What Accountants Need to Know About AI
Artificial intelligence is divided into a variety of sub-fields, including but not limited to machine reasoning, machine learning, deep learning, and natural language processing. Each of these sub-fields has essential applications in accounting. In effect, Al has fields that may even be nested as sub-fields, such as Deep Learning (DL) is a sub-field of Machine Learning (ML), as illustrated in Figure 1.1.
As stated in the introduction, artificial intelligence is defined broadly as a computer program or software application that can imitate or simulate human behavior. This definition works well with a variety of technologies that we use every day. For example, the words that you are currently reading were spoken into a dictation tool in a popular word processing software. This form of Al uses speech recognition software to translate audio data into text, mimicking the actions of a human transcriber.
However, it is important to recognize that a generally accepted definition of artificial intelligence remains open for debate (Dobrev, 2012). An influential textbook in Al research entitled Artificial Intelligence: A Modern Approach offers four possibilities: systems that think like humans, systems that act like humans, systems that think rationally, and systems that act rationally (Russell & Norvig, 2010).
Expert Systems and other systems that mimic human behavior using rules-based or fuzzy-based algorithms.
Algorithms that enable machines to learn and improve without explicit programming based on training data sets.
Examples: supervised, unsupervised, semi-supervised, reinforcement learning.
Artificial Neural Networks: algorithms using models inspired by the biological neural networks found in human brains to analyze and learn from large volumes of data.
Figure 1.1 Relationship between Al, ML, and DL.
Today speech recognition is considered a form of Al and tomorrow it may be relegated to straightforward automation. The bottom line here is that what constitutes Al continues to evolve with time and technological advances. Although a consensus may not exist on a formal definition of Al, experts often make a distinction between two subtypes of Al - artificial narrow intelligence (AN1) and artificial general intelligence (AG1).
Artificial Narrow Intelligence (ANI)
Artificial narrow intelligence (ANI), also known as weak or narrow Al, focuses on a task, such as speech recognition, computers that can play chess, or autonomous vehicles. Virtually all of the Al with which we are familiar can be classified as AN1, as computer scientists have yet to create machines that can experience emotion and the ability to perceive or feel things (Jajal, 2018). AN I machines that are programmed for these specific tasks are typically much better than their humans at doing the same task.
Examples of narrow Al include chatbots that are built for the narrow purpose of conversation. Chatbots are capable of verbal responses, using algorithms and a database, based on a statement or question.
Artificial General Intelligence (AGI)
The goal of artificial general intelligence (AGI), also known as strong or broad Al, is to create machines capable of performing all the cognitive tasks of the human brain. The essential elements of AGI include: (1) the ability to apply knowledge from one domain to another; (2) the ability to plan for the future based on experience and knowledge; and (3) the ability to adapt based on changing circumstances (Walch, 2019).
Machine reasoning (MR) is the ability of a computer to draw conclusions from a knowledge base using automated inference techniques that can imitate or simulate human inference, such as deduction and induction. León Bottou, a reasoning expert, provides a more technical definition of MR as an “algebraic manipulation of previously acquired knowledge in order to answer a new question” (Bottou, 2014, p. 136).
The initial applications of reasoning systems were referred to as expert systems (ES). Expert systems are an early form of Al, first developed in the 1970s. They are computer systems that store knowledge from human experts to emulate human decision-making. To become an expert, one must have deep knowledge of both facts and rules as well as practical experience in a particular domain (Negnevitsky, 2011). Such expertise can be gained through formal training or hands-on experience. ES works by making recommendations or drawing conclusions on a narrow topic using a rules-based engine applied against a database. ES is used in a variety of disciplines in accounting, such as
What Accountants Need to Know 7 audit, tax, management accounting, and personal financial planning (Yang & Vasarhelyi, 1993).
Expert systems can be classified as rules-based or fuzzy-based. Rules-based expert systems employ a set of detailed instructions, called algorithms, that apply rules to interpret how to react to various scenarios. For example, a tax software program may use the following rule: if the adjusted gross income equals X dollars, then the applicable federal tax is equal to Y dollars.
Fuzzy-based expert systems use fuzzy logic - a logical system in which conditions cannot be described in binary terms such as true or false, 0 or 1. Fuzzy logic is useful in situations that involve imprecision and vagueness (Negnevitsky, 2011). For example, an algorithm using fuzzy logic was used in an expert system designed to assist with fraud detection of settled insurance claims (Pathak et al., 2005). These researchers also suggest that fuzzy logic expert systems could be used by auditors to assess control risks and detection risks on an audit engagement.
Today’s machine reasoning systems go beyond the rules-based capabilities of ES. Al-enabled MR systems can now learn using machine learning algorithms and artificial neural networks to drawconclusions through data analysis. As a result, there are vast opportunities for accountants to utilize MR systems to enhance productivity.