Cutting through the AI hype

A book review of ‘Prediction Machines — the Simple Economics of Artificial Intelligence’ by Ajay Agrawal, Joshua Gans, and Avi Goldfarb


By means of the famously cited ’10 man’ Dartmouth Summer Conference, the research field of Artificial Intelligence (AI) was already established in 1956. The elite workshop was attended by John McCharty (Dartmouth College), Marvin Minsky (Harvard University), Nathaniel Rochester (IBM) and Claude Shannon (Bell Telephone Laboratories), amongst others.

However, the public debate about AI has just begun. After more than half a century of intensive research and development, the technology finally made it into mainstream business: Netflix and Amazon are recommending films and books to you. Due to massive amounts of big data, generated by millions of individual users, it is calculated whether you could potentially like ‘Bird Box’ when you already watched ‘La Casa del Papel’ and ‘Narcos’. The system behind is based on a machine learning (ML) algorithm.

At latest since the presentation of the iPhone by Steve Jobs at the 2007 Mac World Conference in San Francisco, we are living in a world dominated by smartphones — the invention which shows to have the fastest mass adoption in human history. Thereby, we are getting used to taking advantage of many inventions based on AI, even if we are not aware of it. For instance, when you are doing a selfie with your smartphone, an AI-based algorithm calculates the location of your head, which helps the camera to know where to focus.

Today, the public debate concerning AI is heavily skewed by either utopian or rather dystopian views of researchers, technologists, venture capitalists, journalists, and bloggers. Pessimists like Cathy O’Neil argue that in the worst case scenario, erroneously trained machines could become ‘weapons of math destruction’, thereby increasing inequality and threatening democracy. Optimists are framing AI as a never-ending dream to free humans from both manual as well as mental labor. In this sense, Google’s director of engineering Ray Kurzweil is predicting the momentum of ‘Singularity’ to occur around 2045.

This means that there will be ‘conscious machines’, which surpass human-level intelligence. Consequently, this will translate into an age of run-away economic growth, where human work becomes obsolete. Indeed, the Oxford researchers Frey and Osborne conclude that 47 percent of the US jobs are at high risks from automatisation. They write, that within the first wave of future industry adaptation ‘most workers in transportation and logistics occupations, together with the bulk of office and administrative support workers, and labour in production occupations, are likely to be substituted by computer capital.’ Due to such dramatic visions, it is not surprising that AI has the potential to cause widespread public fears.


If you want to read a straightforward guide in order to cut through the hype, ‘Prediction Machines’ by the Toronto based professors Ajay Agrawal, Joshua Gans, and Avi Goldfarb is the book you should read!

Source of the book front side.

The authors provide a soft and sound economic analysis of what AI is currently capable of and what hasn’t been achieved yet. In a sober way of thinking, they state: ‘If economists are good at one thing, it is cutting through hype. Whereas others see transformational new innovation, we see a simple fall in price.’

In their point of view, the advent and commercialization of computers has made arithmetic cheap. In other words, solving a complex mathematical equation was done more easily and in less time than before. Sounds boring and less promising? Hold on.

So, what will AI technologies make cheap?

From the authors’ perspective, it is prediction. Prediction is central to decision-making under uncertainty and thereby elementary for any kind of business. Better prediction provide new opportunities for all kind of companies from insurance to health care and retail.

The authors did a great job in showing a bunch of real-world use cases and examples, which helps to break down your ‘AI moment’. For instance, the deep-learning powered service iFlytek is used by over 500 million Chinese users to translate, transcribe and communicate natural language. The Apple watch application Cardiogram is able to save lives, because it detects an irregular heart rhythm with 97 percent accuracy, using a deep neural network. ZipRecruiter matches job requirements with user profiles of potential applicants and Atomwise is aiming to shorten the time involved in discovering promising pharmaceutical drug prospects. Besides, applications like Uber, Lyft or Waze are forecasting the fastest way of getting from point A to point B by using real-world data. Here, the performance of the app gets better and more valuable the more users it has. Such an opportunity is called ‘network effect’ by economists.

As we can see, former semantic problems of translation and interpretation have been applied to probabilistic calculations. An AI system gets trained by large amount of input data or linguistic features in order to provide an output which figures out the meaning of a sentence. In similar settings, former navigation problems of relying on the knowledge of a cab driver became probabilistic challenges, matching the vast amount of information on the streets in order to generate the optimal route.


So, what about jobs? How will AI influence labor markets?

The most intuitive answer when prediction becomes cheap due to AI systems is that humans will be used less for predictive tasks. The notion of ‘tasks’ is very important and probably the main failure of the Oxford study mentioned before, because Frey and Osborne were trying to forecast the automatisation potential for a ‘job’ category as a whole. However, a ‘job’ consists of hundreds of different tasks. Imagine a secretary, which has to take calls, manage agendas, understand, categorize and archive documents and e-mails, give hints to his or her boss, prepare meeting rooms and welcome clients, amongst others. Not all of these tasks can be automated yet.

Besides, not every client would be happy to be welcomed and served by a robot, which is why many businesses have no interest in automating everything. The authors use the example of a school bus driver. In case, autonomous cars would enter our roads, the school bus driver would probably not lose his or her job. This is because many parents would not like the idea that their kids would drive to school autonomously, without any kind of adult supervision. Hence, the job of the bus driver would rather be redesigned. Instead of driving, his new main tasks would consist in mentoring the kids as well as managing any kind of conflict and uncertain event. This is particularly the case because machines are very bad in predicting rare situations. Overall, human judgement is still needed.

Historically, such a redesign of jobs happened for example when automated teller machines (ATMs) were invented and installed all around the globe. As shown be James Bessen and cited by the authors, this evolution did not result in bank tellers losing their jobs but in more of them becoming employed. Instead of handing in cash to their clients, their new job was to provide consulting in order to advise them on loans and work out credit card options.

Source of Figure 16–1 from the book.

Hence, when prediction becomes cheaper, complementary inputs to decision-making like data as well as human interpretation, judgement and action will become more valuable. To sum it up, instead of full job automation realized throughout AI systems, it is more likely that we can expect many jobs to be redesigned. Some jobs will be replaced, but others will be augmented. Many companies but also public administration will need to rethink their firm-level structures, hierarchies, chains of command and workflows. When you are an entrepreneur, Agrawal, Gans and Goldfarb provide plenty of advice on how to do that. Prediction Machines is a sober economic analysis and comprehensive read full of plausible and helpful examples.

Further readings:

Agrawal, A., Goldfarb, A., & Gans, J. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business School Press Books.

Bessen, J. E. (2016). How Computer Automation Affects Occupations: Technology, Jobs, and Skills. Boston University School of Law. Law and Economics Research Paper.

Buhr, S. (2017). Apple’s Watch can detect an abnormal heart rhythm with 97% accuracy, UCSF study says. Retrieved from: https://techcrunch.com/2017/05/11/apples-watch-can-detect-an-abnormal-heart-rhythm-with-97-accuracy-ucsf-study-says/

Frey, C. B., & Osborne, M. A. (2013). The Future of Employment. Oxford University.

Futurism (2017). Kurzweil Claims That the Singularity Will Happen by 2045. Get ready for humanity 2.0. 5th of Oktober 2017. Retrieved from: https://futurism.com/kurzweil-claims-that-the-singularity-will-happen-by-2045/

O’Neil, C. (2016). Weapons of Math Destruction : How Big Data Increases Inequality and Threatens Democracy. New York: Broadway Books.