The Next Chapter of AI Empowerment

Background

The Wall Street Journal (WSJ) reported on the impressive potential of AI empowerment eight years ago. On January 5, 2017, the WSJ reported a story on AlphaGo, an AI-empowered computer program that can play the game of Go, which is popular in East Asia (“Humans Mourn Loss After Google Is Unmasked as China’s Go Master,” Wall Street Journal, Jan. 5, 2017, https://tinyurl.com/ekv676uk). Go is very complex when compared to chess because there are an enormous number of alternatives and strategies available for every move, but AlphaGo made history in March 2016 by beating Lee Sedol, South Korea’s top Go player. In a 2018 article in The CPA Journal, Allan Koltin predicted that using AI would enable the Big Four to reduce their campus recruiting of accountants (“The State of the Profession,” The CPA Journal, December 2018, pp. 26–33).

Created by DeepMind Technologies in 2014, the virtual master AlphaGo is a machine learning computer program which played thousands of games against itself and improved its strategy from every game. Eventually, in merely seven consecutive days, the tireless AlphaGo accomplished a stunning 60-0 record against top professional players, including 3-0 over Ke Jie, the world’s best player at the time (Steven Borowiec, Tracey Lien, “Alphago Beats Human Go Champ in Milestone for Artificial Intelligence,” Los Angeles Times, March 12, 2016, https://tinyurl.com/4ck6hsxb). The WSJ article on AlphaGo revealed that AI empowerment could have the following implications for accounting professionals:

• ▪ All top human Go players spent years or decades learning and training to perfect their strategies before reaching the summit, but AlphaGo arrived there just two years after it was created. This means that in complicated areas such as fraud detection, a properly developed AI-empowered virtual worker might be able to learn quickly and deliver results for firms.
• ▪ No human can play sixty Go games against world-class opponents in seven days. The implication of the AlphaGo story is that firms can use tireless virtual workers to perform tasks continuously anytime anywhere to increase productivity by deploying AI technology.
• ▪ The WSJ article predicted that computers will perform better than humans in increasingly complex skills and will continue to take on more roles now performed by humans.

The AlphaGo team published a paper in Nature that described the algorithms used (D. Silver et al., “Mastering the Game of Go with Deep Neural Networks and Tree Search,” Nature, Jan. 27, 2016, https://tinyurl.com/5t5xtbxn). Accordingly, the AI empowerment of AlphaGo was using the algorithms that focused on classifying historical data, recognizing patterns, refining analytics, and making decisions. As the next chapter of AI empowerment, GenAI is capable of creating new content that is similar to existing data including: research reports, grant proposals, images, music, or functioning computer programs (i.e., coding) (“What Is Generative AI?,” McKinsey, April 2, 2024, https://tinyurl.com/4kxu5w78).

In a recent report, Forrester found that Generative AI could augment and enhance existing business processes in ways that were previously impossible. It also predicts that ignoring Generative AI will be a costly mistake for businesses (“Ignoring Generative AI Will Be a Costly Mistake for Enterprises,” Forrester, Feb. 14, 2023, https://tinyurl.com/464x9udc). Thus, it is important for accounting professionals to understand the capabilities of GenAI and identify the best niche for their firms. To help accountants keep abreast of advancing technology, this article shows how ChatGPT works, offers a few accounting-related examples to demonstrate its prowess, and describes why ChatGPT is a good tool for accountants to stay productive. In addition, an example is provided to explain why ChatGPT’s answers might seem authoritative but could be wrong at times. Finally, several Generative AI tools are presented for accountants to start their AI empowerment journey.

ChatGPT

ChatGPT is an advanced chatbot that is capable of processing and responding in natural (i.e., human-like) language. Based on its prior training on large and diverse datasets, its ability to learn from prior conversations with humans, and its predictive capability, ChatGPT can generate new content within chat conversations. Specifically, ChatGPT’s training datasets included 570GB of data obtained from books, web texts, Wikipedia, articles, and other pieces of writing on the Internet. In other words, about 300 billion words have been fed into ChatGPT’s training datasets (N. Arya, “ChatGPT: Everything You Need to Know,” KD Nuggets, Jan. 16, 2023, https://tinyurl.com/mpa8x68h). ChatGPT 3.5 was trained using the data from various sources published until 2021, but more powerful versions of ChatGPT (e.g., 4, 4o, and 4.1) trained on larger datasets and with access to the Internet have also been released.

Using natural language prompts from humans, ChatGPT can generate summaries of text, help rewrite text, develop program code in different computer languages, provide helpful hints or problem-solving steps, plan itineraries for business trips, and create various types of text including poems, stories, and answers—pretty much anything that humans can accomplish using their intelligence, experience, and skills. ChatGPT has attracted considerable attention from knowledge workers across many fields—individuals who rely on their intellectual and analytical abilities to gather, process, and interpret data; generate valuable insights or solutions; and communicate useful information to users. For example, while ChatGPT can generate answers such as a firm’s inventory turnover, it does not perform the calculations that are often required for accountants to complete the same task. Instead, ChatGPT just finds the ratios in the existing datasets. In Exhibit 1, ChatGPT presents Costco’s inventory turnover ratio for the 2020 fiscal year along with the related formula/equation.

Although ChatGPT can provide answers quickly, it is heavily dependent on the underlying data sources because it does not use the current data available on the Internet (and thus relies on its training data sources). For example, ChatGPT 3.5 cannot provide Costco’s inventory turnover of 2022, which falls after the 2021 cut-off timeline of the training datasets it used.

ChatGPT is able to effectively use available data sources as needed to answer queries. For example, in Exhibit 2, ChatGPT compares the PPE turnover ratios, which are crucial for Marriott and Hilton, due to their extensive holdings in real estate. It might take an accountant a while to find all the required financial numbers to calculate those five annual PPE turnover ratios each for Marriott and Hilton.

In addition to looking up numeric data, ChatGPT is capable of handling unstructured and descriptive natural language text in crafting its response. In Exhibit 3, ChatGPT compares the cost structures of Marriott and Hilton by extracting similarities and differences found in different documents of the two hotel chains in its datasets.

Although ChatGPT can offer human-like responses to a wide range of questions, there is a certain randomness in its answers. To illustrate, ChatGPT offered a generic comparison between the two hotel chains in Exhibit 3 when the cost structures question was asked by one of the authors (addressed as PA by ChatGPT). As shown in Exhibit 4, when the same cost structures question was asked by a different person (addressed as AN), ChatGPT compared some numbers in their income statements but did not discuss cost behaviors, the underlying concepts of a firm’s cost structure.

As it continues to evolve and expand its natural language and reasoning capabilities, ChatGPT is likely to assist accountants in creating new content from updated datasets, synthesizing content from different sources, summarizing information from reports, translating content in different languages, and obtaining assistance to solve problems.

ChatGPT Hallucinations

One concern of using ChatGPT is that it does not provide references for its answers. Described as a “pioneer in deep learning and the self-supervised learning used in large language models,” (Craig Smith, “Hallucinations Could Blunt ChatGPT’s Success: OpenAI Says the Problem’s Solvable, Yann LeCun Says We’ll See,” IEEE Spectrum, Mar. 13, 2023, https://tinyurl.com/krvz7kus) Yann LeCunn explained, “Those systems generate text that sounds fine, grammatically, semantically, but they don’t really have some sort of objective other than just satisfying statistical consistency with the prompt.” For example, in response to a request for listing the articles on AI in The CPA Journal, ChatGPT’s answers are presented in Exhibit 5. ChatGPT said “I do not have access to the most current articles from the CPA Journal beyond the 2021 knowledge cutoff date. However, I can provide a list of some of the articles related to AI in the CPA Journal published before 2021.” All eight references provided by ChatGPT (in Exhibit 5) are non-existent, demonstrating a ChatGPT hallucination.

In natural language processing such as ChatGPT, a hallucination refers to generated content that is non-existent or false. Consequently, all users, including accountants, must use ChatGPT’s answers with caution. It is worth noting that Microsoft Copilot (previously Bing Chat) will generate an answer with references leading to the exact sources from which Copilot prepared the answers. Users can follow the links provided to view and verify these sources.

GenAI and the Big Four

CPA firms can potentially increase their productivity, enhance efficiency, and even reduce their headcounts by adopting GenAI. It is still an evolving technology and not every firm is willing to integrate GenAI into their practice. Exhibit 6 describes how each of the Big Four has already started projects using it.

Ethical and Legal Implications of Using Generative AI Tools

The early implementation of Generative AI has raised questions about its ultimate impact. “We need regulations that shield us during this large-scale world-wide change” (S. Mohasseb, “Post-ChatGPT, We Must Consider Human Purpose Beyond Work,” The Hill, Mar. 18, 2023, https://tinyurl.com/3vcdezmd). Citing an example of using Generative AI for marketing purposes by CarMax, Rowan Curran, a research analyst at Forrester, said the tools should be used with the same amount of “governance, controls, testing, and validation” that any other enterprise application or machine learning model needs (M. Holland, “The Accelerating Use of Generative AI May Prompt U.S. Action,” TechTarget, Feb. 3, 2023, https://tinyurl.com/bddx9kvm).

In 2023, the Federal Trade Commission (FTC) launched an investigation into OpenAI, probing whether the software maker had violated consumer protection laws by placing personal reputations and data at risk. Specifically, the FTC demanded that OpenAI provide detailed descriptions of all complaints it had received about its products generating false, misleading, disparaging, or harmful information. The FTC investigated whether the firm engaged in unfair or deceptive practices that resulted in “reputational harm” to people (“FTC Investigates OpenAI over Data Leak and ChatGPT’s Inaccuracy,” The Washington Post, July 13, 2023, https://tinyurl.com/mr3cnbea). Consequently, while using GenAI tools, CPAs need to be aware of protecting clients’ privacy and complying with any government regulations and their firm’s policy on GenAI.

Embracing GenAI

GenAI such as ChatGPT has the potential to revolutionize accounting practice by automating tasks and generating content, including creating accounting reports, analyzing financial statements, meeting compliance filing requirements, and developing insights for auditors and regulators (D. Beerbaum, “Generative Artificial Intelligence (GAI) with Chat GPT for Accounting – a Business Case,” SSRN, Mar. 11, 2023, https://tinyurl.com/3x6me2u6). The applications of Generative AI could be limitless since it could potentially be integrated in different ways wherever humans create content. For example, ChatGPT can compose replies to email messages, create social media posts, and write reviews. ChatGPT will disrupt the ways in which content is created and blur the lines between machines and humans due to its natural language capabilities.

Currently, ChatGPT functions as a versatile chatbot with general-purpose knowledge from a variety of domains in order to answer a wide spectrum of questions from users. In the same way that large firms like KPMG and PwC are working with Microsoft (Exhibit 6), smaller CPA firms can team up with a partner to develop a chatbot that is trained in specific domains including accounting, auditing, laws, or tax so that it can offer specialized content much like human experts in specific business disciplines.

Given that natural language processing is built into ChatGPT, users can ask any kind of question and expect to receive an answer, but there is no guarantee that the results will be accurate and reliable. The key is to ask the right question in order to receive good answers (J. Snow, “ChatGPT Can Give Great Answers. But Only If You Know How to Ask the Right Question,” Wall Street Journal, Apr. 12, 2023, https://tinyurl.com/3yf298pa). The WSJ article also reported that “asking the right question” is why some firms are hiring “prompt engineers”—people with the skills, background, and experiences to use Generative AI tools effectively. Thus, at least in the near term, prompt engineering may well become a desirable skill for many professionals, including CPAs.

Looking to Tomorrow

In today’s IT-intensive workplace, CPAs need to be conscious of technology trends and understand how to leverage emerging IT tools to enhance their businesses. CPAs are advised to start their AI empowerment journey sooner rather than later and begin adapting to a rapidly advancing technology. Specifically, CPAs can use one of the GenAI tools listed in Exhibit 7 and familiarize themselves with new technologies that could facilitate another level of AI empowerment. Exhibit 7 also offers examples of accounting-related tasks that can help CPAs select preferred tools according to their needs. Ultimately, CPAs seeking to advance their career should assess the potential use cases for GenAI.