Last October, the Creative Destruction Lab (CDL) of the University of Toronto’s Rotman School of Management organized a conference on Machine Learning and the Market for Intelligence. TFS was one of the selected participants in the event. Each year, the presentations delivered by speakers from the realm of machine intelligence are centred around a chosen theme. Last year’s 2019 conference focused on the technological advances in artificial intelligence and their impact on socio-economic power of AI corporations. I was particularly inspired by the speech “Power and Prediction” by Dr Ajay Agrawal, the founder of the CDL. Professor Agrawal analysed, using important economic concepts, how AI shifts power. Specifically, he concentrated on AIs used to make predictions for medical diagnosis.
According to Professor Agrawal, out of different types of AIs, there exist two main categories: efficiency AI and strategic AI. Efficiency AI is artificial intelligence used to replace an existing technique in order to increase efficiency and productivity, having a limited impact on power. Whereas, strategic AI has a significant impact on shifting power, targeting more than just inefficiency. It is strategic AI that focuses our discussion on power shifting.
Power can be shifted due to data feedback loops. Artificial intelligence systems optimize predictions and performance through feedback loops: by analyzing data and information continuously collected from its exposure. Sometimes, these feedback loops occur in isolation without the influence of consumers (e.g. a robotic vacuum cleaner navigating in a house), however, they often occur while interacting with market forces - customer demand and competitors. This interaction between feedback loops and market forces can lead to a power shift. As more feedback data is provided, the AI is able to make better, smarter and more accurate predictions which attract more users. Having more users provides the system with more varied and realistic data, thereby improving the quality of future predictions. As this cycle continues, power shifts into the hands of only a few AI companies that gathered a larger share of customers in the market. For example, if one company secures a competitive edge over other competitors due to a small advancement or innovation in their technology, users will begin to migrate from other competitors to this company. This is because users do not want to use the “second-best” AI. Once the company receives more users, the feedback loop begins, resulting in better predictions, more users, and ultimately, the concentration of more control and thus power over the consumers market. The small gap between competitors caused by the small innovation grows into a large difference. A real-life example of this power shift is the ride-hailing industry. Taxis used to be the most common form of ride-hailing. The power in the industry used to be distributed between different taxi companies around the world. Nowadays, companies like Uber have taken much of the power away from these companies. Uber uses AI to predict the demand for rides in different neighbourhoods. As it gains more users, the feedback cycle continues, providing more data and more accurate predictions. This shift in power is enabled because, in the domain of AI predictions, it is simple to migrate to a company providing higher-quality predictions without paying a higher price - high-quality predictions cost the same as low-quality predictions.
Further, the shifting of power can be caused by the minimum efficient scale. The minimum efficient scale is the scale of output where internal economies of scale have been fully exploited. In other words, it is the scale of output where the average cost of producing each unit is minimized. It is difficult for a company that has not achieved minimum efficient scale to compete with a competitor that has done so because its production costs are higher. This concept also applies to machine intelligence, where instead of changing the scale to lower per-unit costs, the amount of data (scale) is changed to increase prediction accuracy. The minimum efficient scale sets a bar for companies to compete in the market. The higher the minimum efficient scale within the industry, the harder it becomes to enter the market, because competitors must increase their scale and prediction accuracy until they meet the minimum efficient scale in order to compete. In addition to competition, regulations can set a certain minimum efficient scale to require a certain prediction accuracy (e.g. for greater safety in medical applications). A high minimum efficient scale creates a barrier to entry into the market. With this barrier, production and economic power concentrate in the hands of the few high producers that have reached minimum efficient scale. These companies operating at minimum efficient scale recognize that it is difficult for competitors to enter the market. They invest to move the threshold up to make it even more difficult for competitors to enter the market. An example of this power concentrated within the hands of companies operating at the minimum efficient scale is the airplane production industry. The threshold for the minimum efficient scale in the industry has been raised so high that companies would need to invest hundreds of billions of dollars to enter the industry and compete with Airbus and Boeing.
Based on the information above, it appears that at some level, technology leads to a gain of socioeconomic power by top AI companies. Whenever strategic AI is applied and has the ability to shift power, it is important to reflect upon the optimal strategies from the perspective of companies. Similarly, it is important to evaluate what can be done to optimize and enhance the public good from the perspective of governments. Before we enter a regime where technology can potentially lead to the concentration of power in the hands of a few companies, it is crucial for companies and governments to plan their strategies in order to lay the foundation for a society that is good and fair for everyone.