Artificial Intelligence and Article 33.4 VCLT
By Tarcisio Gazzini
Published on 14 March 2024
Introduction
Artificial intelligence (AI) is set to radically change legal education and the legal profession. Suffice it to mention the administration of justice through machines. This comment focuses on a much more specific issue, namely the interpretation of multilingual treaties. Article 33.4 of the Vienna Convention on the Law of Treaties, in particular, directs the interpreter in overcoming differences between equally authoritative texts. It is unanimously considered as unsatisfactory (see Tabory, 213; Gardiner, 419) and its application remains fraught with difficulties, especially when authentic texts are numerous. AI-driven tools could unleash its potential to spectacularly enhance the performance of interpreters such as tribunals, governments and civil servants.
Interpretation of multilingual treaties
Interpretation is a complex process involving “the deployment of analytical and other skills: it cannot be reduced to a few propositions capable of purely automatic application in all circumstances” (see Sinclair, 153). The process is even more challenging when treaties are concluded in multiple languages since “[f]ew plurilingual treaties containing more than one or two articles are without some discrepancy between the texts”, possibly due to “the different genius of the languages, the absence of a complete consensus ad idem, or lack of sufficient time to co-ordinate” (see International Law Commission, 225).
Considering all authentic texts of a multilingual treaty is not indispensable (see Waldock, 100), although it would be the prudent course (see Gardiner, 419). It becomes arduous if not impossible when the treaty has been concluded in six equally authentic languages—as within the United Nations (UN) —or even more than 20—as most of the trade agreements concluded by the European Union (EU). Attempts by Tribunals to rely on dictionaries to establish the meaning of treaty provisions have been rather frustrating, if not counterproductive. Indeed, it has been observed that “[t]he cult of the dictionary in treaty interpretation leads to the erosion of settled meanings for international legal concepts and, instead, fixates upon the lowest common denominator of meaning generated by a sterile linguistic analysis of the treaty terms” (see Douglas, 101).
Yet, a multilingual treaty remains a single legal instrument with a “single set of terms” and the interpreter must search for its single meaning (see International Law Commission, 225). Its interpretation is governed by Article 33 according to which the terms of the treaty are presumed to have the same meaning in each authentic text. The contracting parties may agree that in case of divergence between the different authentic texts one of them would prevail. Otherwise, in accordance with Article 33.4, each authentic text has the same legal value. And when their comparison discloses a difference of meaning which the application of Articles 31 and 32 does not remove, the interpreter must adopt the meaning which best reconciles the texts, having regard to the object and purpose of the treaty.
The reconciliation of the texts does not mean selecting one or several languages deemed to express correctly the meaning of the text, but rather extracting from the different texts “the best reconciliation of the differences” (see Gardiner, 442-3). In so doing, the interpreter should understand and compare the different authentic texts and respect their different genius.
Arguably, the methodological approach adopted by some tribunals has been hardly consistent with Article 33.4. In Slovakia v. Council and Hungary v. Council, for instance, the Court of Justice of the EU developed its legal argument essentially from the perspective of the English language. It translated into English the terms used in other languages and eventually relied on the division of the (then) 15 authentic languages in two English terminological alternatives, which eventually it considered in good substance as equivalent (see Pietrobon and Gazzini)
The reconciliation required under Article 33.4 implies an adequate command by the interpreter of the different authentic languages. The proficiency of the members of tribunals, however, cannot be taken for granted. In Kiliç v. Turkmenistan, for instance, none of the arbitrators could speak the authentic languages of the relevant treaty and under the circumstances the tribunal took the very questionable decision of relying on non-official English translations of the treaty.
Artificial intelligence and the interpretation of multilingual treaties
AI may be expected to be a game changer in the application of Article 33 and in particular Article 33.4. It provides an arsenal of valuable tools that may assist the interpreter. Natural language processing (NLP) tools may particularly suitable. According to IBM, NLP “refers to the branch of computer science—and more specifically, the branch of artificial intelligence or AI—concerned with giving computers the ability to understand text and spoken words in much the same way human beings can. [NLP] combines computational linguistics—rule-based modelling of human language—with statistical, machine learning, and deep learning models”.
Equally promising is the development of increasing sophisticated large language models (LLMs). Through a training process of immense proportions (in billions of pages), LLMs learn grammar, semantics and conceptual relationships through zero-shot and self-supervised learning. Although currently mostly employed to generate texts, there is no reason why LLMs could not be resorted to for the purpose of interpreting multilingual treaties (see Wright Nelson).
AI-driven tools can therefore analyse and compare the different texts, determine the meaning of their provisions, and consider the context and scope of the treaty. The process may be facilitated by NLP such as GPT-3, which uses AI and statistic to predict the next word in a provision on the basis of preceding words.
Consider, for instance, the Free Trade Agreement between the EU and its members and Korea, which is made up of 273 articles plus several side-documents and annexes, and has been concluded in 23 equally authentic languages, including Korean. Analysing the relevant parts of the text and, if appropriate, the supplementary means, in all authentic languages for the purpose of Articles 31 and 32 without AI-driven tools would be a titanic—if not an impossible—enterprise. AI may be indispensable also to systematically examine the subsequent practice of all the parties, including official documents and judicial decisions. This has the great advantage of connecting the analysis of the text of the treaty to its practical application, rather than limiting the examination of the text to linguistic issues as is the case with the use of dictionaries. From the standpoint of Article 33.4, furthermore, AI can meaningfully assist in seeking a true reconciliation of linguistic differences, if any. This presupposes a thorough command of the 23 languages and a deep knowledge of their different genius. During the process, finally, AI can facilitate the correction of grammatical and lexical mistakes.
The use of AI must not be reduced to a mechanical exercise. Interpreters (judges, civil servants, parliaments, etc.), may engage in a process where they resort to AI-driven tools to progressively establish the meaning of the treaty following the logical sequence of Articles 31 to 33. With regard specifically to Article 33, they can ask those tools to interpret the treaty in each authentic language, then to detect any possible discrepancies, to identify and group the different plausible interpretations, and finally to attempt a reconciliation in accordance with Article 33.4. With high numbers of authentic languages—as in the example above—these tasks are simply beyond human capacity.
While the reliability and accuracy of AI improve every day, it would be clearly premature and indeed dangerous to blindly entrust them with the whole interpretative process. At least for the time being, interpreters must continue to play an active role and exercise their critical skills over such a process. From this perspective, AI-driven tools can provide a preliminary interpretation of the treaty based on a volume of information that only machines can handle, especially when the authentic texts are numerous, and the treaty made of hundreds of provisions.
Interpreters then intervene by double checking the interpretation offered by AI-driven tools in order to avoid any absurd, unreasonable or obscure outcome. If necessary, they make choices between different plausible interpretations by using critical thinking that still belongs exclusively to humans. Both functions will require the development of the skills needed to detect biases in the AI-driven interpretation and, if necessary, to complete the interpretation process facilitated by AI-driven tools.
Conclusions
The impact of AI on legal education and the legal profession will take a multitude of forms and provoke a quantum leap. From the standpoint of Article 33.4, such impact is already palpable due to technological progress, especially in NLP and LLMs. If the application of Article 33.4 is extremely arduous for human beings, especially with large numbers of equally authentic texts, AI will be able to deliver a much more efficient, accurate and rapid job. The advantages in terms of legal accuracy, predictability, and certainty as well as optimization of resources and reduction of costs are evident. So are the risks and the importance of continuing supervision by interpreters, who must learn how to ensure the appropriate use of AI-driven tools, detect biases and discharge unacceptable interpretations.
One last point relates to how AI may also prevent or minimise the risk of discrepancies between different authentic texts that may later trigger the application of Article 33.4. With regard to existing treaties, AI may facilitate the adoption of joint interpretative statements by the parties to the treaty—expressing the authentic interpretation of any of its provisions. From this perspective, AI could detect errors, ambiguities and potential frictions without waiting for the interpreter to be confronted with the application of Article 33.4. Perhaps even more importantly, AI will be a formidable companion for those involved in the negotiations of new treaties. From this perspective, AI-driven tools may be developed to enhance the coherence and accuracy of all texts of the treaty.
Tarcisio Gazzini is Professor of International Law at the University of Padua. He has previously taught at the Universities of Glasgow, VU Amsterdam and East Anglia. He is a founding editor of the book series International Investment Law, published by Brill, and of the Journal du droit transnational, published by Editoriale Scientifica (online)