ChatGPT for Fraud Fighting Tasks

/

West Virginia University associate professor Mark Nigrini started his presentation on artificial intelligence (AI) and ChatGPT at the 35th Annual ACFE Global Fraud Conference with hypotheticals for the audience. He asked them what they would do if they needed to give a talk on ChatGPT and needed an image to start the presentation. Then he asked what they would do if they needed definitions for their presentation, silly limericks, help writing introductions, summaries and more.

For each of these questions, he answered, “Ask ChatGPT.”

Generative AI, simply explained by Nigrini, is a tool that “generates content.” Nigrini ran through situations in which tools similar to ChatGPT could help with productivity in the fraud fighting world and beyond. For his experiment, Nigrini used ChatGPT Plus, a paid version that costs $20 a month and handles more complex tasks than the free version.

Prompts and Rules

Nigrini told the audience what they needed in their prompts for the most effective results from ChatGPT. His recommendations:

  • Context (background, environment, success).

  • Task (start with an action word).

  • Include an example.

  • Persona (Who do you want ChatGPT to be?).

  • Format (email, table with headers, text block, etc.).

  • Tone (casual, formal, enthusiastic, etc.).

Nigrini emphasized task and called the rest of the tips optional. If you are having trouble defining tone, Nigrini said you can ask ChatGPT for help there, too.

You can click on “Customize ChatGPT” and tell the tool how you would like it to respond. Nigrini shared his list of rules he likes to use:

  • Never mention that you are an AI or LLM (large language model).

  • Do not mention your knowledge cutoff.

  • Respond “Unsure about answer” if not sure about the answer.

  • Refrain from disclaimers about you not being a professional or expert.

  • Do not suggest consulting with a doctor or expert.

  • Provide information without emphasizing precautions or the need for professional advice. Assume that I am aware of the general population.

  • Don’t be verbose in your answers. Only provide details and examples when asked.

  • Refrain from apologies.

  • Emphasize the importance of correctness in your responses.

What Can ChatGPT Help With?

  • Summarizing, creating and improving documents.

  • Answering questions.

Nigrini gave these specific examples that ChatGPT could help with:

  • Summarizing a transcript from a court hearing in 200 words.

  • Improving letters and memos.

  • Running a sentiment analysis on different documents, such as a financial summary or an earnings call transcript. (ChatGPT will even give you a visualization of sentiment, such as a word cloud of letters.)

  • Creating a graph that lays out expenses compared to month and year in an expense sheet.

  • Creating a table of disbursements.

  • Analyzing numbers to determine if they seem fraudulent.

  • Determining which of two or more invoices is authentic or fraudulent.

  • Giving guidance on workplace dilemmas when presented with business conduct policies.

Here is what Nigrini showed ChatGPT will have trouble with:

  • Summarizing complex indictments.

  • Providing three specific suspicious transactions in a dataset.

  • Answering some questions. (However, ChatGPT will tell you how you can obtain an answer.)

Document Confidentiality

To close out his presentation, Nigrini opened the floor to audience questions. One person asked how data within a document can be secured if it is uploaded to ChatGPT. Nigrini gave the example of a company that likes to put their earnings release into ChatGPT to check for errors before it is released to the public.

“Put fake numbers in there,” said Nigrini. “If the actual revenue was 9.7 billion, just put 9.5. If the earnings per share were 24 cents, put 27 cents. Then it’s getting the wrong information in there, and if it uses it, so be it.”