The emergence of ChatGPT has put artificial intelligence in the spotlight again, arousing continuous heated discussions and attention. As an important support, the GPT model has intensively released many large language models at home and abroad recently. OpenAI launched GPT-4, Google launched large models such as LaMDA and PaLM, and Meta launched an open source large model LLaMA. Also followed closely behind, Baidu launched "Wen Xin Yi Yan", Ali launched "Synonymous Thousand Questions", Huawei launched "Pangu Large Model", Tencent launched "Hunyuan Large Model", JD.com launched "ChatJD", SenseTime also launched Recently, "Daily New SenseNova" was launched, and HKUST Xunfei, Megvii, Kuaishou, etc. also have large-scale model layouts. As more and more manufacturers and individuals start to access GPT applications, the AI ecosystem and scene-based layout are becoming more and more perfect, and various GPT tools that have sprung up like mushrooms after the spring rain officially mark that we have entered the real artificial intelligence. Smart age.
A research report recently released by OpenAI predicts that accountants and auditors are among the occupations that are most likely to be subverted, which naturally adds a lot of crisis and panic to us.
Will GPT replace auditors?
This topic has been under discussion for a long time. The three core elements of the development of the GPT model are data, algorithms, and computing power. Data is continuously generated every day, and the algorithm is also self-iterating in the feedback again and again, and the computing power also follows the application. The volume is constantly increasing, as if artificial intelligence can finally replace us. We know that GPT, as a learning model, is actually an expression of human will and thinking to some extent. Strictly speaking, at least at the current stage, the self-awareness of artificial intelligence cannot be awakened. Although according to the latest foreign research, artificial intelligence consciousness There have been signs of "independence" slowly, so a group of bigwigs led by Musk began to jointly urgently stop the development of "GPT-5".
The core reason for holding the view that "GPT will not replace auditing" is that "AI will not take the blame." In fact, regarding this point, the core is whether to define AI as a tool, or as a subject of consciousness, or even at the legal level. It is obvious that artificial intelligence cannot exist independently of human subject consciousness at this stage, because the subject identity and relationship of auditing in the corresponding relationship cannot be replaced. Secondly, when discussing the degree and speed of audit replacement, there is another main line object, that is, a series of models established based on financial digitization, and the effectiveness of their operation in practice. Obviously, relying on "machines" alone is temporarily cannot handle such complex scenarios.
Another part of the people's point of view is that GPT will replace the work of auditors. The core reason is that all behavior patterns of people in specific audit projects can be optimized in continuous algorithm iterations, and GPT's own understanding And processing, learning ability to a large extent far exceeds that of humans. When all the results output by GPT can be bound to a specific subject identity, and the corresponding independence and professionalism are guaranteed to the greatest extent , and guarantee the underlying security, reliability, adaptability, etc., at least on this basis, the "GPT+human" model will indeed replace the simple human model.
So the core of this issue at this time is actually to discuss which part of the GPT replaces people, and how GPT is integrated into the audit work.
Application of GPT in auditing
1. GPT application logic framework
Li Yingqi, Director of the Auditing Department of Shanghai National Accounting Institute, mentioned the introduction of ChatGPT's auditing work, emphasizing the degree of prudence and the appropriateness of the introduction. All automatically generated reports must be manually reviewed and proofread to ensure the accuracy of relevant results. Accuracy and completeness. Chen Yangyang, director of the Institute of Accounting and Law at East China University of Political Science and Law, also believes that we should pay attention to the possible risks caused by the deviation of data and semantic understanding in the operation of ChatGPT. So what kind of application architecture system should be built to better introduce the value of GPT tools in a reasonable way? The following is from a consulting company on how to combine artificial intelligence tools such as ChatGPT with our service scenarios. The following is Classification according to the application architecture of Ai technology:
- Products use Ai as a certain capability or module, embedded in our products or services, to empower or enhance the experience of end customers.
- Process artificial intelligence technology can greatly optimize the daily workflow and operations at the structural level, upgrade the original operational level to the strategic level, and realize bottom-level automated operations, thereby improving productivity. Automation can be expressed in two forms, which can directly reduce manpower and increase the dimension of human work.
- Insight artificial intelligence can assist in making more informed strategic decisions. Based on the underlying more powerful data collection and processing capabilities, and continuously optimized intelligent decision-making models, it can provide companies with more intelligent, more targeted, and constructive suggestions.
2. Application scenarios in GPT audit
We raised the question of "ChatGPT's application scenarios in auditing work" again, and got the following responses:
3. Application examples and tool recommendations in GPT auditing
-
ChatExcel
official experience address: http://chatexcel.com/
In the audit work, many underlying matters are carried out around the excel table. In order to improve efficiency, I have learned a lot of programming and table formulas. With the arrival of ChatExcel, You will be surprised to find that all form processing operations can be as simple as dialogue, and any operation you want to achieve can complete the entire form processing process accurately and efficiently. Of course, we can also use ChatGPT to generate the vba code or python code we need at any time, and secondly, we can automatically realize the calculation and processing of some corresponding financial analysis models. -
ChatPDF
official address: https://www.chatpdf.com/
ChatPDF is an AI tool that can quickly extract useful information from PDF files and interpret the information through ChatGPT. In audit work, we often need to quickly analyze and understand a large number of audit texts, such as financial statements, business contracts, management systems, laws and regulations, etc. With ChatPDF, we can quickly upload the corresponding content in PDF file format , and then you can ask questions freely, and get answers within the range you specify most quickly.
GPT applies risk in audit
Although GPT will indeed bring us some huge changes in efficiency in actual audit work, I must also clearly recognize that there are many risk factors in at least the current application.
- Data Privacy and Information Security
When we use GPT tools to assist in auditing, the first thing we need to do is to upload data, and the corresponding data is often used as important financial data of the enterprise. Whether it is restricted by legal behavior or the confidentiality requirements of this part of the data, etc., we When using such tools, there is no guarantee that these tools can meet our basic requirements for data privacy and information security. The Samsung leak incident some time ago is a good proof.
- Inability to guarantee effective independence and professionalism
At this time, the bottom layer first involves that the effectiveness of the algorithm model depends on the size of the data volume, the parameters of the model, the size of the training volume, and the feedback in practice. In fact, when we use it, we cannot guarantee the overall system, especially Whether the learning ability of GPT meets our requirements, that is, to be able to correctly understand the meaning we want to express and the thinking mode, then in a longer-term process, it must be relatively inefficient and "stupid".
- AI can't handle complex systems for auditors
At this stage, GPT tools are still relatively defined as tool attributes, but due to the existence of "autonomous" consciousness, there will be "self-deception". If the auditors lose the ability of independent analysis if they rely too much on GPT tools, Even if it is the underlying job ability, it will greatly reduce the flexibility of people in dealing with complex scenarios.
- The regulatory and compliance system cannot break through the GPT black box
The existence of GPT intelligent applications has surpassed the construction of the previous human-centered work system, and there are often black-box attributes inside GPT, which cannot accurately understand the data processing and algorithm logic, even if the whole process is under human control, so Compliance with internal systems cannot be guaranteed.
There are bound to be risks in reform, and because of the existence of risks, it also limits the partial trial use of GPT tools in audits, and even in many cases, any conclusions cannot really be effectively combined with audit work. At a time when GPT-like tools are in full bloom and in full swing, on April 11, the Cyberspace Administration of China drafted the "Management Measures for Generative Artificial Intelligence Services (Draft for Comment)", which is now open to the public for comments. Everyone can make more suggestions.