“The world-altering powers that technology has delivered into our hands now require a degree of consideration and foresight that has never before been asked of us.”— Carl Sagan, Pale Blue Dot: A Vision of the Human Future in Space
Since the introduction of ChatGPT in November 2022, artificial intelligence (AI) has taken the legal and accounting professions by storm. The use of AI in legal proceedings has already made headlines (“Lawyer’s AI Blunder Shows Perils of ChatGPT in ‘Early Days,’” Bloomberg Law, May 31, 2023, https://bit.ly/3KkdsvM) and research now suggests that most legal and tax professionals believe ChatGPT can be applied to legal or tax work. [A series of surveys of more than 1,800 legal and tax professionals in the U.S., U.K., and Canada conducted between March and May 2023 found that 82% of legal professionals and 73% of tax professionals believe ChatGPT can be applied to legal or tax work. (Future of Professionals: ChatGPT and Generative AI in Legal, Corporate & Tax Markets, Thomson Reuters, 2023).] Although the revolutionary potential of AI is obvious, little has been said about the technology’s implications for professionals bound by standards like the American Bar Association’s Model Rules of Professional Conduct (MRPC) and the Treasury Department’s Circular 230.
The MRPC and Circular 230 both require professional and ethical judgment, but Circular 230 appears to prohibit the use of ChatGPT when tax advisors are providing “written advice,” creating difficulties for tax accounting firms and professionals seeking to streamline costs or provide quicker service. Given this obstacle, and numerous others, tax professionals should tread cautiously and treat ChatGPT as a small piece of the puzzle rather than a cure-all.
What is ChatGPT and How Does It Work?
ChatGPT is a type of Generative Pre-trained Transformer (GPT) that was introduced by OpenAI, an American artificial intelligence organization, on November 30, 2022. The latest model of ChatGPT (GPT-4) was released on March 14, 2023, and is available to paid subscribers on a limited basis.
GPTs are trained on large datasets of unlabeled text and can generate novel human-like content in many formats, including images, text, and audio. GPTs are inherently “generative,” meaning that they expand on their initial training by learning from user interactions. Despite its novelty and adaptive nature, ChatGPT has drawn criticism due to confidentiality concerns and inaccuracies resulting from “AI hallucination” (described below). These issues demand a critical discussion on the ethics of incorporating ChatGPT (and many other forms of generative AI) into one’s professional work.
In addition to confidentiality concerns, ChatGPT tends to hallucinate. “AI hallucination” occurs when an AI system is unable to properly interpret data and, as a result, generates inaccurate or unusual outputs. AI hallucinations can take many shapes, from reports of fake news to false documents or assertions about people, historical events, or scientifically proven facts. For example, a program like ChatGPT can create a fake 20th-century jazz saxophonist with a full biography and discography that is entirely fictitious. To make matters worse, these inaccuracies are often communicated convincingly and in misleading ways. In addition to hallucination, the most recent version of ChatGPT (GPT-4, the version currently available to the public) generally lacks knowledge of events that occurred after September 2021. OpenAI acknowledges these deficiencies in a recent technical report (GPT-4 Technical Report, OpenAI, p. 10, Mar. 27, 2023, https://arxiv.org/pdf/2303.08774.pdf).
ChatGPT’s confidentiality and accuracy issues are significant and have direct consequences under the ethical rules of the MRPC and Circular 230.
ChatGPT and the MRPC
Rule 1.1 of the MRPC requires an attorney to provide “competent representation” by possessing the legal knowledge, skill, thoroughness, and preparation reasonably necessary for a given legal matter. Comments to this rule highlight two immediate issues for attorneys seeking to leverage ChatGPT.
First, to competently handle a matter, an attorney must use “methods and procedures meeting the standards of competent practitioners” (MRPC, r. 1.1 cmt. 5). This determination largely depends on “what is at stake,” with “major litigation and complex transactions” requiring more extensive treatment. Even in complex cases where the use of ChatGPT to manage large data sets or files might be justified, the ABA’s comments could not have envisioned the emergence of AI or its tendency to produce inaccurate results. [Apart from Resolution 112, which the ABA passed in August 2019 to urge the legal industry to grapple with “emerging ethical and legal issues related to the usage of artificial intelligence (‘AI’) in the practice of law,” the most recent amendment to the MRPC occurred in August 2020, long before the public introduction of generative AI.] OpenAI also admits that ChatGPT is “not fully reliable” and should be used with “great care … particularly in high-stakes contexts” (GPT-4 Technical Report, OpenAI, p. 10, Mar. 27, 2023, https://arxiv.org/pdf/2303.08774.pdf). ChatGPT in its current form arguably does not comport with the MRPC’s standards, even in the most complex of cases.
Second, attorneys must “maintain competence” by keeping abreast of changes in the law and its practice, “including the benefits and risks associated with relevant technology” (MRPC, r. 1.1 cmt. 8). For all attorneys (not just those actively using ChatGPT), the MRPC explicitly requires an active awareness of developments relating to generative AI and its practicality under a variety of circumstances. As the technology develops, attorneys ought to determine when the use of ChatGPT and other generative AI models may be appropriate for their legal services. This will require awareness of any efforts to improve the confidentiality and accuracy issues inherent in ChatGPT’s work product, as well as the status of ongoing investigations into the practices of AI development companies such as OpenAI.
Rule 1.2 requires an attorney to abide by a client’s decisions concerning the objectives of the representation and “consult with the client as to the means by which [the objectives] are to be pursued.” Most clients are likely unaware of ChatGPT’s capabilities in legal settings, so it is important for any attorney using the technology to obtain informed client consent by explaining the technology’s strengths and weaknesses, including the possibility of confidentiality and accuracy issues. Relatedly, Rule 1.4 requires attorneys to “consult and explain” when communicating with clients about legal matters, so that the client has “sufficient information to participate intelligently in decisions concerning the objectives of the representation and the means by which they are to be pursued” (MRPC, r. 1.4 cmt. 5). As with Rule 1.2, when using ChatGPT or any other form of generative AI, an attorney ought to fully explain the technology to allow the client to provide informed consent to its use.
The use of ChatGPT presents confidentiality issues for professionals who maintain sensitive client information.
Rule 1.6 imposes arguably the most prohibitive requirement concerning the use of generative AI in legal settings—confidentiality. Specifically, an attorney must not reveal information relating to the representation of a client and must “make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client” [MRPC, r. 1.6(a), (c)]. As explained above, ChatGPT and other generative AI models “learn” from user interactions by taking whatever information users provide and using it to influence future outputs. If an attorney or an attorney’s assistant inputs sensitive or confidential client information into a generative AI model, the attorney likely violates Rule 1.6 and the requirement to “make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client.”
ChatGPT and Circular 230
Although the MRPC raises important considerations for attorneys seeking to use ChatGPT, Circular 230 imposes additional and arguably more stringent requirements on persons who “practice before the Internal Revenue Service,” including “attorneys, certified public accountants, enrolled agents, enrolled retirement plan agents, registered tax return preparers, and other persons representing taxpayers before the IRS” [Circular 230, section 10.30(a)].
Section 10.22 of Circular 230 generally requires practitioners to exercise “diligence as to accuracy” in preparing or assisting in the preparation of, approving, and filing tax returns, documents, affidavits, and other papers relating to IRS matters, as well as in determining the correctness of oral or written representations made by the practitioner to the Treasury Department [Circular 230, section 10.22(a)]. For these purposes, a tax practitioner is presumed “diligent” if they rely on the work product of another person and use reasonable care in engaging, supervising, training, and evaluating the person [Circular 230, section 10.22(b)]. Setting aside the question of whether ChatGPT could or should be considered a “person” under this rule, it is technically impossible for a tax practitioner to train or supervise a large language model like ChatGPT, which runs on a mix of past training data provided by a development team and subsequent information provided by users. Additionally, in many circumstances it may be impractical to expect the typical tax practitioner to detect nuanced biases or inaccuracies in the results of ChatGPT’s output, because AI hallucinations and other inaccuracies are typically presented convincingly and without obvious warning.
Circular 230 also imposes significant requirements on tax practitioners when they provide “written advice” (including by electronic means) concerning federal tax matters. (Notably, written advice does not include government submissions on matters of general policy or continuing education presentations provided solely to enhance an audience’s federal tax knowledge.) For any tax practitioner providing written advice, section 10.37 requires that they (i) base the written advice on reasonable, factual, and legal assumptions (including as to future events); (ii) reasonably consider all relevant facts and circumstances that the practitioner knows or should know; (iii) use reasonable efforts to identify and ascertain the facts relevant to written advice on the subject matter of the advice; and (iv) not rely upon representations, statements, findings, or agreements (including projections, financial forecasts, or appraisals) of the taxpayer or any other person if reliance on them would be unreasonable [Circular 230, section 10.37(a)(2)].
Given these requirements, it seems practically impossible for a tax practitioner to rely on ChatGPT to form the basis of written advice in compliance with Circular 230. With ChatGPT’s known tendency to hallucinate and generate inaccurate responses with confidence, any written advice provided partly using ChatGPT would presumably be based on unreasonable factual or legal assumptions—OpenAI itself acknowledges that ChatGPT is “not fully reliable” (GPT-4 Technical Report, OpenAI, p. 10, Mar. 27, 2023, https://arxiv.org/pdf/2303.08774.pdf). Written advice generated using ChatGPT would also fail to consider “all relevant facts and circumstances known to the practitioner,” because ChatGPT does not consider taxpayer-specific circumstances or events post-September 2021 (presumably including routine, daily IRS guidance). ChatGPT has also been known to respond differently to identical prompts, raising additional concerns about accuracy and reliability (“Four Tax Questions for ChatGPT and Other Language Models,” Tax Notes Federal, May 8, 2023, https://bit.ly/3rQK7Td). For these reasons, providing written advice based even partly on representations, statements, or findings of ChatGPT would violate the requirements of Circular 230.
Any tax practitioner considering using technology like ChatGPT in their practice should consider Circular 230’s significant and largely prohibitive requirements. Until OpenAI has demonstrated improvements in reliability and accuracy, tax practitioners would do well to avoid relying on ChatGPT for written advice or other high-stakes matters, and instead consider using the technology for other purposes, such as analyzing large data sets or conducting background research using open-ended inquiries with no clear correct or incorrect answer.
Beyond the immediate obstacles caused by the interaction of AI and current professional ethical standards, attorneys and tax professionals should also pay careful attention to ongoing investigations and lawsuits concerning the development and use of AI. The Federal Trade Commission (FTC) recently opened an investigation to determine whether OpenAI violated consumer protection laws while developing its AI models (including ChatGPT) by engaging in “unfair or deceptive privacy or data security practices” or engaging in “unfair or deceptive practices relating to risks of harm to consumers, including reputational harm” (FTC Civil Investigative Demand Schedule, FTC File No. 232-3044, https://wapo.st/3OwQ40y). Several AI companies have also been the subject of recent class action lawsuits alleging the use of proprietary material in the training and development of AI models. For example, Microsoft, OpenAI, and GitHub are currently facing a class action lawsuit claiming that an AI-powered coding tool named “Copilot” committed “software piracy” because it is trained on existing programming code gathered from public websites. These actions could impair the development and proliferation of AI in its early stages, but they may also compel improvements that make the technology a viable practical tool for attorneys and tax professionals.
In addition to the aforementioned investigative and legal actions, the ABA issued Resolution 604 earlier this year to urge developers and users of AI to adhere to three general principles: 1) developers ofAll Users AI (e.g., integrators, suppliers, operators) should ensure that their products and services are subject to human authority, oversight, and control; 2) responsible individuals and organizations should be accountable for the consequences caused by their use of AI, unless they have taken reasonable measures to mitigate against harm or injury; and 3) developers should ensure the transparency and traceability of their AI products and services, while protecting associated intellectual property, by documenting key decisions made with regard to the design and risk of the data sets, procedures, and outcomes underlying their AI products and services [ABA, “Resolution 604” (Feb. 6, 2023)].
As noted above, the MRPC requires attorneys to “maintain competence” by staying informed of changes in the law and its practice, “including the benefits and risks associated with relevant technology” (MRPC, r. 1.1 cmt. 8). Tracking investigations and legal proceedings involving AI to determine whether the technology can become feasible for professional use would align with the spirit of this requirement.
As a final point, any attorney or tax professional who desires to use ChatGPT should be mindful of how one accesses the technology. Like other popular websites, numerous imitations have emerged that replicate ChatGPT’s user interface and use spoofed URLs to trick users into providing private information or downloading malware. To access OpenAI’s ChatGPT (within ethical boundaries), use the following link: https://openai.com/blog/chatgpt.
Any tax practitioner considering using technology like ChatGPT in their practice should consider Circular 230’s significant and largely prohibitive requirements.
ChatGPT is a Piece of the Puzzle, Not a Cure-All
ChatGPT (like all generative AI) is a powerful tool with the potential to streamline the work of attorneys and CPAs, but the technology must be reconciled with existing ethical frameworks. The numerous ethical and accuracy related issues raised by using ChatGPT strongly suggest that, at least for now, professionals should treat ChatGPT as a small piece of the puzzle, not a cure-all.