Hello World: How to Be Human in the Age of the Machine. By H. FRY (London, UK: Doubleday, 2018). pp. 244 + xii, hardback (ISBN 978-0-8575-2524-6), £18.99.
“In the age of the algorithm, humans have never been more important.” With these words, Hannah Fry concludes her exploration of the state of the art of AI (but see below) as it affects everyday life. Mathematician Fry will be best known to people in the UK, where she’s built a reputation as an excellent science communicator. She presented the documentary The Joy of Data for BBC4, and co-presents The Curious Cases of Rutherford & Fry on BBC Radio 4. She also recently started a regular slot, The Maths of Life, on Lauren Laverne’s breakfast show on BBC Radio 6 Music.
There’s Artificial Intelligence, Machine Learning and Computational Statistics
Hello World focuses mostly on Machine Learning (ML) rather than AI per se, although the mainstream media (and some tech companies, who really should know better) often equate the two. And, as Fry notes, when people say “Machine Learning” what they are really talking about is computational statistics. Don’t let this put you off, however. You don’t need to be an expert in maths or statistics to understand what’s going on here. Where a little extra knowledge is needed to grasp a particular concept, Fry provides it in a concise, easy-to-understand way. So those who aren’t familiar with (as well as those who can’t remember!) Bayes’ Theorem, for example, will come away with a basic idea of how it works, as well as how and where it’s being used.
No Silver Bullet
The Big Tech companies, amongst others, would have you believe that AI (read ML) is the solution to all our problems. Fry debunks this myth in a clear and highly readable manner that will appeal to people of all abilities. She begins with a high level description of ML, and the algorithms involved, before moving on to critique how it is being (and has been) deployed, using a plethora of examples showing how the algorithms are having an impact on our everyday lives.
For the most part, ML is used in areas where we rely on human decision-making. We know from the work of people like Dan Kahneman that human decision-making is far from perfect, particularly when there’s a lot of data, and not a lot of time to come up with a suitable answer. In these situations we often rely on what Kahneman calls System One thinking (in his excellent book, Thinking, Fast and Slow), which is quick and works fine a lot of the time. It can lead to problems, however, in those cases where we should really have spent more time deliberating over the problem using the slower System Two thinking. One classic, albeit simple, example used in the book is:
A bat and ball together cost £1.10.
The bat costs £1 more than the ball.
How much does the ball cost?
(The answer is given at the foot of this review.)
From Chess to the Arts via Self-Driving Cars
Fry examines the power of ML algorithms, and the ways in which data is being harvested to feed them, some of which she describes as creepy. She illustrates the power starting with the familiar tale of the chess match between Garry Kasparov and Big Blue. While many will know the original outcome, they may be less aware of some of the clever behaviours that Big Blue was programmed to exhibit, and how it changed the way that Kasparov plays chess. The discussion about data will be a bit of an eye-opener for some people when they discover that there are companies — data brokers — whose sole business model is to sell your personal data; it may also make you think twice about signing up with genealogy companies that ask for your DNA.
It would be almost impossible to provide comprehensive coverage of all the areas where ML is being used, but Fry offers a very good cross-section. She shows how ML is already affecting our lives, for better or worse, in areas such as determining the length of custodial sentences in criminal justice; diagnosing cancer; helping cars to drive themselves; aiding the prevention and detection of crime; and even in making works of art, including music. In each of the different areas Fry critically appraises what the algorithms can do, always making sure that we’re aware of the bad as well as the good. While some of the examples will be familiar, having received coverage in the press, others will not. There is enough detail in the main text, however, to satisfy the interest of both casual readers and passing experts. Anyone wanting a more detailed exposition can follow the endnotes to the extensive list of references at the back of the book.
Fry shows that the consequences of using ML are often relatively benign, but sometimes they can be very disturbing, as in the case of Steve Talley. He was severely beaten when he was arrested for carrying out two armed bank robberies. Talley was mistakenly identified, using face recognition (AI) software, and it took him over a year to clear his name, even though he had a cast-iron alibi.
The chapter on cars was of particular interest to me. Although research into self-driving cars has been around since the 1950s, it has really taken off in the last 20 years or so. These cars employ a plethora of sensors, and use Bayes Theorem to try to make sense of the world that the car perceives as it drives along. Current forecasts suggest that there will be autonomous vehicles on the roads from about 2020. They’ll be rather more limited than some people would have you believe, however, and there are still many problems to be solved. These include technological ones, such as how to cope with faulty or failed sensors, and how to break the car’s pre-programmed “rules of the road” so that you can do things like drive on the pavement to let an ambulance pass. But there are others too, such as the ethical problem of choosing between two possible outcomes, when both will lead to fatalities.
Rather than just allowing the cars to take over, Fry suggests that we should perhaps be thinking more in terms of the car working together with the driver. So, in some (limited) situations the car would be in control and act as a sort of chauffeur, while in others it would act more as a monitor to make sure that the driver, who is in control, is paying appropriate attention. Toyota is following this approach, which gives you the best of both worlds.
People + Transparent Technology
Fry picks up on the need for people and technology to work together from Bainbridge’s classic Ironies of Automation. Many of the problems described in that paper are still prevalent some 35 years later. Fry goes on to suggest that we should be exploiting ML as a tool to support human decision-making, rather than replacing it. That way we can exploit the fact that people are better at doing some things, while technology is better at others. It also means that the responsibility for taking actions based on the final decision ultimately rests with the human. This strongly resonates with our approach when designing the FLORENCE (Fuzzy LOgic REspiratory Neonatal Care Expert) system a few years ago. FLORENCE was designed to help junior doctors make decisions when treating premature babies with Respiratory Distress Syndrome. Even though FLORENCE could have automatically implemented its decisions, we decided that they would always be carried out by the doctors. We also let staff use their own discretion in implementing FLORENCE’s suggestions. As long as they gave their reasons for doing so, they were allowed to override FLORENCE.
The other main thing that Fry singles out is the need for algorithms to be transparent, so that we can understand how they arrive at their decisions. It’s just like school, where you always had to show your working in mathematics… at least in my day! We designed FLORENCE to be able to provide its line of reasoning on request. This need for explanation is critical, because when any technology goes wrong—which it will, at some point—it is invariably the user (pilot, driver, whoever) that has to step in. In order to do so, however, they need to know what the system is currently doing.
Hello World is an excellent book. It’s a very easy read, and requires no prior knowledge of ML (or AI, more generally). For novices and casual readers, it provides a very good introduction to ML and its strengths and weaknesses. For those with a background in ML, it provides plenty of pointers for futher investigation in the references. The book won’t teach you how to write an ML algorithm, although it can be used to generate a set of principles (covering data collection, transparency, accuracy thresholds and so on) to help guide anyone who does. The book is very current, using several recent examples to illustrate its points. There is now a paperback edition available too, which includes an extra chapter. I highly recommend Hello World to everyone with an interest in ML, from those wanting to find out a little bit more about AI — ML, in particular — to those who are (or will be) developing ML algorithms.
ANSWER: For those who are still wondering about the bat and ball problem, the ball costs 5p (and the bat costs £1.05). Many people when they first encounter the problem think that the ball costs 10p, but that would mean that the bat costs £1.10, so the cost of both bat and ball would be £1.20, rather than £1.10.