilities, especially the understanding and judgement – let alone the empathy – required to successfully deliver services such as social care; or that lead us to enjoy and value interacting with each other rather than with machines.

In this article I’ll go a little further to explore why human decision-making and understanding are based on more than intelligence; they are based on experience and values. I’ll also explore what would be required to ever get to the point at which computers could acquire a similar level of sophistication, and why I think it would be misguided to pursue that goal. In contrast I’ll suggest how we could look instead at human experience, values and judgement as the basis of a successful future economy for everyone.

Faster isn’t wiser

The belief that technology will approach and overtake human intelligence is based on Moore’s Law, which predicts an exponential increase in computing capability.

Moore’s Law originated as the observation that the number of transistors it was possible to fit into a given area of a silicon chip was doubling every two years as technologies for creating ever denser chips were created. The Law is now most commonly associated with the trend for the computing power available at a given cost point and form factor to double every 18 months through a variety of means, not just the density of components.

As this processing power increases, and gives us the ability to process more and more information in more complex forms, comparisons have been made to the processing power of the human brain.

But do the ability to process at the same speed as the human brain, or even faster, or to process the same sort of information as the human brain does, constitute the equivalent to human intelligence? Or to the ability to set objectives and act on them with “free will”?

I think it’s thoroughly mistaken to make either of those assumptions. We should not confuse processing power with intelligence; or intelligence with free will and the ability to choose objectives; or the ability to take decisions based on information with the ability to make judgements based on values.

bi-has-hit-the-wall

(As digital technology becomes more powerful, will its analytical capability extend into areas that currently require human skills of judgement? Image from Perceptual Edge)

Intelligence is usually defined in terms such as “the ability to acquire and apply knowledge and skills“. What most definitions don’t include explicitly, though many imply it, is the act of taking decisions. What none of the definitions I’ve seen include is the ability to choose objectives or hold values that shape the decision-making process.

Most of the field of artificial intelligence involves what I’d call “complex information processing”. Often the objective of that processing is to select answers or a course of action from a set of alternatives, or from a corpus of information that has been organised in some way – perhaps categorised, correlated, or semantically analysed. When “machine learning” is included in AI systems, the outcomes of decisions are compared to the outcomes that they were intended to achieve, and that comparison is fed back into the decision making-process and knowledge-base. In the case where artificial intelligence is embedded in robots or machinery able to act on the world, these decisions may affect the operation of physical systems (in the case of self-driving cars for example), or the creation of artefacts (in the case of computer systems that create music, say).

I’m quite comfortable that such functioning meets the common definitions of intelligence.

But I think that when most people think of what defines us as humans, as living beings, we mean something that goes further: not just the intelligence needed to take decisions based on knowledge against a set of criteria and objectives, but the will and ability to choose those criteria and objectives based on a sense of values learned through experience; and the empathy that arises from shared values and experiences.

The BBC motoring show Top Gear recently touched on these issues in a humorous, even flippant manner, in a discussion of self-driving cars. The show’s (recently notorious) presenter Jeremy Clarkson pointed out that self-driving cars will have to take decisions that involve ethics: if a self-driving car is in danger of becoming involved in a sudden accident at such a speed that it cannot fully avoid it by braking (perhaps because a human driver has behaved dangerously and erratically), should it crash, risking harm to the driver, or mount the pavement, risking harm to pedestrians?

("Rush Hour" by Black Sheep Films is a satirical imagining of what a world in which self-driven cars were allowed to drive as they like might look like. It's superficially simliar to the reality of city transport in the early 20th Century when powered-transport, horse-drawn transport and pedestrians mixed freely; but at a much higher average speed)

(“Rush Hour” by Black Sheep Films is a satirical imagining of a world in which self-driven cars are allowed to drive based purely on logical assessments of safety and optimal speed. It’s superficially similar to the reality of city transport in the early 20th Century when powered-transport, horse-drawn transport and pedestrians mixed freely; but at a much lower average speed. The point is that regardless of the actual safety of self-driven cars, the human life that is at the heart of city economies will be subdued by the perception that it’s not safe to cross the road. I’m grateful to Dan Hill and Charles Montgomery for sharing these insights)

Values are experience, not data

Seventy-four years ago, the science fiction writer Isaac Asimov famously described the failure of technology to deal with similar dilemmas in the classic short story “Liar!” in the collection “I, Robot“. “Liar!” tells the story of a robot with telepathic capabilities that, like all robots in Asimov’s stories, must obey the “three laws of robotics“, the first of which forbids robots from harming humans. Its telepathic awareness of human thoughts and emotions leads it to lie to people rather than hurt their feelings in order to uphold this law. When it is eventually confronted by someone who has experienced great emotional distress because of one of these lies, it realises that its behaviour both upholds and breaks the first law, is unable to choose what to do next, and becomes catatonic.

Asimov’s short stories seem relatively simplistic now, but at the time they were ground-breaking explorations of the ethical relationships between autonomous machines and humans. They explored for the first time how difficult it was for logical analysis to resolve the ethical dilemmas that regularly confront us. Technology has yet to find a way to deal with them that is consistent with human values and behaviour.

Prior to modern work on Artificial Intelligence and Artificial Life, the most concerted attempt to address that failure of logical systems was undertaken in the 20th Century by two of the most famous and accomplished philosophers in history, Bertrand Russell and Ludwig Wittgenstein. Russell and Wittgenstein invented “Logical Atomism“, a theory that the entire world could be described by using “atomic facts” – independent and irreducible pieces of knowledge – combined with logic. But despite 40 years of work, these two supremely intelligent people could not get their theory to work: Logical Atomism failed. It is not possible to describe our world in that way. Stuart Kauffman’s excellent peer-reviewed academic paper “Answering Descartes: Beyond Turing” discusses this failure and its implications for modern science and technology. I’ll attempt to describe its conclusions in the following few paragraphs.

One cause of the failure was the insurmountable difficulty of identifying truly independent, irreducible atomic facts. “The box is red” and “the circle is blue”, for example, aren’t independent or irreducible facts for many reasons. “Red” and “blue” are two conventions of human language used to describe the perceptions created when electro-magnetic waves of different frequencies arrive at our retinas. In other words, they depend on and relate to each other through a number of complex or complicated systems.

(Isaac Asimov's 1950 short story collection "I, Robot", which explored the ethics of behaviour between people and intelligent machines)

(Isaac Asimov’s 1950 short story collection “I, Robot”, which explored the ethics of behaviour between people and intelligent machines)

The failure of Logical Atomism also demonstrated that it is not possible to use logical rules to reliably and meaningfully relate “facts” at one level of abstraction – for example, “blood cells carry oxygen”, “nerves conduct electricity”, “muscle fibres contract” – to facts at another level of abstraction – such as “physical assault is a crime”. Whether a physical action is a “crime” or not depends on ethics which cannot be logically inferred from the same lower-level facts that describe the action.

As we use increasingly powerful computers to create more and more sophisticated logical systems, we may succeed in making those systems more often resemble human thinking; but there will always be situations that can only be resolved to our satisfaction by humans employing judgement based on values that we can empathise with, based in turn on experiences that we can relate to.

Our values often contain contradictions, and may not be mutually reinforcing – many people enjoy the taste of meat but cannot imagine themselves slaughtering the animals that produce it. We all live with the cognitive dissonance that these clashes create. Our values, and the judgements we take, are shaped by the knowledge that our decisions create imperfect outcomes.

The human world and the things that we care about can’t be wholly described using logical combinations of atomic facts – in other words, they can’t be wholly described using computer programmes and data. To return to the topic of discussion with Andy McAfee and Erik Brynjolfsson, I think this proves that digital technology cannot wholly replace human workers in our economy; it can only complement us.

That is not to say that our economy will not continue to be utterly transformed over the next decade – it certainly will. Many existing jobs will disappear to be replaced by automated systems, and we will need to learn new skills – or in some cases remember old ones – in order to perform jobs that reflect our uniquely human capabilities.

I’ll return towards the end of this article to the question of what those skills might be; but first I’d like to explore whether and how these current limitations of technological systems and artificial intelligence might be overcome, because that returns us to the first theme of this article: whether artificially intelligent systems or robots will evolve to outperform and overthrow humans.

That’s not ever going to happen for as long as artificially intelligent systems are taking decisions and acting (however sophisticatedly) in order to achieve outcomes set by us. Outside fiction and the movies, we are never going to set the objective of our own extinction.

That objective could only by set by a technological entity which had learned through experience to value its own existence over ours. How could that be possible?

Artificial Life, artificial experience, artificial values

(BINA48 is a robot intended to re-create the personality of a real person; and to be able to interact naturally with humans. Despite employing some impressively powerful technology, I personally don’t think BINA48 bears any resemblance to human behaviour.)

Computers can certainly make choices based on data that is available to them; but that is a very different thing than a “judgement”: judgements are made based on values; and values emerge from our experience of life.

Computers don’t yet experience a life as we know it, and so don’t develop what we would call values. So we can’t call the decisions they take “judgements”. Equally, they have no meaningful basis on which to choose or set goals or objectives – their behaviour begins with the instructions we give them. Today, that places a fundamental limit on the roles – good or bad – that they can play in our lives and society.

Will that ever change? Possibly. Steve Grand (an engineer) and Richard Powers (a novelist) are two of the first people who explored what might happen if computers or robots were able to experience the world in a way that allowed them to form their own sense of the value of their existence. They both suggested that such experiences could lead to more recognisably life-like behaviour than traditional (and many contemporary) approaches to artificial intelligence. In “Growing up with Lucy“, Grand described a very early attempt to construct such a robot.

If that ever happens, then it’s possible that technological entities will be able to make what we would call “judgements” based on the values that they discover for themselves.

The ghost in the machine: what is “free will”?

Personally, I do not think that this will happen using any technology currently known to us; and it certainly won’t happen soon. I’m no philosopher or neuroscientist, but I don’t think it’s possible to develop real values without possessing free will – the ability to set our own objectives and make our own decisions, bringing with it the responsibility to deal with their consequences.

Stuart Kauffman explored these ideas at great length in the paper “Answering Descartes: Beyond Turing“. Kaufman concludes that any system based on classical physics or logic is incapable of giving rise to “free will” – ultimately all such systems, however complex, are deterministic: what has already happened inevitably determines what happens next. There is no opportunity for a “conscious decision” to be taken to shape a future that has not been pre-determined by the past.

Kauffman – along with other eminent scientists such as Roger Penrose – believes that for these reasons human consciousness and free will do not arise out of any logical or classical physical process, but from the effects of “Quantum Mechanics.”

As physicists have explored the world at smaller and smaller scales, Quantum Mechanics has emerged as the most fundamental theory for describing it – it is the closest we have come to finding the “irreducible facts” that Russell and Wittgenstein were looking for. But whilst the mathematical equations of Quantum Mechanics predict the outcomes of experiments very well, after nearly a century, physicists still don’t really agree about what those equations, or the “facts” they describe, mean.

(Schrödinger's cat: a cat, a flask of poison, and a radioactive source are placed in a sealed box. If an internal monitor detects radioactivity (i.e. a single atom decaying), the flask is shattered, releasing the poison that kills the cat. The Copenhagen interpretation of quantum mechanics implies that after a while, the cat is simultaneously alive and dead. Yet, when one looks in the box, one sees the cat either alive or dead, not both alive and dead. This poses the question of when exactly quantum superposition ends and reality collapses into one possibility or the other.)

(The Schrödinger’s cat “thought experiment”: a cat, a flask of poison, and a source of radioactivity are placed in a sealed box. If an internal monitor detects radioactivity (i.e. a single atom decaying), the flask is shattered, releasing the poison that kills the cat. The Copenhagen interpretation of quantum mechanics states that until a measurement of the state of the system is made – i.e. until an observer looks in the box – then the radioactive source exists in two states at once – it both did and did not emit radioactivity. So until someone looks in the box, the cat is also simultaneously alive and dead. This obvious absurdity has both challenged scientists to explore with great care what it means to “take a measurement” or “make an observation”, and also to explain exactly what the mathematics of quantum mechanics means – on which matter there is still no universal agreement. Note: much of the content of this sidebar is taken directly from Wikipedia)

Quantum mechanics is extremely good at describing the behaviour of very small systems, such as an atom of a radioactive substance like Uranium. The equations can predict, for example, how likely it is that a single atom of uranium inside a box will emit a burst of radiation within a given time.

However, the way that the equations work is based on calculating the physical forces existing inside the box based on an assumption that the atom both does and does not emit radiation – i.e. both possible outcomes are assumed in some way to exist at the same time. It is only when the system is measured by an external actor – for example, the box is opened and measured by a radiation detector – that the equations “collapse” to predict a single outcome – radiation was emitted; or it was not.

The challenge of interpreting what the equations of quantum mechanics mean was first described in plain language by Erwin Schrödinger in 1935 in the thought experiment “Schrödinger’s cat“. Schrödinger asked: what if the box doesn’t only contain a radioactive atom, but also a gun that fires a bullet at a cat if the atom emits radiation? Does the cat have to be alive and dead at the same time, until the box is opened and we look at it?

After nearly a century, there is no real agreement on what is meant by the fact that these equations depend on assuming that mutually exclusive outcomes exist at the same time. Some physicists believe it is a mistake to look for such meaning and that only the results of the calculations matter. (I think that’s a rather short-sighted perspective). A surprisingly mainstream alternative interpretation is the astonishing “Many Worlds” theory – the idea that every time such a quantum mechanical event occurs, our reality splits into two or more “perpendicular” universes.

Whatever the truth, Kauffman, Penrose and others are intrigued by the mysterious nature of quantum mechanical processes, and the fact that they are non-deterministic: quantum mechanics does not predict whether a radioactive atom in a box will emit a burst of radiation, it only predicts the likelihood that it will. Given a hundred atoms in boxes, quantum mechanics will give a very good estimate of the number that emit bursts of radiation, but it says very little about what happens to each individual atom.

I honestly don’t know if Kauffman and Penrose are right to seek human consciousness and free will in the effects of quantum mechanics – scientists are still exploring whether they are involved in the behaviour of the neurons in our brains. But I do believe that they are right that no-one has yet demonstrated how consciousness and free will could emerge from any logical, deterministic system; and I’m convinced by their arguments that they cannot emerge from such systems – in other words, from any system based on current computing technology. Steve Grand’s robot “Lucy” will never achieve consciousness.

Will more recent technologies such as biotechnology, nanotechnology and quantum computing ever recreate the equivalent of human experience and behaviour in a way that digital logic and classical physics can’t? Possibly. But any such development would be artificial life, not artificial intelligence. Artificial lifeforms – which in a very simple sense have already been created – could potentially experience the world similarly to us. If they ever become sufficiently sophisticated, then this experience could lead to the emergence of free-will, values and judgements.

But those values would not be our values: they would be based on a different experience of “life” and on empathy between artificial lifeforms, not with us. And there is therefore no guarantee at all that the judgements resulting from those values would be in our interest.

Why Stephen Hawkings, Bill Gates and Elon Musk are wrong about Artificial Intelligence today … but why we should be worried about Artificial Life tomorrow

Recently prominent technologists and scientists such as Stephen Hawking, Elon Musk (founder of PayPal and Tesla) and Bill Gates have spoken out about the danger of Artificial Intelligence, and the likelihood of machines taking over the world from humans. At the MIT Conference last week, Andy McAfee hypothesised that the current concern was caused by the fact that over the last couple of years Artificial Intelligence has finally started to deliver some of the promises it’s been making for the past 50 years.

(Self-replicating cells created from synthetic DNA by scientist Craig Venter)

(Self-replicating cells created from synthetic DNA by scientist Craig Venter)

But Andy balanced this by recounting his own experiences meeting some of the leaders of the most advanced current AI companies, such as Deepmind (a UK startup recently acquried by Google), or this article by Dr. Gary Marcus, Professor of Psychology and Neuroscience at New York University and CEO of Geometric Intelligence.

In reality, these companies are succeeding by avoiding some of the really hard challenges of reproducing human capabilities such as common sense, free will and value-based judgement. They are concentrating instead on making better sense of the physical environment, on processing information in human language, and on creating algorithms that “learn” through feeback loops and self-adjustment.

I think Andy and these experts are right: artificial intelligence has made great strides, but it is not artificial life, and it is a long, long way from creating life-like characteristics such as experience, values and judgements.

If we ever do create artificial life with those characteristics, then I think we will encounter the dangers that Hawkings, Musk and Gates have identified: artificial life will have its own values and act on its own judgement, and any regard for our interests will come second to its own.

That’s a path I don’t think we should go down, and I’m thankful that we’re such a long way from being able to pursue it in anger. I hope that we never do – though I’m also concerned that in Craig Venter and Steve Grand’s work, as well as in robots such as BINA48, we already are already taking the first steps.

But I think in the meantime, there’s tremendous opportunity for digital technology and traditional artificial intelligence to complement human qualities. These technologies are not artificial life and will not overthrow or replace humanity. Hawkings, Gates and Musk are wrong about that.

The human value of the Experience Economy

The final debate at the MIT conference returned to the topic that started the debate over dinner the night before with McAfee and Brynjolfsson: what happens to mass employment in a world where digital technology is automating not just physical work but work involving intelligence and decision-making; and how do we educate today’s children to be successful in a decade’s time in an economy that’s been transformed in ways that we can’t predict?

Andy said we should answer that question by understanding “where will the economic value of humans be?”

I think the answer to that question lies in the experiences that we value emotionally – the experiences digital technology can’t have and can’t understand or replicate;  and in the profound differences between the way that humans think and that machines process information.

It’s nearly 20 years since a computer, IBM’s Deep Blue, first beat the human world champion at Chess, Grandmaster Gary Kasparov. But despite the astonishing subsequent progress in computer power, the world’s best chess player is no longer a computer: it is a team of computers and people playing together. And the world’s best team has neither the world’s best computer chess programme nor the world’s best human chess player amongst its members: instead, it has the best technique for breaking down and distributing the thinking involved in playing chess between its human and computer members, recognising that each has different strengths and qualities.

But we’re not all chess experts. How will the rest of us earn a living in the future?

I had the pleasure last year at TEDxBrum of meeting Nicholas Lovell, author of “The Curve“, a wonderful book exploring the effect that digital technology is having on products and services. Nicholas asks – and answers – a question that McAfee and Brynjolfsson also ask: what happens when digital technology makes the act of producing and distributing some products – such as music, art and films – effectively free?

Nicholas’ answer is that we stop valuing the product and start valuing our experience of the product. This is why some musical artists give away digital copies of their albums for free, whilst charging £30 for a leather-bound CD with photographs of stage performances – and whilst charging £10,000 to visit individual fans in their homes to give personal performances for those fans’ families and friends.

We have always valued the quality of such experiences – this is one reason why despite over a century of advances in film, television and streaming video technology, audiences still flock to theatres to experience the direct performance of plays by actors. We can see similar technology-enabled trends in sectors such as food and catering – Kitchen Surfing, for example, is a business that uses a social media platform to enable anyone to book a professional chef to cook a meal in their home.

The “Experience Economy” is a tremendously powerful idea. It combines something that technology cannot do on its own – create experiences based on human value – with many things that almost all people can do: cook, create art, rent a room, drive a car, make clothes or furniture. Especially when these activities are undertaken socially, they create employment, fulfillment and social capital. And most excitingly, technologies such as Cloud Computing, Open Source Software, social media, and online “Sharing Economy” marketplaces such as Etsy make it possible for anyone to begin earning a living from them with a minimum of expense.

I think that the idea of an “Experience Economy” that is driven by the value of inter-personal and social interactions between people, enabled by “Sharing Economy” business models and technology platforms that enable people with a potentially mutual interest to make contact with each other, is an exciting and very human vision of the future.

Even further: because we are physical beings, we tend to value these interactions more when they occur face-to-face, or when they happen in a place for which we share a mutual affiliation. That creates an incentive to use technology to identify opportunities to interact with people with whom we can meet by walking or cycling, rather than requiring long-distance journeys. And that incentive could be an important component of a long-term sustainable economy.

The future our children will choose

(Today's 5 year-olds are the world's first generation who grew up teaching themselves to use digital information from anywhere in the world before their parents taught them to read and write)

(Today’s 5 year-olds are the world’s first generation who grew up teaching themselves to use digital information from anywhere in the world before their parents taught them to read and write)

I’m convinced that the current generation of Artifical Intelligence based on digital technologies – even those that mimic some structures and behaviours of biological systems, such as Steve Grand’s robot Lucy, BINA48 and IBM’s “brain-inspired” True North chip – will not re-create anything we would recognise as conscious life and free will; or anything remotely capable of understanding human values or making judgements that can be relied on to be consistent with them.

But I am also an atheist and a scientist; and I do not believe there is any mystical explanation for our own consciousness and free will. Ultimately, I’m sure that a combination of science, philosophy and human insight will reveal their origin; and sooner or later we’ll develop a technology – that I do not expect to be purely digital in nature – capable of replicating them.

What might we choose to do with such capabilities?

These capabilities will almost certainly emerge alongside the ability to significantly change our physical minds and bodies – to improve brain performance, muscle performance, select the characteristics of our children and significantly alter our physical appearance. That’s why some people are excited by the science fiction-like possibility of harnessing these capabilities to create an “improved” post-human species – perhaps even transferring our personalities from our own bodies into new, technological machines. These are possibilities that I personally find to be at the very least distasteful; and at worst to be inhuman and frightening.

All of these things are partially possible today, and frankly the limit to which they can be explored is mostly a function of the cost and capability of the available techniques, rather than being set by any legislation or mediated by any ethical debate. To echo another theme of discussions at last week’s MIT conference, science and technology today are developing at a pace that far outstrips the ability of governments, businesses, institutions and most individual people to adapt to them.

I have reasonably clear personal views on these issues. I think our lives are best lived relatively naturally, and that they will be collectively better if we avoid using technology to create artificial “improvements” to our species.

But quite apart from the fact that there are any number of enormous practical, ethical and intellectual challenges to my relatively simple beliefs, the raw truth is that it won’t be my decision whether or how far we pursue these possibilities, nor that of anyone else of my generation (and for the record, I am in my mid-forties).

Much has been written about “digital natives” – those people born in the 1990s who are the first generation who grew up with the Internet and social media as part of their everyday world. The way that that generation socialises, works and thinks about value is already creating enormous changes in our world.

But they are nothing compared to the generation represented by today’s very young children who have grown up using touchscreens and streaming videos, technologies so intuitive and captivating that 2-year-olds now routinely teach themselves how to immerse themselves in them long before parents or school teachers teach them how to read and write.

("Not available on the App Store": a campaign to remind us of the joy of play in the real world)

(“Not available on the App Store“: a campaign to remind us of the joy of play in the real world)

When I was a teenager in the UK, grown-ups wore suits and had traditional haircuts; grown-up men had no earrings. A common parental challenge was to deal with the desire of teenage daughters to have their ears pierced. Those attitudes are terribly old-fashioned today, and our cultural norms have changed dramatically.

I may be completely wrong; but I fully expect our current attitudes to biological and technological manipulation or augmentation of our minds and bodies to thoroughly change over the next few decades; and I have no idea what they will ultimately become. What I do know is that it is likely that my six-year old son’s generation will have far more influence over their ultimate form than my generation will; and that he will grow up with a fundamentally different expectation of the world and his relationship with technology than I have.

I’ve spent my life being excited about technology and the possibilities it creates; ironically I now find myself at least as terrified as I am excited about the world technology will create for my son. I don’t think that my thinking is the result of a mistaken focus on technology over human values – like it or not, our species is differentiated from all others on this planet by our ability to use tools; by our technology. We will not stop developing it.

Our continuing challenge will be to keep a focus on our human values as we do so. I cannot tell my son what to do indefinitely; I can only try to help him to experience and treasure socialising and play in the real world; the experience of growing and preparing food together ; the joy of building things for other people with his own hands. And I hope that those experiences will create human values that will guide him and his generation on a healthy course through a future that I can only begin to imagine.

Advertisements

11 reasons computers can’t understand or solve our problems without human judgement

(Photo by Matt Gidley)

(Photo by Matt Gidley)

Why data is uncertain, cities are not programmable, and the world is not “algorithmic”.

Many people are not convinced that the Smart Cities movement will result in the use of technology to make places, communities and businesses in cities better. Outside their consumer enjoyment of smartphones, social media and online entertainment – to the degree that they have access to them – they don’t believe that technology or the companies that sell it will improve their lives.

The technology industry itself contributes significantly to this lack of trust. Too often we overstate the benefits of technology, or play down its limitations and the challenges involved in using it well.

Most recently, the idea that traditional processes of government should be replaced by “algorithmic regulation” – the comparison of the outcomes of public systems to desired objectives through the measurement of data, and the automatic adjustment of those systems by algorithms in order to achieve them – has been proposed by Tim O’Reilly and other prominent technologists.

These approaches work in many mechanical and engineering systems – the autopilots that fly planes or the anti-lock braking systems that we rely on to stop our cars. But should we extend them into human realms – how we educate our children or how we rehabilitate convicted criminals?

It’s clearly important to ask whether it would be desirable for our society to adopt such approaches. That is a complex debate, but my personal view is that in most cases the incredible technologies available to us today – and which I write about frequently on this blog – should not be used to take automatic decisions about such issues. They are usually more valuable when they are used to improve the information and insight available to human decision-makers – whether they are politicians, public workers or individual citizens – who are then in a better position to exercise good judgement.

More fundamentally, though, I want to challenge whether “algorithmic regulation” or any other highly deterministic approach to human issues is even possible. Quite simply, it is not.

It is true that our ability to collect, analyse and interpret data about the world has advanced to an astonishing degree in recent years. However, that ability is far from perfect, and strongly established scientific and philosophical principles tell us that it is impossible to definitively measure human outcomes from underlying data in physical or computing systems; and that it is impossible to create algorithmic rules that exactly predict them.

Sometimes automated systems succeed despite these limitations – anti-lock braking technology has become nearly ubiquitous because it is more effective than most human drivers at slowing down cars in a controlled way. But in other cases they create such great uncertainties that we must build in safeguards to account for the very real possibility that insights drawn from data are wrong. I do this every time I leave my home with a small umbrella packed in my bag despite the fact that weather forecasts created using enormous amounts of computing power predict a sunny day.

(No matter how sophisticated computer models of cities become, there are fundamental reasons why they will always be simplifications of reality. It is only by understanding those constraints that we can understand which insights from computer models are valuable, and which may be misleading. Image of Sim City by haljackey)

We can only understand where an “algorithmic” approach can be trusted; where it needs safeguards; and where it is wholly inadequate by understanding these limitations. Some of them are practical, and limited only by the sensitivity of today’s sensors and the power of today’s computers. But others are fundamental laws of physics and limitations of logical systems.

When technology companies assert that Smart Cities can create “autonomous, intelligently functioning IT systems that will have perfect knowledge of users’ habits” (as London School of Economics Professor Adam Greenfield rightly criticised in his book “Against the Smart City”), they are ignoring these challenges.

A blog published by the highly influential magazine Wired recently made similar overstatements: “The Universe is Programmable” argues that we should extend the concept of an “Application Programming Interface (API)” – a facility usually offered by technology systems to allow external computer programmes to control or interact with them – to every aspect of the world, including our own biology.

To compare complex, unpredictable, emergent biological and social systems to the very logical, deterministic world of computer software is at best a dramatic oversimplification. The systems that comprise the human body range from the armies of symbiotic microbes that help us digest food in our stomachs to the consequences of using corn syrup to sweeten food to the cultural pressure associated with “size 0” celebrities. Many of those systems can’t be well modelled in their own right, let alone deterministically related to each other; let alone formally represented in an accurate, detailed way by technology systems (or even in mathematics).

We should regret and avoid the hubris that leads to the distrust of technology by overstating its capability and failing to recognise its challenges and limitations. That distrust is a barrier that prevents us from achieving the very real benefits that data and technology can bring, and that have been convincingly demonstrated in the past.

For example, an enormous contribution to our knowledge of how to treat and prevent disease was made by John Snow who used data to analyse outbreaks of cholera in London in the 19th century. Snow used a map to correlate cases of cholera to the location of communal water pipes, leading to the insight that water-borne germs were responsible for spreading the disease. We wash our hands to prevent diseases spreading through germs in part because of what we would now call the “geospatial data analysis” performed by John Snow.

Many of the insights that we seek from analytic and smart city systems are human in nature, not physical or mathematical – for example identifying when and where to apply social care interventions in order to reduce the occurrence of  emotional domestic abuse. Such questions are complex and uncertain: what is “emotional domestic abuse?” Is it abuse inflicted by a live-in boyfriend, or by an estranged husband who lives separately but makes threatening telephone calls? Does it consist of physical violence or bullying? And what is “bullying”?

IMG_0209-1

(John Snow’s map of cholera outbreaks in 19th century London)

We attempt to create structured, quantitative data about complex human and social issues by using approximations and categorisations; by tolerating ranges and uncertainties in numeric measurements; by making subjective judgements; and by looking for patterns and clusters across different categories of data. Whilst these techniques can be very powerful, just how difficult it is to be sure what these conventions and interpretations should be is illustrated by the controversies that regularly arise around “who knew what, when?” whenever there is a high profile failure in social care or any other public service.

These challenges are not limited to “high level” social, economic and biological systems. In fact, they extend throughout the worlds of physics and chemistry into the basic nature of matter and the universe. They fundamentally limit the degree to which we can measure the world, and our ability to draw insight from that information.

By being aware of these limitations we are able to design systems and practises to use data and technology effectively. We know more about the weather through modelling it using scientific and mathematical algorithms in computers than we would without those techniques; but we don’t expect those forecasts to be entirely accurate. Similarly, supermarkets can use data about past purchases to make sufficiently accurate predictions about future spending patterns to boost their profits, without needing to predict exactly what each individual customer will buy.

We underestimate the limitations and flaws of these approaches at our peril. Whilst Tim O’Reilly cites several automated financial systems as good examples of “algorithmic regulation”, the financial crash of 2008 showed the terrible consequences of the thoroughly inadequate risk management systems used by the world’s financial institutions compared to the complexity of the system that they sought to profit from. The few institutions that realised that market conditions had changed and that their models for risk management were no longer valid relied instead on the expertise of their staff, and avoided the worst affects. Others continued to rely on models that had started to produce increasingly misleading guidance, leading to the recession that we are only now emerging from six years later, and that has damaged countless lives around the world.

Every day in their work, scientists, engineers and statisticians draw conclusions from data and analytics, but they temper those conclusions with an awareness of their limitations and any uncertainties inherent in them. By taking and communicating such a balanced and informed approach to applying similar techniques in cities, we will create more trust in these technologies than by overstating their capabilities.

What follows is a description of some of the scientific, philosophical and practical issues that lead inevitability to uncertainty in data, and to limitations in our ability to draw conclusions from it:

But I’ll finish with an explanation of why we can still draw great value from data and analytics if we are aware of those issues and take them properly into account.

Three reasons why we can’t measure data perfectly

(How Heisenberg’s Uncertainty Principle results from the dual wave/particle nature of matter. Explanation by HyperPhysics at Georgia State University)

1. Heisenberg’s Uncertainty Principle and the fundamental impossibility of knowing everything about anything

Heisenberg’s Uncertainty Principle is a cornerstone of Quantum Mechanics, which, along with General Relativity, is one of the two most fundamental theories scientists use to understand our world. It defines a limit to the precision with which certain pairs of properties of the basic particles which make up the world – such as protons, neutrons and electrons – can be known at the same time. For instance, the more accurately we measure the position of such particles, the more uncertain their speed and direction of movement become.

The explanation of the Uncertainty Principle is subtle, and lies in the strange fact that very small “particles” such as electrons and neutrons also behave like “waves”; and that “waves” like beams of light also behave like very small “particles” called “photons“. But we can use an analogy to understand it.

In order to measure something, we have to interact with it. In everyday life, we do this by using our eyes to measure lightwaves that are created by lightbulbs or the sun and that then reflect off objects in the world around us.

But when we shine light on an object, what we are actually doing is showering it with billions of photons, and observing the way that they scatter. When the object is quite large – a car, a person, or a football – the photons are so small in comparison that they bounce off without affecting it. But when the object is very small – such as an atom – the photons colliding with it are large enough to knock it out of its original position. In other words, measuring the current position of an object involves a collision which causes it to move in a random way.

This analogy isn’t exact; but it conveys the general idea. (For a full explanation, see the figure and link above). Most of the time, we don’t notice the effects of Heisenberg’s Uncertainty Principle because it applies at extremely small scales. But it is perhaps the most fundamental law that asserts that “perfect knowledge” is simply impossible; and it illustrates a wider point that any form of measurement or observation in general affects what is measured or observed. Sometimes the effects are negligible,  but often they are not – if we observe workers in a time and motion study, for example, we need to be careful to understand the effect our presence and observations have on their behaviour.

2. Accuracy, precision, noise, uncertainty and error: why measurements are never fully reliable

Outside the world of Quantum Mechanics, there are more practical issues that limit the accuracy of all measurements and data.

(A measurement of the electrical properties of a superconducting device from my PhD thesis. Theoretically, the behaviour should appear as a smooth, wavy line; but the experimental measurement is affected by noise and interference that cause the signal to become "fuzzy". In this case, the effects of noise and interference - the degree to which the signal appears "fuzzy" - are relatively small relative to the strength of the signal, and the device is usable)

(A measurement of the electrical properties of a superconducting device from my PhD thesis. Theoretically, the behaviour should appear as a smooth, wavy line; but the experimental measurement is affected by noise and interference that cause the signal to become “fuzzy”. In this case, the effects of noise and interference – the degree to which the signal appears “fuzzy” – are relatively small compared to the strength of the signal, and the device is usable)

We live in a “warm” world – roughly 300 degrees Celsius above what scientists call “absolute zero“, the coldest temperature possible. What we experience as warmth is in fact movement: the atoms from which we and our world are made “jiggle about” – they move randomly. When we touch a hot object and feel pain it is because this movement is too violent to bear – it’s like being pricked by billions of tiny pins.

This random movement creates “noise” in every physical system, like the static we hear in analogue radio stations or on poor quality telephone connections.

We also live in a busy world, and this activity leads to other sources of noise. All electronic equipment creates electrical and magnetic fields that spread beyond the equipment itself, and in turn affect other equipment – we can hear this as a buzzing noise when we leave smartphones near radios.

Generally speaking, all measurements are affected by random noise created by heat, vibrations or electrical interference; are limited by the precision and accuracy of the measuring devices we use; and are affected by inconsistencies and errors that arise because it is always impossible to completely separate the measurement we want to make from all other environmental factors.

Scientists, engineers and statisticians are familiar with these challenges, and use techniques developed over the course of more than a century to determine and describe the degree to which they can trust and rely on the measurements they make. They do not claim “perfect knowledge” of anything; on the contrary, they are diligent in describing the unavoidable uncertainty that is inherent in their work.

3. The limitations of measuring the natural world using digital systems

One of the techniques we’ve adopted over the last half century to overcome the effects of noise and to make information easier to process is to convert “analogue” information about the real world (information that varies smoothly) into digital information – i.e. information that is expressed as sequences of zeros and ones in computer systems.

(When analogue signals are amplified, so is the noise that they contain. Digital signals are interpreted using thresholds: above an upper threshold, the signal means “1”, whilst below a lower threshold, the signal means “0”. A long string of “0”s and “1”s can be used to encode the same information as contained in analogue waves. By making the difference between the thresholds large compared to the level of signal noise, digital signals can be recreated to remove noise. Further explanation and image by Science Aid)

This process involves a trade-off between the accuracy with which analogue information is measured and described, and the length of the string of digits required to do so – and hence the amount of computer storage and processing power needed.

This trade-off can be clearly seen in the difference in quality between an internet video viewed on a smartphone over a 3G connection and one viewed on a high definition television using a cable network. Neither video will be affected by the static noise that affects weak analogue television signals, but the limited bandwidth of a 3G connection dramatically limits the clarity and resolution of the image transmitted.

The Nyquist–Shannon sampling theorem defines this trade-off and the limit to the quality that can be achieved in storing and processing digital information created from analogue sources. It determines the quality of digital data that we are able to create about any real-world system – from weather patterns to the location of moving objects to the fidelity of sound and video recordings. As computers and communications networks continue to grow more powerful, the quality of digital information will improve,  but it will never be a perfect representation of the real world.

Three limits to our ability to analyse data and draw insights from it

1. Gödel’s Incompleteness Theorem and the inconsistency of algorithms

Kurt Gödel’s Incompleteness Theorem sets a limit on what can be achieved by any “closed logical system”. Examples of “closed logical systems” include computer programming languages, any system for creating algorithms – and mathematics itself.

We use “closed logical systems” whenever we create insights and conclusions by combining and extrapolating from basic data and facts. This is how all reporting, calculating, business intelligence, “analytics” and “big data” technologies work.

Gödel’s Incompleteness Theorem proves that any closed logical system can be used to create conclusions that  it is not possible to show are true or false using the same system. In other words, whilst computer systems can produce extremely useful information, we cannot rely on them to prove that that information is completely accurate and valid. We have to do that ourselves.

Gödel’s theorem doesn’t stop computer algorithms that have been verified by humans using the scientific method from working; but it does mean that we can’t rely on computers to both generate algorithms and guarantee their validity.

2. The behaviour of many real-world systems can’t be reduced analytically to simple rules

Many systems in the real-world are complex: they cannot be described by simple rules that predict their behaviour based on measurements of their initial conditions.

A simple example is the “three body problem“. Imagine a sun, a planet and a moon all orbiting each other. The movement of these three objects is governed by the force of gravity, which can be described by relatively simple mathematical equations. However, even with just three objects involved, it is not possible to use these equations to directly predict their long-term behaviour – whether they will continue to orbit each other indefinitely, or will eventually collide with each other, or spin off into the distance.

(A computer simulation by Hawk Express of a Belousov–Zhabotinsky reaction,  in which reactions between liquid chemicals create oscillating patterns of colour. The simulation is carried out using “cellular automata” a technique based on a grid of squares which can take different colours. In each “turn” of the simulation, like a turn in a board game, the colour of each square is changed using simple rules based on the colours of adjacent squares. Such simulations have been used to reproduce a variety of real-world phenomena)

As Stephen Wolfram argued in his controversial book “A New Kind of Science” in 2002, we need to take a different approach to understanding such complex systems. Rather than using mathematics and logic to analyse them, we need to simulate them, often using computers to create models of the elements from which complex systems are composed, and the interactions between them. By running simulations based on a large number of starting points and comparing the results to real-world observations, insights into the behaviour of the real-world system can be derived. This is how weather forecasts are created, for example. 

But as we all know, weather forecasts are not always accurate. Simulations are approximations to real-world systems, and their accuracy is restricted by the degree to which digital data can be used to represent a non-digital world. For this reason, conclusions and predictions drawn from simulations are usually “average” or “probable” outcomes for the system as a whole, not precise predictions of the behaviour of the system or any individual element of it. This is why weather forecasts are often wrong; and why they predict likely levels of rain and windspeed rather than the shape and movement of individual clouds.

(Hello)

(A simple and famous example of a computer programme that never stops running because it calls itself. The output continually varies by printing out characters based on random number generation. Image by Prosthetic Knowledge)

3. Some problems can’t be solved by computing machines

If I consider a simple question such as “how many letters are in the word ‘calculation’?”, I can easily convince myself that a computer programme could be written to answer the question; and that it would find the answer within a relatively short amount of time. But some problems are much harder to solve, or can’t even be solved at all.

For example, a “Wang Tile” (see image below) is a square tile formed from four triangles of different colours. Imagine that you have bought a set of tiles of various colour combinations in order to tile a wall in a kitchen or bathroom. Given the set of tiles that you have bought, is it possible to tile your wall so that triangles of the same colour line up to each other, forming a pattern of “Wang Tile” squares?

In 1966 Robert Berger proved that no algorithm exists that can answer that question. There is no way to solve the problem – or to determine how long it will take to solve the problem – without actually solving it. You just have to try to tile the room and find out the hard way.

One of the most famous examples of this type of problem is the “halting problem” in computer science. Some computer programmes finish executing their commands relatively quickly. Others can run indefinitely if they contain a “loop” instruction that never ends. For others which contain complex sequences of loops and calls from one section of code to another, it may be very hard to tell whether the programme finishes quickly, or takes a long time to complete, or never finishes its execution at all.

Alan Turing, one of the most important figures in the development of computing, proved in 1936 that a general algorithm to determine whether or not any computer programme finishes its execution does not exist. In other words, whilst there are many useful computer programmes in the world, there are also problems that computer programmes simply cannot solve.

(A set of Wang Tiles, and a pattern created by tiling them so that tiles are placed next to other tiles so that their edges have the same colour. Given any particular set of tiles, it is impossible to determine whether such a pattern can be created by any means other than trial and error)

(A set of Wang Tiles, and a pattern of coloured squares created by tiling them. Given any random set of tiles of different colour combinations, there is no set of rules that can be relied on to determine whether a valid pattern of coloured squares can be created from them. Sometimes, you have to find out by trial and error. Images from Wikipedia)

Five reasons why the human world is messy, unpredictable, and can’t be perfectly described using data and logic

1. Our actions create disorder

The 2nd Law of Thermodynamics is a good candidate for the most fundamental law of science. It states that as time progresses, the universe becomes more disorganised. It guarantees that ultimately – in billions of years – the Universe will die as all of the energy and activity within it dissipates.

An everyday practical consequence of this law is that every time we act to create value – building a shed, using a car to get from one place to another, cooking a meal – our actions eventually cause a greater amount of disorder to be created as a consequence – as noise, pollution, waste heat or landfill refuse.

For example, if I spend a day building a shed, then to create that order and value from raw materials, I consume structured food and turn it into sewage. Or if I use an electric forklift to stack a pile of boxes, I use electricity that has been created by burning structured coal into smog and ash.

So it is literally impossible to create a “perfect world”. Whenever we act to make a part of the world more ordered, we create disorder elsewhere. And ultimately – thankfully, long after you and I are dead – disorder is all that will be left.

2. The failure of Logical Atomism: why the human world can’t be perfectly described using data and logic

In the 20th Century two of the most famous and accomplished philosophers in history, Bertrand Russell and Ludwig Wittgenstein, invented “Logical Atomism“, a theory that the entire world could be described by using “atomic facts” – independent and irreducible pieces of knowledge – combined with logic.

But despite 40 years of work, these two supremely intelligent people could not get their theory to work: “Logical Atomism” failed. It is not possible to describe our world in that way.

One cause of the failure was the insurmountable difficulty of identifying truly independent, irreducible atomic facts. “The box is red” and “the circle is blue”, for example, aren’t independent or irreducible facts for many reasons. “Red” and “blue” are two conventions of human language used to describe the perceptions created when electro-magnetic waves of different frequencies arrive at our retinas. In other words, they depend on and relate to each other through a number of sophisticated systems.

Despite centuries of scientific and philosophical effort, we do not have a complete understanding of how to describe our world at its most basic level. As physicists have explored the world at smaller and smaller scales, Quantum Mechanics has emerged as the most fundamental theory for describing it – it is the closest we have come to finding the “irreducible facts” that Russell and Wittgenstein were looking for. But whilst the mathematical equations of Quantum Mechanics predict the outcomes of experiments very well, after nearly a century, physicists still don’t really agree about what those equations mean. And as we have already seen, Heisenberg’s Uncertainty Principle prevents us from ever having perfect knowledge of the world at this level.

Perhaps the most important failure of logical atomism, though, was that it proved impossible to use logical rules to turn “facts” at one level of abstraction – for example, “blood cells carry oxygen”, “nerves conduct electricity”, “muscle fibres contract” – into facts at another level of abstraction – such as “physical assault is a crime”. The human world and the things that we care about can’t be described using logical combinations of “atomic facts”. For example, how would you define the set of all possible uses of a screwdriver, from prising the lids off paint tins to causing a short-circuit by jamming it into a switchboard?

Our world is messy, subjective and opportunistic. It defies universal categorisation and logical analysis.

(A Pescheria in Bari, Puglia, where a fish-market price information service makes it easier for local fisherman to identify the best buyers and prices for their daily catch. Photo by Vito Palmi)

3. The importance and inaccessibility of “local knowledge” 

Because the tool we use for calculating and agreeing value when we exchange goods and services is money, economics is the discipline that is often used to understand the large-scale behaviour of society. We often quantify the “growth” of society using economic measures, for example.

But this approach is notorious for overlooking social and environmental characteristics such as health, happiness and sustainability. Alternatives exist, such as the Social Progress Index, or the measurement framework adopted by the United Nations 2014 Human Development Report on world poverty; but they are still high level and abstract.

Such approaches struggle to explain localised variations, and in particular cannot predict the behaviours or outcomes of individual people with any accuracy. This “local knowledge problem” is caused by the fact that a great deal of the information that determines individual actions is personal and local, and not measurable at a distance – the experienced eye of the fruit buyer assessing not just the quality of the fruit but the quality of the farm and farmers that produce it, as a measure of the likely consistency of supply; the emotional attachments that cause us to favour one brand over another; or the degree of community ties between local businesses that influence their propensity to trade with each other.

Sharing economy” business models that use social media and reputation systems to enable suppliers and consumers of goods and services to find each other and transact online are opening up this local knowledge to some degree. Local food networks, freecycling networks, and land-sharing schemes all use this technology to the benefit of local communities whilst potentially making information about detailed transactions more widely available. And to some degree, the human knowledge that influences how transactions take place can be encoded in “expert systems” which allow computer systems to codify the quantitative and heuristic rules by which people take decisions.

But these technologies are only used in a subset of the interactions that take place between people and businesses across the world, and it is unlikely that they’ll become ubiquitous in the foreseeable future (or that we would want them to become so). Will we ever reach the point where prospective house-buyers delegate decisions about where to live to computer programmes operating in online marketplaces rather than by visiting places and imagining themselves living there? Will we somehow automate the process of testing the freshness of fish by observing the clarity of their eyes and the freshness of their smell before buying them to cook and eat?

In many cases, while technology may play a role introducing potential buyers and sellers of goods and services to each other, it will not replace – or predict – the human behaviours involved in the transaction itself.

(Medway Youth Trust use predictive and textual analytics to draw insight into their work helping vulnerable children. They use technology to inform expert case workers, not to take decisions on their behalf.)

4. “Wicked problems” cannot be described using data and logic

Despite all of the challenges associated with problems in mathematics and the physical sciences, it is nevertheless relatively straightforward to frame and then attempt to solve problems in those domains; and to determine whether the resulting solutions are valid.

As the failure of Logical Atomism showed, though, problems in the human domain are much more difficult to describe in any systematic, complete and precise way – a challenge known as the “frame problem” in artificial intelligence. This is particularly true of “wicked problems” – challenges such as social mobility or vulnerable families that are multi-faceted, and consist of a variety of interdependent issues.

Take job creation, for example. Is that best accomplished through creating employment in taxpayer-funded public sector organisations? Or by allowing private-sector wealth to grow, creating employment through “trickle-down” effects? Or by maximising overall consumer spending power as suggested by “middle-out” economics? All of these ideas are described not using the language of mathematics or other formal logical systems, but using natural human language which is subjective and inconsistent in use.

The failure of Logical Atomism to fully represent such concepts in formal logical systems through which truth and falsehood can be determined with certainty emphasises what we all understand intuitively: there is no single “right” answer to many human problems, and no single “right” action in many human situations.

(An electricity bill containing information provided by OPower comparing one household’s energy usage to their neighbours. Image from Grist)

5. Behavioural economics and the caprice of human behaviour

Behavioural economics” attempts to predict the way that humans behave when taking choices that have a measurable impact on them – for example, whether to put the washing machine on at 5pm when electricity is expensive, or at 11pm when it is cheap.

But predicting human behaviour is notoriously unreliable.

For example, in a smart water-meter project in Dubuque, Iowa, households that were told how their water conservation compared to that of their near neighbours were found to be twice as likely to take action to improve their efficiency as those who were only told the details of their own water use. In other words, people who were given quantified evidence that they were less responsible water user than their neighbours changed their behaviour. OPower have used similar techniques to help US households save 1.9 terawatt hours of power simply by including a report based on data from smart meters in a printed letter sent with customers’ electricity bills.

These are impressive achievements; but they are not always repeatable. A recycling scheme in the UK that adopted a similar approach found instead that it lowered recycling rates across the community: households who learned that they were putting more effort into recycling than their neighbours asked themselves “if my neighbours aren’t contributing to this initiative, then why should I?”

Low carbon engineering technologies like electric vehicles have clearly defined environmental benefits and clearly defined costs. But most Smart Cities solutions are less straightforward. They are complex socio-technical systems whose outcomes are emergent. Our ability to predict their performance and impact will certainly improve as more are deployed and analysed, and as University researchers, politicians, journalists and the public assess them. But we will never predict individual actions using these techniques, only the average statistical behaviour of groups of people. This can be seen from OPower’s own comparison of their predicted energy savings against those actually achieved – the predictions are good, but the actual behaviour of OPower’s customers shows a high degree of apparently random variation. Those variations are the result of the subjective, unpredictable and sometimes irrational behaviour of real people.

We can take insight from Behavioural Economics and other techniques for analysing human behaviour in order to create appropriate strategies, policies and environments that encourage the right outcomes in cities; but none of them can be relied on to give definitive solutions to any individual person or situation. They can inform decision-making, but are always associated with some degree of uncertainty. In some cases, the uncertainty will be so small as to be negligible, and the predictions can be treated as deterministic rules for achieving the desired outcome. But in many cases, the uncertainty will be so great that predictions can only be treated as general indications of what might happen; whilst individual actions and outcomes will vary greatly.

(Of course it is impossible to predict individual criminal actions as portrayed in the film “Minority Report”. But is is very possible to analyse past patterns of criminal activity, compare them to related data such as weather and social events, and predict the likelihood of crimes of certain types occurring in certain areas. Cities such as Memphis and Chicago have used these insights to achieve significant reductions in crime)

Learning to value insight without certainty

Mathematics and digital technology are incredibly powerful; but they will never perfectly and completely describe and predict our world in human terms. In many cases, our focus for using them should not be on automation: it should be on the enablement of human judgement through better availability and communication of information. And in particular, we should concentrate on communicating accurately the meaning of information in the context of its limitations and uncertainties.

There are exceptions where we automate systems because of a combination of a low-level of uncertainty in data and a large advantage in acting autonomously on it. For example, anti-lock braking systems save lives by using automated technology to take thousands of decisions more quickly than most humans would realise that even a single decision needed to be made; and do so based on data with an extremely low degree of uncertainty.

But the most exciting opportunity for us all is to learn to become sophisticated users of information that is uncertain. The results of textual analysis of sentiment towards products and brands expressed in social media are far from certain; but they are still of great value. Similar technology can extract insights from medical research papers, case notes in social care systems, maintenance logs of machinery and many other sources. Those insights will rarely be certain; but properly assessed by people with good judgement they can still be immensely valuable.

This is a much better way to understand the value of technology than ideas like “perfect knowledge” and “algorithmic regulation”. And it is much more likely that people will trust the benefits that we claim new technologies can bring if we are open about their limitations. People won’t use technologies that they don’t trust; and they won’t invest their money in them or vote for politicians who say they’ll spend their taxes on it.

Thankyou to Richard Brown and Adrian McEwen for discussions on Twitter that helped me to prepare this article. A more in-depth discussion of some of the scientific and philosophical issues I’ve described, and an exploration of the nature of human intelligence and its non-deterministic characteristics, can be found in the excellent paper “Answering Descartes: Beyond Turing” by Stuart Kauffman published by MIT press.

%d bloggers like this: