In recent times, regulators have begun to explore the use of technology to help them perform their regulatory and supervisory functions. Known as RegTech (a contraction of the terms “regulatory” and “technology”) and also SupTech (a contraction of the terms ‘supervision’ and ‘technology”), innovation in this area includes the use of natural language processing (NLP) – a form of artificial intelligence – to facilitate and enhance the review of documents by regulators to assess compliance with disclosure requirements. There is a broad range of documents to which such technology might be applied, including corporate accounts, corporate announcements, company prospectuses and financial product disclosure documents.
Developments in RegTech have accompanied developments in FinTech (for a discussion about FinTech and smart contracts, see China Business Law Journal volume 7 issue 8: FinTech and smart contracts).
This column explores the potential that NLP offers in the area of corporate disclosure, and the legal and regulatory implications that arise as a result. These implications include the following: (1) whether technology will change the way in which the language of corporate disclosure and disclosure standards are interpreted by regulators; (2) whether regulators will be able to maintain transparency in relation to how technology is used to monitor and review corporate disclosure; and (3) how to maintain an appropriate degree of human involvement and guarantee trust in the process.
RegTech and its benefits. RegTech is increasingly being used by both firms and regulators for a range of purposes. The use of RegTech by regulators to assist them with their own functions has developed significant momentum over the past three years.
RegTech promises a range of benefits, including enabling firms to undertake compliance work in a more efficient and cost-effective manner, improving managerial discipline by improving the quality of information provided to boards of financial institutions and promoting good corporate governance generally. A range of technologies have been utilized for the purposes of RegTech, including artificial intelligence (AI), blockchain (distributed ledger technology), machine learning, NLP and data analytics.
In recent times, RegTech has been enthusiastically embraced by corporate and financial regulators, which have launched various initiatives and trials with a view to operationalizing technology for regulatory purposes.
Natural language processing (NPL). The digitization of data has enabled the development of technologies involving NLP. For our purposes, the term ‘natural language’ is synonymous with human language (i.e., language that has developed naturally) and is used to distinguish human language from other languages, such as artificial language or computer code. The term ‘processing’ refers to the various ways in which technology can be used to process natural language. This ranges from simple processing, such as tracking the frequency of words and phrases and extracting information for the purpose of comparison, to complex processing, such as interpreting the meaning of natural language.
NLP is a subset of machine learning, which is itself a subset of artificial intelligence. In recent times, there has been a move away from supervised or ‘logic-based’ machine learning, where an algorithm responds to rules and logic designed by humans, towards unsupervised learning, where the algorithm is programmed to learn directly from data and to identify patterns and does not need to identify the reasons as to why those patterns exist. Unsupervised learning is also referred to as ‘deep learning’.
Various methods are adopted for NPL. These include programming techniques, which extract words and phrases to define patterns in documents; topic modelling, which identifies topics and the relationship between topics; text categorization, which sorts texts into specific categories; text clustering, which groups text or documents based on similarities in content; and sentiment analysis, which interprets the meaning behind human language.
Among the NLP methods, topic modelling has begun to be used extensively by regulators such as the United States Securities and Exchange Commission (SEC) to assess the adequacy of corporate disclosure and to identify abnormal or fraudulent disclosure.
The digitization of data and the advent of technologies such as NLP mean that regulators can review and monitor a vastly greater amount of information than before in a fraction of the time required in the case of human review.
Legal and regulatory implications of RegTech. One of the potential implications of RegTech is that the use of technology will affect the interpretation of language. By way of example, let’s say that text analysis undertaken by NLP identifies a problem with the language of a prospectus and associates it with previous prospectuses that have been found to be misleading or deceptive. However, the problem subsequently turns out to be a false alarm. This is a particular risk with the ‘data-up’ approach, where the algorithm does not follow any rules but, instead, sorts through a large volume of historical data to make predictions based on patterns.
There may not be any adverse consequence if the results are examined and the false alarm is identified by an appropriately qualified human. However, even in these circumstances, there is a risk that the use of NLP technology will reinforce biases on the part of the humans who are interpreting the results.
A second potential implication is that RegTech will affect the way in which regulators interpret the disclosure requirements in respect of corporate disclosure documents. For example, the general disclosure test for prospectuses in Australia is set out in section 710 of the Corporations Act 2001 as follows:
710 Prospectus content–general disclosure test
(1) A prospectus for a body’s securities must contain all the information that investors and their professional advisers would reasonably require to make an informed assessment of the matters set out in the table below. The prospectus must contain this information:
(a) only to the extent to which it is reasonable for investors and their professional advisers to expect to find the information in the prospectus; and
(b) only if a person whose knowledge is relevant (see subsection (3)):
(i) actually knows the information; or
(ii) in the circumstances ought reasonably to have obtained the information by making enquiries.
The above test incorporates an objective element; namely, it refers to “all the information that investors and their professional advisers would reasonably require to make an informed assessment” and “only to the extent to which it is reasonable for investors and their professional advisers to expect to find the information in the prospectus.”
The concern is that NLP technology will affect or undermine the interpretation and application of objective disclosure standards that are tied to notions of “reasonableness” or the “reasonable investor”. This concern arises because of the potential for technology to expand the interpretation of language from one that is defined by reference to human parameters to one that incorporates algorithmic understanding.
What might be perceived as reasonable and complete from a human perspective might be perceived as unreasonable and incomplete from an algorithmic perspective. The specific concern is that questions of reasonableness will no longer be determined simply by humans (and human understanding) but by humans who are informed (and influenced) by technology (i.e., algorithmic understanding). Technology will thus redefine notions of the “reasonable person” and replace objectives standards with standards that are tied to the notion of the “reasonable algorithm” or the “reasonable computer”.
We have already started to see discussion about the possible reorientation of standards towards computer standards in the context of tort law, where humans are no longer judged against the standard of the hypothetical reasonable person but instead against the higher standards of computers. An example of this is the emergence of driverless cars, where liability for accidents is likely to be measured by reference to the standard of a “reasonable algorithm” instead of by reference to a human standard. Although this might be a good outcome in the context of driverless cars as it increases the standard of care on the part of those who invent them, it runs the risk of confusing the way in which corporate disclosure standards are interpreted and applied and it may also increase the potential difficulty of complying with them.
Another potential implication is that RegTech might make it difficult for regulators to maintain transparency in relation to the process by which they review corporate disclosure documents. This is particularly relevant in the case of unsupervised machine learning. This concern might be overcome by ensuring that there are qualified persons within the regulators who are familiar with the technology and can provide the necessary transparency from a technical perspective.
A fourth potential implication is the need to maintain an appropriate level of human involvement to guarantee trust in the process. The emergence of RegTech has led some people to speculate that human regulators will ultimately give way to robo-regulators or robo-cops. Such speculation, however, appears to be premature and misguided. The importance of maintaining human judgment in RegTech has been widely acknowledged. This is particularly relevant in circumstances where complex decision-making must be reserved to human decision-makers.
Some concluding thoughts
There appears to be little doubt that technology has huge potential to assist firms to comply with regulatory requirements and to assist regulators with their supervisory functions.
As is often the case, technological developments are accompanied by certain risks and challenges. In the case of the use of RegTech by regulators to monitor and review corporate disclosure, these risks and challenges include the impact of RegTech on the way in which language and disclosure standards are interpreted and the need to ensure regulatory transparency, particularly in circumstances where human judgment relies on, or is influenced by, algorithmic judgment. Technology can certainly assist in identifying fraudulent or abnormal disclosure, which is good. However, the use of technology to assess compliance with disclosure requirements may be problematic.
Looking forward, several questions are likely to require further research and analysis: do we need standards to ensure transparency in the use of RegTech and to ensure that natural language is preserved and clarity is maintained from a human perspective? Do we also need processes to ensure that there is sufficient human involvement in, and oversight of, RegTech to maintain trust and confidence in the process? If so, what might those standards or processes look like and how can we avoid unintended consequences?
In the area of corporate disclosure, few people would be unfamiliar with the famous statement of Louis Brandeis, the former US Supreme Court Justice: “Sunlight is said to be the best of disinfectants; electric light the most efficient policeman.”. As RegTech becomes more pervasive, one might well replace the reference to “electric light” with “artificial intelligence”. However, whether artificial intelligence proves to be an equally transparent and impartial policeman remains to be determined.
The above is based on a presentation delivered by the author at a seminar entitled “Trends and Challenges in Corporate Law – A Comparative Perspective”, held in Singapore on 23-24 May 2019 and hosted by Singapore Management University and Melbourne Law School.
A former partner of Linklaters Shanghai, Andrew Godwin teaches law at Melbourne Law School in Australia, where he is an associate director of its Asian Law Centre. Andrew’s new book is a compilation of China Business Law Journal’s popular Lexicon series, entitled China Lexicon: Defining and translating legal terms. The book is published by Vantage Asia and available at www.vantageasia.com.