1. The technical landscape

A short history of artificial intelligence

In 1950, British mathematician Alan Turing published a paper on computing machinery and intelligence (Turing, 1950[1]) posing the question of whether machines can think. He developed a simple heuristic to test his hypothesis: could a computer have a conversation and answer questions in a way that would trick a suspicious human into thinking the computer was actually a human?1 The resulting “Turing test” is still used today. That same year, Claude Shannon proposed the creation of a machine that could be taught to play chess (Shannon, 1950[2]). The machine could be trained by using brute force or by evaluating a small set of an opponent’s strategic moves (UW, 2006[3]).

Many consider the Dartmouth Summer Research Project in the summer of 1956 as the birthplace of artificial intelligence (AI). At this workshop, the principle of AI was conceptualised by John McCarthy, Alan Newell, Arthur Samuel, Herbert Simon and Marvin Minsky. While AI research has steadily progressed over the past 60 years, the promises of early AI promoters proved to be overly optimistic. This led to an “AI winter” of reduced funding and interest in AI research during the 1970s.

New funding and interest in AI appeared with advances in computation power that became available in the 1990s (UW, 2006[3]). Figure 1.1 provides a timeline of AI’s early development.

Figure 1.1. Timeline of early AI developments (1950s to 2000)

Source: Adapted from Anyoha (28 August 2017[4]), “The history of artificial intelligence”, http://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/.

The AI winter ended in the 1990s as computational power and data storage were advancing to the point that complex tasks were becoming feasible. In 1995, AI took a major step forward with Richard Wallace’s development of the Artificial Linguistic Internet Computer Entity that could hold basic conversations. Also in the 1990s, IBM developed a computer named Deep Blue that used a brute force approach to play against world chess champion Gary Kasparov. Deep Blue would look ahead six steps or more and could calculate 330 million positions per second (Somers, 2013[5]). In 1996, Deep Blue lost to Kasparov, but won the rematch a year later.

In 2015, Alphabet’s DeepMind launched software to play the ancient game of Go against the best players in the world. It used an artificial neural network that was trained on thousands of human amateur and professional games to learn how to play. In 2016, AlphaGo beat the world’s best player at the time, Lee Sedol, four games to one. AlphaGo’s developers then let the program play against itself using trial and error, starting from completely random play with a few simple guiding rules. The result was a program (AlphaGo Zero) that trained itself faster and was able to beat the original AlphaGo by 100 games to 0. Entirely from self-play – with no human intervention and using no historical data – AlphaGo Zero surpassed all other versions of AlphaGo in 40 days (Silver et al., 2017[6]) (Figure 1.2).

Figure 1.2. AlphaGo’s rapid self-learning to become best Go player in the world in 40 days

Source: adapted from Silver et al. (2017[6]), “Mastering the game of Go without human knowledge”, https://doi.org/10.1038/nature24270.

What is AI?

There is no universally accepted definition of AI. In November 2018, the AI Group of Experts at the OECD (AIGO) set up a subgroup to develop a description of an AI system. The description aims to be understandable, technically accurate, technology-neutral and applicable to short- and long-term time horizons. It is broad enough to encompass many of the definitions of AI commonly used by the scientific, business and policy communities. As well, it informed the development of the OECD Recommendation of the Council on Artificial Intelligence (OECD, 2019[8]).

Conceptual view of an AI system

The present description of an AI system is based on the conceptual view of AI detailed in Artificial Intelligence: A Modern Approach (Russel and Norvig, 2009[9]). This view is consistent with a widely used definition of AI as “the study of the computations that make it possible to perceive, reason, and act” (Winston, 1992[10]) and with similar general definitions (Gringsjord and Govindarajulu, 2018[11]).

A conceptual view of AI is first presented as the high-level structure of a generic AI system (also referred to as “intelligent agent”) (Figure 1.3). An AI system consists of three main elements: sensors, operational logic and actuators. Sensors collect raw data from the environment, while actuators act to change the state of the environment. The key power of an AI system resides in its operational logic. For a given set of objectives and based on input data from sensors, the operational logic provides output for the actuators. These take the form of recommendations, predictions or decisions that can influence the state of the environment.

Figure 1.3. A high-level conceptual view of an AI system

Source: As defined and approved by AIGO in February 2019.

A more detailed structure captures the main elements relevant to the policy dimensions of AI systems (Figure 1.4). To cover different types of AI systems and different scenarios, the diagram separates the model building process (such as ML), from the model itself. Model building is also separate from the model interpretation process, which uses the model to make predictions, recommendations and decisions; actuators use these outputs to influence the environment.

Figure 1.4. Detailed conceptual view of an AI System

Source: As defined and approved by AIGO in February 2019.

Environment

An environment in relation to an AI system is a space observable through perceptions (via sensors) and influenced through actions (via actuators). Sensors and actuators are either machines or humans. Environments are either real (e.g. physical, social, mental) and usually only partially observable, or else virtual (e.g. board games) and generally fully observable.

AI system

An AI system is a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. It does so by using machine and/or human-based inputs to: i) perceive real and/or virtual environments; ii) abstract such perceptions into models through analysis in an automated manner (e.g. with ML, or manually); and iii) use model inference to formulate options for information or action. AI systems are designed to operate with varying levels of autonomy.

AI model, model building and model interpretation

The core of an AI system is the AI model, a representation of all or part of the system’s external environment that describes the environment’s structure and/or dynamics. A model can be based on expert knowledge and/or data, by humans and/or by automated tools (e.g. ML algorithms). Objectives (e.g. output variables) and performance measures (e.g. accuracy, resources for training, representativeness of the dataset) guide the building process. Model inference is the process by which humans and/or automated tools derive an outcome from the model. These take the form of recommendations, predictions or decisions. Objectives and performance measures guide the execution. In some cases (e.g. deterministic rules), a model can offer a single recommendation. In other cases (e.g. probabilistic models), a model can offer a variety of recommendations. These recommendations are associated with different levels of, for instance, performance measures like level of confidence, robustness or risk. In some cases, during the interpretation process, it is possible to explain why specific recommendations are made. In other cases, explanation is almost impossible.

The AI system lifecycle

In November 2018, AIGO established a subgroup to inform the OECD Recommendation of the Council on Artificial Intelligence (OECD, 2019[8]) by detailing the AI system lifecycle. This lifecycle framework does not represent a new standard for the AI lifecycle2 or propose prescriptive actions. However, it can help contextualise other international initiatives on AI principles.3

An AI system incorporates many phases of traditional software development lifecycles and system development lifecycles more generally. However, the AI system lifecycle typically involves four specific phases. The design, data and models phase is a context-dependent sequence encompassing planning and design, data collection and processing, as well as model building and interpretation. This is followed by verification and validation, deployment, and operation and monitoring (Figure 1.5. AI system lifecycle). These phases often take place in an iterative manner and are not necessarily sequential. The decision to retire an AI system from operation may occur at any point during the operation and monitoring phase.

The AI system lifecycle phases can be described as follows:

1. 1. Design, data and modelling includes several activities, whose order may vary for different AI systems:

   • Planning and design of the AI system involves articulating the system’s concept and objectives, underlying assumptions, context and requirements, and potentially building a prototype.

   • Data collection and processing includes gathering and cleaning data, performing checks for completeness and quality, and documenting the metadata and characteristics of the dataset. Dataset metadata include information on how a dataset was created, its composition, its intended uses and how it has been maintained over time.

   • Model building and interpretation involves the creation or selection of models or algorithms, their calibration and/or training and interpretation.

2. 2. Verification and validation involves executing and tuning models, with tests to assess performance across various dimensions and considerations.

3. 3. Deployment into live production involves piloting, checking compatibility with legacy systems, ensuring regulatory compliance, managing organisational change and evaluating user experience.

4. 4. Operation and monitoring of an AI system involves operating the AI system and continuously assessing its recommendations and impacts (both intended and unintended) in light of objectives and ethical considerations. This phase identifies problems and adjusts by reverting to other phases or, if necessary, retiring an AI system from production.

Figure 1.5. AI system lifecycle

Source: As defined and approved by AIGO in February 2019.

The centrality of data and of models that rely on data for their training and evaluation distinguishes the lifecycle of many AI systems from that of more general system development. Some AI systems based on ML can iterate and evolve over time.

AI research

This section reviews some technical developments with regard to AI research in academia and the private sector that are enabling the AI transition. AI, and particularly its subset called ML, is an active research area in computer science today. A broader range of academic disciplines is leveraging AI techniques for a wide variety of applications.

There is no agreed-upon classification scheme for breaking AI into research streams that is comparable, for example, to the 20 major economics research categories in the Journal of Economic Literature’s classification system. This section aims to develop an AI research taxonomy for policy makers to understand some recent AI trends and identify policy issues.

Research has historically distinguished symbolic AI from statistical AI. Symbolic AI uses logical representations to deduce a conclusion from a set of constraints. It requires that researchers build detailed and human-understandable decision structures to translate real-world complexity and help machines arrive at human-like decisions. Symbolic AI is still in widespread use, e.g. for optimisation and planning tools. Statistical AI, whereby machines induce a trend from a set of patterns, has seen increasing uptake recently. A number of applications combine symbolic and statistical approaches. For example, natural language processing (NLP) algorithms often combine statistical approaches (that build on large amounts of data) and symbolic approaches (that consider issues such as grammar rules). Combining models built on both data and human expertise is viewed as promising to help address the limitations of both approaches.

AI systems increasingly use ML. This is a set of techniques to allow machines to learn in an automated manner through patterns and inferences rather than through explicit instructions from a human. ML approaches often teach machines to reach an outcome by showing them many examples of correct outcomes. However, they can also define a set of rules and let the machine learn by trial and error. ML is usually used in building or adjusting a model, but can also be used to interpret a model’s results (Figure 1.6). ML contains numerous techniques that have been used by economists, researchers and technologists for decades. These range from linear and logistic regressions, decision trees and principle component analysis to deep neural networks.

Figure 1.6. The relationship between AI and ML

Source: Provided by the Massachusetts Institute of Technology (MIT)’s Internet Policy Research Initiative (IPRI).

In economics, regression models use input data to make predictions in such a way that researchers can interpret the coefficients (weights) on the input variables, often for policy reasons. With ML, people may not be able to understand the models themselves. Additionally, ML problems tend to work with many more variables than is common in economics. These variables, known as “features”, typically number in the thousands or higher. Larger data sets can range from tens of thousands to hundreds of millions of observations. At this scale, researchers rely on more sophisticated and less-understood techniques such as neural networks to make predictions. Interestingly, one core research area of ML is trying to reintroduce the type of explainability used by economists in these large-scale models (see Cluster 4 below).

The real technology behind the current wave of ML applications is a sophisticated statistical modelling technique called “neural networks”. This technique is accompanied by growing computational power and the availability of massive datasets (“big data”). Neural networks involve repeatedly interconnecting thousands or millions of simple transformations into a larger statistical machine that can learn sophisticated relationships between inputs and outputs. In other words, neural networks modify their own code to find and optimise links between inputs and outputs. Finally, deep learning is a phrase that refers to particularly large neural networks; there is no defined threshold as to when a neural net becomes “deep”.

This evolving dynamic in AI research is paired with continual advances in computational abilities, data availability and neural network design. Together, they mean the statistical approach to AI will likely continue as an important part of AI research in the short term. As a result, policy makers should focus their attention on AI developments that will likely have the largest impact over the coming years and represent some of the most difficult policy challenges. These challenges include unpacking the machines’ decisions and making the decision-making process more transparent. Policy makers should also keep in mind that most dynamic AI approaches – statistical AI, specifically “neural networks” – are not relevant for all types of problems. Other AI approaches, and coupling symbolic and statistical methods, remain important.

There is no widely agreed-upon taxonomy for AI research or for the subset of ML. The taxonomy proposed in the next subsection represents 25 AI research streams. They are organised into four broad categories and nine sub-categories, mainly focused on ML. In traditional economic research traditions, researchers may focus on a narrow research area. AI researchers commonly work across multiple clusters simultaneously to solve open research problems.

Cluster 3: Ways of improving ML/optimisations

The third broad category of research focuses on ways to improve and optimise ML tools. It breaks down research streams based on the time horizon for results (current, emerging and future) (Table 1.3). Short-term research is focusing on speeding up the deep-learning process. It does this either via better data collection or by using distributed computer systems to train the algorithm.

Table 1.3. Cluster 3: Ways of improving ML/optimisations
Ways of improving ML	Enabling factors (current)	Faster deep learning
		Better data collection
		Distributed training algorithms
	Enabling factors (emerging)	Performance on low-power devices
		Learning to learn/meta learning
		AI developer tools
	Enabling factors (future)	Understanding neural networks
		One-shot learning
Source: Provided by the MIT’s IPRI.

Researchers are also focused on enabling ML on low-power devices such as mobile phones and other connected devices. Significant progress has been made on this front. Projects such as Google’s Teachable Machine now offer open-source ML tools light enough to run in a browser (Box 1.2). Teachable Machine is just one example of emerging AI development tools meant to expand the reach and efficiency of ML. There are also significant advances in the development of dedicated AI chips for mobile devices.  

ML research with a longer time horizon includes studying the mechanisms that allow neural networks to learn so effectively. Although neural networks have proven to be a powerful ML technique, understanding of how they operate is still limited. Understanding these processes would make it possible to engineer neural networks on a deeper level. Longer-term research is also looking at ways to train neural networks using much smaller sets of training data, sometimes referred to as “one-shot learning”. It is also generally trying to make the training process more efficient. Large models can take weeks or months to train and require hundreds of millions of training examples.

Box 1.2. Teachable Machine

Teachable Machine is a Google experiment that allows people to train a machine to detect different scenarios using a camera built into a phone or computer. The user takes a series of pictures for three different scenarios (e.g. different facial expressions) to train the teachable machine. The machine then analyses the photos in the training data set and can use them to detect different scenarios. For example, the machine can play a sound every time the person smiles in a camera range. Teachable Machine stands out as an ML project because the neural network runs exclusively in the user’s browser without any need for outside computation or data storage (Figure 1.7).

Figure 1.7. Training a machine using a computer’s camera

Source: https://experiments.withgoogle.com/ai/teachable-machine.

Cluster 3: Policy relevance

The policy relevance of the third cluster includes the implications of running ML on stand-alone devices and thus of not necessarily sharing data on the cloud. It also includes the potential to reduce energy use, and the need to develop better AI tools to expand its beneficial uses.