Report on a session at the 14th Annual IBA Finance & Capital Markets Tax Conference held in London on 21 January 2025

Session Chair

Niklas Schmidt, Wolf Theiss, Vienna

Panelists

Jorge Correa Cervera, Creel García-Cuéllar Aiza y Enríquez, Mexico City
Idoya Fernández, Cuatrecasas, Madrid
Nils Harbeke, PrimeTax, Zürich
Pallav Raghuvanshi, Greenberg Traurig, New York
Caitlin R Tharp, Steptoe, Washington, DC

Reporter

Andrew Haikal, Fasken Martineau DuMoulin, Montréal

Introduction

Niklas Schmidt opened the session by introducing the panel and presenting the agenda. The panel discussed the future of artificial intelligence (AI) and its role in the legal profession, the increased use of Big Data by tax administrations and the taxation of cryptocurrency.

Panel discussion

The future of AI

Caitlin R Tharp provided an overview of generative AI, which she presented as a new development in the world of AI. Tharp highlighted the principal differences between AI, which is already a part of and very present in our day-to-day life, and generative AI, which is now starting to impact the practice of the law. Generative AI is different as it is able to generate new content. It is not a research tool of itself, but it can make predictions based on patterns. Its main flaw is that it incorporates data that may be incorrect.

Tharp used ChatGPT as an example of generative AI and considered whether taxpayers and tax professionals can rely on it for advice or in regard to their practice. ChatGPT states that it should not be substituted for qualified and licensed legal tax advice. Tharp gave the example of a legal question posed to ChatGPT, which was answered by referring to case law that does not exist. Tharp also gave the example of an attorney in the case of Mata v Avianca Inc. No. 22-cv-1461 (S.D.N.Y.), who relied on similar non-existent case law, highlighting the importance of verifying the accuracy of the information provided by ChatGPT. Schmidt interjected that generative AI should be viewed as a very creative junior, prompting Tharp to highlight the ethical duties described in American Bar Association Model Rule 1.6, including the importance of not sharing confidential information with tools such as ChatGPT. Pallav Raghuvanshi intervened to explain that law firms such as his are now dealing with this issue by inserting a paragraph on the use of AI in engagement letters, or by obtaining specific licences to use AI outside of the public domain.

Idoya Fernández discussed her experience with AI at Cuatrecasas. The firm is seeing a very high adoption rate for its internal AI tools, with approximately 80 per cent of its lawyers using them, three to four times per week, in addition to having a dedicated team of expert users experimenting with such tools, and internally teaching and explaining the findings provided by these tools. In Fernández’s view, such training is essential in order to increase the tools’ accuracy, through the proper integration of public data and internal knowledge. Notably, Cuatrecasas has its own research tool, which integrates all of the company’s standard internal models, as well as substantial tax information. Its lawyers are able to ask questions and to obtain answers which do not contain any AI hallucinations, as they are based on the firm’s internal database (which is regularly updated).

Jorge Correa asked how Cuatrecasas ensures that it does not share client information and that its tools do not share such information with anybody else. Fernández answered that the firm’s internal models are revised so that they do not contain any sensitive information.

Fernández stated that generative AI is evolving quickly and that the best way to get involved is to find the best tools and providers, and to test them in terms of the quality of their output and data security. It is also important to implement internal procedures in regard to the use of such tools and to make everyone aware that all output needs to be reviewed by an expert lawyer.

Schmidt then provided a quick explanation of how typical large language models work. By way of an analogy, Schmidt explained that these models are like sophisticated parrots, in that they act as statistical next-word predictors. Schmidt emphasised the three factors that determine these models’ performance: (1) the training size (ie, the amount of data used by them), (2) the computing power needed (ie, the amount of computing resources used) and (3) the model size (ie, the size of the artificial brain). Schmidt provided a few examples of popular large language models and noted that they are mostly proprietary systems (ie, users pay to use them, and they cannot be downloaded onto a user’s own computer). Schmidt then provided a list of examples of ways that ChatGPT can be used by lawyers, as well as a list of pros and cons related to such use.

Big Data

Raghuvanshi started by explaining that Big Data is a vast and complex dataset that is challenging to process, characterised by an intense generation speed and which comes in various formats. Raghuvanshi highlighted the main uses of Big Data, namely for analytics, machine learning and personalisation purposes.

Nils Harbeke raised a question about whether anything is different with regard to Big Data compared to an earlier discussion that took place in the context of the Organisation for Economic Co-operation and Development’s base erosion and profit shifting project on the ‘digital economy’. Harbeke noted that when the digital economy discussion started, it was more focused on nexus issues. Now, the focus seems to be more on where the value is actually created when Big Data is used. In other words, the tax analysis issue seems to have shifted towards allocation and attribution issues, rather than nexus issues. A key phrase in this regard is ‘data as an asset’.

Raghuvanshi then provided some thoughts on Big Data as an asset. He queried whether the costs incurred with respect to Big Data can be capitalised or expensed for tax purposes. The answer is likely dependent on its use (ie, the achievement of a long-term benefit through the purchase of proprietary datasets versus day-to-day operations).

Raghuvanshi also touched on the tax treatment of revenue generated through Big Data and noted that the United States Internal Revenue Service recently released final regulations pursuant to which cloud and other digital transactions would be viewed as giving rise to services income (as opposed to royalties). Proposed regulations also aim to address how to source such services.

Harbeke concluded by addressing recommendation agents. He explained that they are built to notice patterns in order to make recommendations, which has the potential to become relevant for tax enforcement, as recommendations help tax audit departments focus their resources more efficiently, for example.

Chaotic crypto taxation

Jorge Correa briefly touched on tokenised assets and decentralised autonomous organisations (DAOs). He noted that Mexico has not enacted any rules to deal with these matters.

He explained that DAOs are akin to a group of people or organisations that operate under a blockchain infrastructure to carry on their activities. Often, a legal wrapper is used, so no new tax issues arise as relates to the characterisation of the entity. However, without a legal wrapper, there is currently no answer as to who the relevant taxpayer is in this context (ie, the DAO itself or its investors), but international models for DAOs have been increasingly characterising them as partnerships (ie, flow-through entities).

Correa mentioned that tokenised assets are subject to more regulation, depending on the jurisdiction concerned. The question is whether they should be treated independently from, or the same as, their underlying assets.