Primer on artificial intelligence and robotics

The International Federation of Robots (IFR), an international industrial group focused on commercial robotics, defines an industrial robot as an “automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes, which can be either fixed in place or mobile for use in industrial automation applications.”⁴ While this definition is a starting point, other roboticists may differ on dimensions such as whether a robot must be automatically controlled or could be autonomous or whether a robot must be reprogrammable. At a broader level, any machine that can be used to carry out complex actions or tasks in an automatic manner may be considered a robot.

Similar to robotics, artificial intelligence is a construct with varying definitions and potentially broad interpretations. For starters, it is useful to distinguish between general and narrow artificial intelligence (Broussard 2018). “General artificial intelligence” refers to computer software that can think and act on its own; nothing like this currently exists. “Narrow artificial intelligence” refers to computer software that relies on highly sophisticated, algorithmic techniques to find patterns in data and make predictions about the future. In this sense, the software “learns” from existing data and hence is sometimes referred to as “machine learning” but this should not be confused with actual learning. Broussard (2018) writes that “machine ‘learning’ is more akin to a metaphor…: it means that the machine can improve at its programmed, routine, automated tasks. It doesn’t mean that the machine acquires knowledge or wisdom or agency, despite what the term learning might imply [p. 89].”

Many applications of machine learning focus on prediction and estimation of unknowns based on a given set of information (Athey 2018; Mullainathan and Spiess 2017). There are a variety of algorithms that can be used for this machine learning. Some of these techniques are relatively straightforward uses of logit models which would be familiar to most organizational scholars, whereas others involve highly sophisticated algorithms that attempt to mimic how a human brain looks for patterns in data (the latter are called “neural networks”). Artificial intelligence technology can be used towards a variety of purposes, including playing abstract strategy games such as Chess or Go; to playing real-time video games such as Atari, Asterix, or Crazy Climber; to image or street number recognition; to natural language translation; and many other uses.

Automation refers to the use of largely automatic, likely computer-controlled, systems and equipment in manufacturing and production processes that replace some or all of the tasks that previously were done by human labor. Automation is not a new concept, as innovations such as the steam engine or the cotton gin can be viewed as automating previously manual tasks. One of the concerns for scholars in this area revolves around how and in what contexts increased use of robotics and artificial intelligence technology may lead to increased automation, and the impact that this form of increased automation may have on the workforce and the design of organizations.

While artificial intelligence, robotics, and automation are all related concepts, it is important to be aware of the distinctions between each of these constructs. Robotics is largely focused on technologies that could be classified as “manipulators” as per the IFR definition, and accordingly, more directly relates to the automation of physical tasks. On the other hand, artificial intelligence does not require physical manipulation, but rather computer-based learning. The distinction between the two technologies can become fuzzier as applications of artificial intelligence may involve robotics or vice versa. For example, “smart robots” are robots that integrate machine learning and artificial intelligence to continuously improve the robots’ performance.

Both artificial intelligence and robotics technologies are capable of automation. However, an open question is how and whether the effects of automation may differ across the two technologies. Some scholars contend that computerization and the increased use of artificial intelligence have the potential to automate certain non-routine tasks compared to the more rote tasks previously subjected to automation (Frey and Osborne 2017; Autor et al. 2006). Accordingly, it is possible that technologies incorporating artificial intelligence may be able to automate far more tasks than pure robotics-based technologies.

In addition to the distinction across the concepts of robotics, artificial intelligence, and automation, we additionally draw readers’ attention to the contrast between artificial intelligence and robotics, and computerization and information technologies more generally. Similarly to robotics and artificial intelligence, information and communication technology (ICT) has been of interest to researchers and policymakers with regards to both its potential to increase productivity and its ability to affect labor (e.g., Autor et al. 2003; Bloom et al. 2014; Akerman et al. 2015). However, while artificial intelligence and robotics may reduce the cost of storing, communicating, and transmitting information much like ICT, they are distinct. ICT can refer to any form of computer-based information system (Powell and Dent-Micallef 1999), while artificial intelligence and robotics may be computer-based but are not necessarily information systems. This distinction can be especially difficult to navigate given the broadness and variation in the definitions used for robotics and artificial intelligence in the literature. Again, we urge organizational scholars to carefully define any of these constructs in their studies.

Research on robotics and artificial intelligence builds off of the substantial body of literature surrounding innovation and technological development. Innovation is a key factor in contributing to economic growth (Solow 1957; Romer 1990) and has been an area of interest for both theorists and policymakers for decades. Literature on robotics and automation has pointed to the impressive potential of these new technologies. Brynjolfsson and McAfee (2017) claim that artificial intelligence has the potential to be “the most important general-purpose technology of our era.” Graetz and Michaels (2018) suggest that robotics added an estimated 0.37 percentage points to annual GDP growth for a panel of 17 countries from 1993 and 2007, an effect similar to that of the adoption of steam engines on economic growth during the industrial revolution.

Historically, excitement around radical new technologies is tempered by anxieties regarding the potential for labor substitution (Mokyr et al. 2015). A body of work has shown that automation spurred by innovation can both complement and substitute for labor. Acemoglu and Restrepo (2018) examine how increased industrial robotics usage has impacted regional US labor markets between 1990 and 2007. Their findings suggest that the adoption of industrial robotics is negatively correlated with employment and wages—specifically that each additional robot reduced employment by six workers and that one new robot across a thousand workers reduced wages by 0.5%. Graetz and Michaels (2018) find that while wages increase with robot use, on average, hours worked drops for low- and middle-skilled workers. A similar study in Germany suggests that each additional industrial robot leads to a loss of two manufacturing job, but these jobs are offset by newly created roles in the service industry (Dauth et al. 2017).

Increasingly, work on automation considers or focuses on artificial intelligence rather than just robotics. Frey and Osborne (2017) predict how increased computerization, in particular, machine learning technologies, will affect non-routine tasks. Based on the tasks most involved in an occupation, the authors propose which occupations may be more or less at risk of automation in the future. Their results suggest that 47% of employment in the USA is at high risk of computerization. Frey and Osborne’s work has been applied by researchers in other countries. Using the same methodology, Brzeski and Burk (2015) suggest that 59% of the German workforce may be highly susceptible to automation, while Pajarinen and Rouvinen (2014) suggest that 35% of Finnish jobs are at high risk. Similar to the task-based approach utilized by Frey and Osborne, Brynjolfsson et al. (2018b) take a task-based approach to assess occupations’ suitability for machine learning. They show that occupations across the wage and wage bill spectrum are equally susceptible, suggesting that machine learning will likely affect different parts of the workforce than earlier waves of automation.

Work on automation and labor has focused on different units of analysis. Much of the existing work in economics has focused on the economy as a whole. For example, Frey and Osborne (2017) measure the risk of automation on an occupation by occupation level but consider the occupations at a global level. Similar work by McKinsey Global Institute (MGI 2017) does the same, and recent work by Accenture considers these at the country level (Accenture 2018). US-specific work has been done by Brynjolfsson et al. (2018b) and Felten et al. (2018). Some research has taken a more focused approach and highlights the effect of artificial intelligence and automation on specific sectors of the economy. For example, Acemoglu and Restrepo (2018) highlight that the largest effects of technology adoption will occur in manufacturing, especially among manual and blue-collar occupations and for workers without a college degree.

Existing work on artificial intelligence and robotics has also attempted to identify “winners” and “losers” and to understand the distributional effects of these new technologies. A body of this work looks at cross-industry effects. Autor and Salomons (2018) show that industry-specific productivity increases are associated with a decrease of employment within the affected industry; however, positive spillovers in other sectors more than offset the negative own-industry effect. Similarly, Mandel (2017) examines brick-and-mortar retail stores during the rise of e-commerce and finds that new jobs created at fulfillment and call centers more than make up for job losses at department stores.

Other work looks at how skill composition can affect the potential complementary or substitution effects of these new technologies. A recent working paper by Choudhury et al. (2018) looks at performance effects of the use of artificial intelligence by workers with different types of training. They find productivity with artificial intelligence technology is highly affected by an individual’s background with computer science and engineering. Individuals who have requisite computer science or engineering skills are better able to unlock superior performance using artificial intelligence technologies than individuals without those skills. Felten et al. (2018) use an abilities-based approach to assess the link between recent advances in artificial intelligence and employment and wage growth. They find that occupations that require a relatively high proportion of software skills see growth in employment when affected by artificial intelligence, while other occupations do not see a meaningful relationship between the impact of artificial intelligence and employment growth.

There is a growing literature in economics, strategy, and information systems that studies the use of machine learning algorithms in decision-making. Some of the authors in this literature use disaggregated, micro-level data to draw insights as to how artificial intelligence affects firms or individuals differently depending on their background. Some of this work examines whether and how the use of artificial intelligence and machine learning tools affects individual biases. For example, machine-based algorithms appear to outperform judges in making decisions regarding potential detainment pre-trial and also reduce inequities (Kleinberg et al. 2018). Hoffman et al. (2017) find that managers who choose to hire against recommendations constructed by machine-based algorithms choose worse hires. Together, these results appear to suggest that machine learning algorithms may have potential in improving decision quality and equity.

However, other research cautions that machine learning algorithms often contain their own form of bias. For example, a machine learning algorithm designed to deliver advertisements for Science, Technology, Engineering, and Math occupations targeted men more than women, despite the fact that the advertisement was explicitly intended to be gender-neutral (Lambrecht and Tucker 2018); Google’s Ad Settings machine learning algorithm displays fewer advertisements for high-paying jobs to females than to males (Datta et al. 2015); and artificial intelligence-based tools used in judicial decision-making appear to display racial biases (Angwin et al. 2016). While these biases are troubling, some argue that compared to the counterfactual of human decision-making, algorithmic processes offer improvements in quality and fairness, and in particular, machine learning tools are best able to mitigate biases when human decision-makers exhibit bias and high levels of inconsistency (Cowgill 2019).

Recommender systems are a common tool on e-commerce platforms and frequently incorporate machine learning or artificial intelligence algorithms in the creation of their recommendations (Adomavicius and Tuzhilin 2005). Barach et al. (2018b) show that the use of recommendation systems for sellers can substitute for explicit monetary incentives in online marketplaces, highlighting one method by which firms can use artificial intelligence technologies to cut costs. Barach et al. (2018a, 2018b) study recommendation systems in online labor marketplaces and find that firms use AI-driven recommendations to identify an initial set of generally acceptable partners before relying on internal capabilities to select the best match. In particular, the use of the recommendation system is used less for specialized jobs and for experienced employees.

In addition to the above areas, research on artificial intelligence and robotics has started to examine a broader range of questions, such as how artificial intelligence may help stimulate innovation (Cockburn et al. 2018), the role of policy in an economy featuring artificial intelligence (Goolsbee 2018), and the role of artificial intelligence in international trade (Brynjolfsson et al. 2018a; Goldfarb and Trefler 2018). There are other important firm strategy and policy questions left to answer in this space such as the impact of artificial intelligence on firm structure, the factors that lead to increased adoption of these technologies, and distributional effects of artificial intelligence across industries, geographies, and occupations. However, aside from literature studying machine learning algorithms, research in this area has been slowed by a lack of available data, especially at the firm level. We discuss future directions of research below.

While there are some data sets containing information on the diffusion of robotics, it is largely at an aggregate level which does not allow for detailed microanalysis and differences across industries and regions can be obscured. There are currently no public data sets on the utilization or adoption of artificial intelligence at either the micro or the macro level, as the most complete sources of information are proprietary and inaccessible to the general public and the academic community (Raj and Seamans 2018; McElheran 2019). Despite these limitations, scholars studying management and organizations have constructed data sets and conducted research using trade magazines and other industry-specific resources. For example, using the industrial robotics industry as a setting, scholars have established that prior technological experience and technological knowledge are associated with greater innovative behavior following the introduction of a disruptive technology (Roy and Sarkar 2016; Roy and Islam 2017). Researchers have also used the industrial robotics industry as a setting to study organizational search and identify two distinct dimensions of search—search scope and search depth (Katila and Ahuja 2002). Nevertheless, the next stage in the evolution of research in this area should involve a proliferation of data to conduct a more focused and rigorous analysis of important questions regarding these technologies, firm adoption, and its consequences in an empirical manner.