As technology continues to advance, artificial intelligence (“AI”) is becoming more prevalent in assisting with a wide range of tasks, including those related to tax and legal matters. One particular AI tool is ChatGPT, a natural language processing model that can generate human-like responses to text prompts. ChatGPT can be a resource in tax and legal matters, providing assistance with research, drafting, and translation of legal documents, and offering answers to common tax and legal questions. In this article, we will explore the current capabilities of ChatGPT in relation to tax matters and answer the question of whether the AI tool has the merits to replace the role of tax advisers and lawyers.
ChatGPT is a large language model developed by OpenAI that uses deep learning techniques to generate natural language responses to text-based prompts. It is based on the “GPT” (Generative Pre-trained Transformer) architecture and refers to a type of neural network architecture that uses self-supervised learning to acquire knowledge from large amounts of text data. Having been trained on a massive amount of text data, the tool is capable of engaging in conversations on a wide range of topics, from general knowledge to specialised domains. With its ability to understand context, infer meaning, and generate coherent responses, ChatGPT represents a significant step forward in natural language processing and has the potential to revolutionise the way we interact with computers and machines.
ChatGPT is part of a family of large language models, starting with the release of GPT-1 in 2018. It was followed by GPT-2, which was released in 2019 and made headlines for its ability to generate highly coherent and natural-sounding text. GPT-2 was so good at generating text that OpenAI initially decided not to release the full version of the model, out of concern that it could be used to generate malicious content such as fake news or spam.
In 2020, OpenAI released GPT-3, the most advanced and largest language model to date. GPT-3 was trained on an unprecedented amount of text data and was able to perform a wide range of natural language tasks with remarkable accuracy, including translation, summarisation, and even programming. ChatGPT specifically refers to a variant of GPT-3.
While ChatGPT is a remarkable achievement in natural language processing, it still has some limitations that should be kept in mind when using it:
Limited knowledge and understanding: Although ChatGPT has been trained on a vast amount of text data, it still has limitations in terms of knowledge and understanding. ChatGPT may not have a deep understanding of certain concepts or topics, and may produce responses that are incomplete, incorrect, or even offensive.
Lack of common sense: ChatGPT lacks common sense, which can lead to errors and misunderstandings in conversation. For example, ChatGPT may not understand the context of a joke, or may provide irrelevant or nonsensical responses to certain prompts.
Biases: ChatGPT may reflect biases present in the data it was trained on, which can lead to problematic or discriminatory responses. For example, ChatGPT may produce sexist, racist, or otherwise offensive responses.
Overreliance on training data: ChatGPT's performance is heavily dependent on the quality and diversity of the training data it was exposed to. If the training data is incomplete, biased, or unrepresentative of the target user population, ChatGPT may not perform well.
Ethical concerns: As with any advanced technology, there are ethical concerns associated with the use of ChatGPT. For example, ChatGPT could be used to generate malicious or harmful content, such as fake news or hate speech, or could be used to impersonate or deceive people.
Time limits and cutoff date: ChatGPT's time limits with regards to information are related to the cut-off date for the training data it was trained on, beyond which new data was not included. The cutoff date for ChatGPT's training data is September 2021.
As a result, ChatGPT may not be up-to-date on certain topics or events that have occurred after September 2021. Additionally, if the conversation involves a topic that has rapidly evolved or changed since that time, ChatGPT may not have the most current information and may provide outdated or inaccurate responses.
Google and ChatGPT are quite different in their purpose and functionality. Google is a search engine that helps users find information on the internet, while ChatGPT is an artificial intelligence language model that can engage in conversation with users and generate text in response to prompts.
Google is designed to help users find relevant websites and other online resources related to their search queries. Users typically enter a search term into Google's search bar and receive a list of results, along with ads, based on the search algorithm's assessment of the relevance and authority of the content.
On the other hand, ChatGPT is designed to engage in natural language conversations with users, generating text in response to their prompts.
While Google and ChatGPT both involve generating text, their underlying algorithms and purposes are quite different. Google's search algorithm is based on analysing website content and linking patterns to determine relevance, while ChatGPT is based on neural networks trained on vast amounts of text data to generate human-like responses.
The author undertook the task to research the tax capabilities of ChatGPT in relation to the Bulgarian tax legal framework. This section provides real examples of questions that the AI tool has been asked and after each example an assessment is provided as to the accuracy of these answers.
When asked what some techniques are for optimising tax liability in Bulgaria, the AI-generated tool was able to provide this answer in roughly 15 seconds:
This question was asked in the most general manner possible on purpose – to determine whether ChatGPT would be able to differentiate between corporate income tax liability and personal income tax liability. To that end, the answers provided are roughly accurate but no differentiation is drawn with respect to corporate and personal tax liability. The answers are rather mixed and as to business expenses, no account is taken as to the nature of the business and the type of business expense one can deduct from corporate tax base.
In addition, although roughly accurate, the answers presented under this question are common for the majority of jurisdictions levying corporate income tax and to that end the author does not agree that ChatGPT takes concrete account and data based on Bulgarian sources but rather assesses on a high level manner what would be sound and accurate answer without exposing the AI to major “mistakes”.
The answer to this question took less than 15 seconds:
Evidently, the answer is wrong. Despite the cutoff date being in September 2021, corporate income tax rates have not changed much. The answer according to the AI tool is Bulgaria with a rate of 10% and then continues to list other jurisdictions among which is the accurate answer, i.e. Hungary with a flat rate of 9%.
Although the author is not acquainted with the underlying models of ChatGPT and how the latter bases its answers, one could think that the tool is biased and provides answers with preference to the location from which the tool is used.
Similar to the first question asked, this inquiry was purposefully phrased in a general manner without differentiation between a tax residency of a legal and a natural person. The answer provided relates only to the determination of a tax residency of a natural person which is inaccurate and misleading.
When reading Article 4 of the said tax treaty, the term "resident of a Contracting State" means any person who, under the laws of that State, is liable to tax therein by reason of his domicile, residence, citizenship, place of management, place of incorporation, or any other criterion of a similar nature, and also includes that State and any political subdivision or local authority thereof. This term, however, does not include any person who is liable to tax in that State in respect only of income from sources in that State.
Only then, when tax residency cannot be determined by reference of the above criteria (i.e. domicile, residence, citizenship..etc.) one may rely on the tie-breaker rules.
As AI continues to advance, there is increasing speculation about whether AI models, such as ChatGPT, could replace tax advisers and lawyers. While these models are to a certain extent helpful tools for gaining general information and insights about taxes and laws, despite the time limitations, they cannot yet replace the expertise and tailored advice provided by trained professionals.
It is claimed that one of the most significant advantages of using ChatGPT is that it can process vast amounts of data in real-time, making it useful for providing instant answers to questions about taxes and laws.
Despite these benefits, ChatGPT and other AI models still have limitations that prevent them from replacing tax advisers and lawyers. As observed in the pervious section these limitations include:
Lack of specialised knowledge: While ChatGPT can provide general information about taxes and laws, it cannot match the depth and breadth of specialised knowledge that comes with years of training and experience. Tax advisers and lawyers are trained to analyse complex cases, interpret legal statutes and regulations, and provide tailored advice to their clients. This expertise cannot be easily replaced by an AI model.
Inability to provide personalised advice: Every tax situation is unique, and ChatGPT cannot provide the kind of personalized advice that a human tax adviser or lawyer can. Tax advisers and lawyers consider each client's specific circumstances and provide guidance based on these individual factors. ChatGPT, on the other hand, can only provide general information that may not be relevant to a particular case.
Lack of emotional intelligence: Tax advisers and lawyers also bring an emotional intelligence to their work that AI models cannot replicate. Clients often seek advice from professionals during times of stress or uncertainty, and the ability to empathise and communicate effectively is critical in these situations. While ChatGPT can simulate conversations with humans, it cannot offer the kind of emotional support that a human professional can.
The potential for errors: Although AI models are designed to be as accurate as possible, they are not infallible, as evidenced from the real examples in section 4. ChatGPT and other models can make mistakes, and the consequences of these mistakes can be severe in tax and legal matters. Trained professionals, on the other hand, can catch errors and make sure that their clients' interests are protected.
Lack of recent data: Given that the tax field is constantly changing it is of utmost importance of any tax adviser to keep up to date with any amendments of legislation. To that end, ChatGPT is not capable to be aligned with recent tax developments.
In conclusion, while ChatGPT and other AI models have many advantages, at present they cannot replace tax advisers and lawyers. These professionals offer specialised knowledge, personalised advice, emotional intelligence, and the ability to catch potential errors. While AI models can provide general information and assist in daily tasks such as draft of contracts, minutes or draft of content, their limitations prevent them from fully replacing human professionals. As such, it is still essential to consult with qualified professionals in these fields for specific advice and guidance.
*Disclaimer: Be advised that 90% of the present article was generated through ChatGPT.