The promise and peril of ChatGPT

Early adopters are finding uses for ChatGPT in accounting, but the chatbot raises concerns about security and accuracy.

ChatGPT, the artificial intelligence-powered chatbot, has captured global attention with its fast-evolving ability to interpret questions and commands to produce coherent, even human-seeming responses. Within five days of ChatGPT’s release in November 2022, more than a million people had tried it out. A few months later, that number had swelled to 100 million monthly users, making it the fastest-growing consumer application ever.

ChatGPT has shown itself capable of composing everything from website copy to college essays and restaurant reviews. It has been used to produce bizarre philosophical conversations and even contest-winning art. In finance and accounting, early adopters already are using language-based AI interfaces to create formulas and programming scripts, write marketing material, and draft responses to clients.

These early experiments have yielded timesaving tricks, intriguing avenues for exploration, and, for some, profound questions about the profession’s future. However, users should proceed with caution: ChatGPT’s responses can contain errors and untruths, and the chatbot also poses security risks.

ACCOUNTANTS ARE PUTTING CHATGPT TO WORK IN EXCEL AND PROGRAMMING

ChatGPT was the culmination of years of recent advances that have put powerful new AI tools into the hands of the general public, such as Midjourney, DALL•E 2, and Stable Diffusion, which can generate images in countless artistic styles in response to users’ prompts, and similar services that focus on audio, text, and video. Some of the most popular tools are derived from the work of OpenAI, the organization that operates ChatGPT and the GPT model that powers it.

ChatGPT is a deceptively simple platform. The chatbot allows a user to input any text they would like. When a message is sent, the platform responds almost instantly with a written response.

There are few limitations or boundaries, with ChatGPT gamely trying to respond to any request based on its unfathomably large knowledge base. The designers of the GPT models have not disclosed the scope or scale of the data library that powers services like ChatGPT, but it is apparent that the platform has ingested everything from pop culture ephemera to texts on economics.

When a user’s prompt is received, GPT-based systems generate an answer based on the pattern it detects in the materials it has already been trained on. GPT stands for “generative pre-trained transformer.” The system is often described as a neural network machine learning model that incorporates billions of datapoints. When it generates content, the system is choosing sequences of words, based on the patterns in the data it has “learned.” Its knowledge has been reinforced by training, during which its incorrect answers are flagged by reviewers.

The results are uncanny, with ChatGPT able to generate content that matches countless styles and themes, as requested by the user. For example, when asked to write a haiku about how automation is affecting accounting, the system quickly typed out:

Automation hums,
Accountants adapt to change,
Efficiency blooms.

Those wide-open possibilities, however, can also mask some of the system’s most useful capabilities.

“I think a lot of people are struggling with trying to even understand what to use it for,” said Byron Patrick, CPA/CITP, CGMA, vice president for client success at The B3 Method Institute.

He and others are quickly finding new uses, though — and they’ve uncovered some surprising capabilities, especially related to Excel formulas and programming.

For example, Patrick recently had a “pile of data” that he needed to structure. The information, including phone numbers and addresses, had inadvertently been condensed into a single column of a spreadsheet, making it difficult to use. Patrick wanted to restore the data into different columns, so he wrote a general description of the messy dataset in ChatGPT’s text entry box and described how he wanted the bot to solve the problem.

“I literally wrote the prompt like I was trying to explain it to a kid,” Patrick said. In response, the system quickly produced a macro for Excel, which Patrick fine-tuned in a matter of minutes by making further requests of ChatGPT.

“It was so cool,” Patrick said. Besides producing functional code, the system included explanations of how that code was identifying phone numbers and addresses for sorting.

“I could have probably spent six hours and gotten to something that was 90% right,” he said. “This thing in 15 minutes got me there in a better path than I ever could have imagined.”

The GPT model has knowledge of the programming interfaces of countless web-integrated technology platforms, allowing it to write code for interacting with those products, Patrick explained. “I think it’s going to lead to a surge in custom tools being built by accountants,” he said.

ACCELERATING EMAIL AND RESEARCH PROJECTS

Jason Staats, an Oregon-based CPA and founder of Realize, an accountancy community, has delved into ChatGPT’s capabilities in written communications. Staats has used the system to improve his emails and speed up the writing process.

“I have it rewrite a lot of the things I write,” he said. “If my first stab is too terse, I’ll have it rewrite it in a more friendly tone.”

Doing so is as simple as pasting text into the program and asking it for a revision — much as you would with a human being.

Staats suggested that accountants can use ChatGPT to improve their marketing materials. The system can be asked to take the style or details of current marketing materials and translate them to a new audience.

“There’s a whole level of creativity there that it has,” he said.

Staats and others also use ChatGPT as a research tool, though with some guardrails. For example, he might ask the system to explain a concept in simple terms and then use its suggestions as a jumping-off point for further research.

Pascal Finette, co-founder of a Coloradobased tech consultancy named be radical, said that extended conversations with ChatGPT can surface new ideas and unexpected connections to other areas.

“Take something you’re interested in, and see what you can do with the system,” Finette said. “There is incredible magic in having the world’s information at your fingertips in an aggregated format.”

CHATGPT IS OFTEN CONFIDENTLY WRONG

For all their capabilities, though, ChatGPT and similar platforms have some critical weaknesses. Most notably, they get things wrong.

ChatGPT has sometimes given the wrong answer to simple math problems and struggles with more complex arithmetic. It also has shown a tendency to get facts wrong. These models can also “hallucinate” — introducing false and even bizarre ideas for no apparent reason. In some cases, the systems may be drawing from dubious knowledge sources, or they may be inappropriately combining different texts.

That raises concerns for CPAs planning on using AI to generate website or client communications material. Anything ChatGPT produces must be carefully vetted for accuracy. Cautionary tales about improper use of AI have already come to light: CNET and Bankrate recently had to post corrections to personal finance articles written by AI after the website Futurism pointed out they contained obvious errors.

Chatbots can also introduce their own errors. In tests by this writer, the system displayed a decent ability to produce code and instructions for projects in Google Sheets, Python, and other products and programming languages. However, the platform also made mistakes — sometimes generating faulty code or referring to a function that didn’t exist in the software in question.

“This is where these systems have, I think, the tendency to danger,” Finette said. “They appear to be incredibly confident. The response has a way of saying, ‘This is it.’”

ChatGPT can also pose security risks. Per the terms of its user agreement, OpenAI is allowed to use data that users input to the free version of ChatGPT in order to improve its products. OpenAI states that it removes personal identification from such data. Nevertheless, there is a risk that if users input sensitive data, it could become part of the database ChatGPT is trained on — and potentially appear in other users’ chats.

WHAT COMES NEXT?

Further advances will come as the technology beyond GPT is adapted from general-purpose models to more specific uses. For example, Patrick said, a model could be given access to a firm’s client files, allowing it to become a customer-facing agent.

“There’s a lot of questions that [CPAs] get from their clients — ‘What’s the status of my return? Where is this? What is this?’” Patrick said. “If I could just point this thing to my internal practice management systems and have it answer, that’d be huge.”

Of course, companies are already using naturallanguage processing for customer service — think of the automated chat windows on many business websites — but the latest versions of the technology will have greater capabilities to actually answer a client’s question.

“I think we are at the very, very beginning of what is going to be possible in the next five to 10 years,” Patrick added.

Advances in AI also will lead to serious upgrades in the internal tools that firms use, Staats said. Tax research companies will deploy bots that can more adeptly answer questions about laws and regulations, Patrick predicted.

And he expects that the major manufacturers of software for accountants will move quickly to incorporate AI technology that can intelligently surface relevant data and make helpful suggestions.

“The industry tools will be a little slower to use [the new models] than the mass-market tools,” Patrick said, but he predicted significant changes within the year.

Jillian Bommarito, CPA, is co-founder of 273 Ventures, a company that is developing a platform to manage disconnected and unstructured data systems for legal and compliance purposes. She expects that AI models will allow automation efforts to break out of their silos. Instead of being deployed in specialized and finely tuned methods, such as robotic process automation, automation technologies will become more flexible and easily applied across an organization. A customerservice bot might be able to rewrite stock answers to fit a specific customer’s problem, or an administrative bot could surface transcripts from previous meetings in real time to aid discussions. A form-processing bot that normally breaks when it encounters an unexpected bit of information might be able to instead look at how similar issues were resolved in the past and draw on the company’s data to suggest possible answers.

The technologies are “going to interact more, and it’s going to basically make everything more powerful,” she said.

But for now, GPT and similar models remain accessible largely in prototype and demonstration versions — and that’s part of why Amy Vetter, CPA/CITP, CGMA, the CEO of The B3 Method Institute, urged caution.

“It’s a nice thing to have at your fingertips, but I’m not real sure yet if it’s reliable information,” she said. She underlined that platforms like ChatGPT are constantly gathering data from their users and building their base of knowledge.

“Be careful when you don’t know the purpose behind a technology,” she said. “Your [intellectual property], all the things that differentiate you as a business, it’s important that you’re careful how you’re sharing something like that. Put the same rigor you would put against any other technology for privacy and security issues.”

It’s difficult to predict just how fast and how far the technology will advance from here, especially considering the leaps it has made in recent years. Indeed, GPT-4 was released in March 2023 and quickly displayed a range of new capabilities and improved performance.

For now, Staats and others advised, the best course of action is to experiment and maintain a sense of skepticism — as well as some imagination. Many of the technology’s potential uses have yet to be discovered, and they’re constantly changing.

“Try not to get too fixated on today,” Staats said, “because, honestly, next week it could be wildly different.”

About the author

Andrew Kenney is a freelance writer based in Colorado.

LEARNING RESOURCES

Ethics and Artificial Intelligence in Business and Finance

Focuses on ethics related to artificial intelligence for business and finance and offers practical examples for what to consider in building a model for ethics and governance.

Artificial Intelligence for Accounting and Finance Professionals

The term artificial intelligence (AI) has been in the news for decades, inspiring a mystique that has rarely lived up to the expectations … until now! AI now offers the potential to revolutionize the way we do business from reinventing customer experiences to prescriptive forecasting leading to a significant competitive advantage.

Artificial Intelligence: Strategy & Applications in Accounting, Audit, and Financial Management

This four-part webcast series will help you learn the business basics of artificial intelligence (AI), machine learning, and the critical function of data management and automation in your enterprise planning. Enhance your understanding of intelligent automation applications within the audit, accounting, and finance world and the crucial role of governance and ethics in the AI revolution.

For more information or to make a purchase, go to aicpa-cima.com/cpe-learning or call the Institute at 888-777-7077.

AICPA RESOURCES

Articles

“Scaling Artificial Intelligence: A Key Strategy for Businesses,” FM magazine, March 23, 2023

“Using ChatGPT with Excel,” JofA, Jan. 30, 2023

“Quiz: The Ethical Pitfalls of AI,” FM magazine, Dec. 1, 2021

“6 Lessons From Audit Experts Who Adopted AI Early,” JofA, Nov. 23, 2021

“Building Ethics Into AI,” FM magazine, April 6, 2020

“Why AI Ethics Help Manage Risk,” FM magazine, Feb. 3, 2020