Authored By: Akhilesh Kakade
High Court of Bombay
ABSTRACT
The emergence of artificial intelligence (AI) as a tool for arbitral proceedings is a revolutionary step forward from human-assisted tools to possible algorithmic adjudication. This paper raises the question of whether AI can meaningfully engage in arbitral decision-making without undermining basic principles of fairness, transparency and human accountability. Through doctrinal and comparative analysis, it explores institutional uses of AI in arbitration, such as guidelines of the International Criminal Court and the Court of International Litigation and recent experiments such as CaseCruncher Alpha and China’s “smart courts.” It also examines the legal and ethical issues of algorithmic adjudication – in particular bias, explainability and due process protections under the UNCITRAL Model Law and the Arbitration and Conciliation Act, 1996. Based on such regulatory developments as the EU Artificial Intelligence Act (2024), Model AI Governance Framework in Singapore (2020), and Responsible AI Strategy in India (2023), the paper assesses the global responses to algorithmic fairness in legal decision-making. The analysis concludes that while AI can help to improve the consistency and efficiency of arbitration, it will not replace the deliberative reasoning and moral judgment of human arbitrators. It proposes a hybrid adjudicative future that is based on three key safeguards: compulsory AI disclosure, human intervention in final determinations and institutional ethical codes with transparency and bias audits. Finally, the paper concludes that the legitimacy of arbitration in the age of smart machines will lie not in man against machine but in man with machine – where fairness and accountability remain the cardinal rules of digital justice.
Keywords: Artificial Intelligence, Arbitration, Algorithmic Fairness, Bias, Explainable AI, Digital Adjudication
INTRODUCTION
Artificial intelligence (“AI”) has evolved into a disruptive force in the legal decision-making process urging the transformation of how disputes are settled and how justice is perceived. AI has gone from document review or e-discovery, to being used in areas where outcomes that directly impact the arbitration and other alternative dispute resolution (ADR) processes are affected. Such a change has furthermore been accelerated by the worldwide thrust towards efficiency, cost reduction and predictability of legal processes through data-driven inputs. No longer a fanciful notion, the idea that in the future an algorithm could serve as a “digital arbitrator” that could sift through evidence and award an award becoming a reality is no longer a say-so but just a matter of time.
The idea of algorithmic adjudication raises complex questions on issues such as fairness, transparency and procedural justice. While AI offers unprecedented speed and analytical accuracy, it also raises issues related to bias, explainability, and accountability. Algorithms trained on data from the past run the risk of perpetuating systemic inequalities but also the lack of transparency in machine learning models creates tension with the need for arbitral awards to be reasoned and understood by the parties. The European Union’s Artificial Intelligence Act (2024) identifies technologies for legal decision-making as “high-risk,” which is indicative of growing apprehension that automation will erode due process safeguards. Scholars like Richard Susskind have warned against reducing justice to computation despite the inevitability of digital transformation.
This article critically analyzes whether AI should be allowed to play an active role in arbitral decision making without undermining the rule of essential principles of fairness and transparency. It examines recent institutional experiments, evaluates emerging regulatory frameworks in the European Union, Singapore and India and examines the question as to whether algorithmic adjudication can coexist with the human core of arbitration.
RISE OF AI IN ARBITRATION – FROM ASSISTANCE TO ADJUDICATION
Artificial intelligence has insidiously become an invaluable assistant in modern arbitration. Over the past decade, arbitral institutions and law firms have been using AI for document analysis, case management and data clustering – activities that have historically been extremely resource-intensive for humans. Predictive analytics now allow counsel to estimate award values, model tribunal behavior and estimate chances of settlement with amazing accuracy. AI-powered tools like Kira Systems, ROSS Intelligence, and Lexis Nexis Context have used unstructured case law to create searchable, pattern-driven insights which represents the first step in the introduction of AI into the resolution of disputes.
The second, and more transformative, stage is in progress: the use of AI not as an assistant but as an adjudicative aid. Experiments such as CaseCruncher Alpha have shown that algorithms using machine learning are able to accurately predict outcomes of claims for financial mis-selling, with higher accuracy than panels of qualified lawyers. Likewise, the much-publicized DoNotPay “robot lawyer” sought to help a litigant walk through traffic court arguing using an AI-generated script before being pulled after regulatory objections. Though these examples were few, they demonstrate an increasing faith in the ability of AI to think creatively within the confines of the law.
China has gone even further, by institutionalizing AI in its judicial system. The Supreme People’s Court of China launched a national AI platform to help judges write and manage judgments in the “smart courts,” signalling a state-backed shift towards algorithmic adjudication. While such systems are still in the hands of humans, they are indicative of the potential for AI assisted decision-making on a larger scale.
For arbitral institutions, the power of AI is efficiency and consistency. Automated tools are well equipped to handle large amounts of evidence, identify contract gaps and draft initial procedural orders. Yet, this shift towards automation raises a fundamental concern: is the need for greater efficiency a sacrifice to procedural fairness? The International Chamber of Commerce (ICC) and the London Court of International Arbitration (LCIA) have both recognized the potential of AI but stressed that any element that involves decision-making must be kept under human control. The move from human-assisted technology into technology-assisted humans is therefore the defining question for the next decade of arbitral evolution.
III. THE LEGAL AND ETHICAL CHALLENGES OF ALGORITHMIC ADJUDICATION
While the introduction of artificial intelligence (“AI”) into arbitration promises obvious gains in efficiency, it also challenges the very foundations of arbitral legitimacy – neutrality, procedural fairness and accountability. The law of arbitration assumes that an award is the product of a conscious human process of reasoning and not a computational output. Article 31 of the UNCITRAL Model Law on International Commercial Arbitration and Section 31 of the Arbitration and Conciliation Act, 1996 in India provides that every award must indicate the reasons on which it is based which is a fundamental principle of transparency by way of explanation. If an AI system were to render or substantially assist in forming an arbitral award, the lack of transparency of machine learning algorithms, the so-called black box problem, would make it practically impossible to comply with these provisions.
The EU Artificial Intelligence Act (2024) addresses this dilemma by enforcing that AI systems which are used in legal decision-making should be considered “high-risk” systems, thereby needing scrupulous human oversight and explainability mechanisms. While the Act’s procedural safeguards are progressive, they are not binding beyond the European Union and so leave cross-border enforcement under the New York Convention (1958) up in the air. The Convention assumes that an “arbitral tribunal” is manned by human arbitrators who are free to exercise independent judgment and are accountable. Whether an AI-assisted award would pass that test – or be denied enforcement under Article V for its violation of public policy – is an open doctrinal question.
Beyond legality, the ethical aspect is also of prime importance. Algorithmic adjudication is prone to inherit the prejudices of the algorithms and the data they are fed. Recent work on the COMPAS algorithm in US sentencing shows how it is easy for training data to replicate racial disparities in the data set, and similar distortions are likely to happen if AI trained on small arbitral data sets systematically privileged specific industries or jurisdictions. These problems assault at the core of equality of arms – a mark of arbitral fairness.
New solutions have a focus on Explainable AI (XAI) and obligatory human-in-command models. Both the Model AI Governance Framework (2020) of Singapore and the Responsible AI Strategy (2023) of India propose traceability, auditability and human accountability towards automated decisions. The ethical duty of disclosure is also very important: whenever AI tools affect procedural or substantive determinations, parties and tribunals need to know. In absence of such transparency, the effectiveness of the arbitration may be lost both among the users and the national courts.
ALGORITHMIC FAIRNESS AND BIAS IN AI-DRIVEN ADJUDICATION
The potential of artificial intelligence (or AI) in arbitration is its capacity to provide efficiency, consistency, and precision in its analysis. But even under that promise lies an inherent danger, algorithms, regardless of how objective they may seem, are only as impartial as the data and design used to produce them. Bias can be insinuated into an AI system in three main ways – data bias, design bias and outcome bias – each presenting unique challenges to arbitral fairness and due process.
Data bias is where certain parts of history that are reflected in the data used to train an algorithm are skewed or incomplete. If a contract-dispute model is trained on largely awards by Western arbitral institutions, it will misinterpret the practices that prevail in developing jurisdictions. Design bias is the result of the implicit assumptions of the developers (for example around weighting of variables or linguistic training inputs) which can reflect cultural or gender biases. Outcome bias occurs when algorithms maximize on such measurable objectives as speed or cost efficiency instead of substantive fairness. All these biases constitute a derailment of the two major characteristics of arbitration: equality of arms and tribunal independence.
The cautionary experience of the COMPAS algorithm in the United States is an example of the real-world implications of such bias. Intended to predict risk of criminal recidivism, for example, the system was later found to overestimate the risk score for Black defendants and underestimate it for white defendants. Although arbitration is not the same as criminal justice, the analogy holds because AI systems based on previous awards may incorporate implicit assumptions of prior tribunals or reproduce socioeconomic disparities that have been embedded in data sets. In arbitral practice, such distortions could be translated into systematically unequal treatment of certain industries, parties or legal traditions.
Recent controversies between generative-AI platforms like Lensa AI and ChatGPT are other examples of the fairness risks of generative AI: hallucination and unverified reasoning. Lensa’s portrait generation function was criticized for sexualizing female portraits and reproducing gender stereotypes, even highlighting the ways in which pattern matching algorithms can reproduce discriminatory tropes. Similarly, a court in New York fined lawyers who blindly supplied falsified references of cases generated by ChatGPT and revealed the unreliability of non-verified AI results in legal arguments. Transposed into the arbitral sphere, such errors could affect the validity of an award itself. An AI that imagines precedents or falsely applies doctrine will fail to meet the requirement of the UNCITRAL Model Law and national arbitration laws on the need to decide in a reasoned way. Moreover, it is the opacity of complex machine learning models – the so-called black box problem – that undermines the principle that parties must understand the reasoning behind a given award, to ensure procedural fairness.
To avoid these threats, arbitration should include algorithmic-accountability measures. The Explainable AI (XAI) framework is a proposal that is based on making machine reasoning interpretable to people. EU Artificial Intelligence Act (2024) defines legal-decision systems as a high-risk category and stipulates that transparency, traceability, and human supervision of AI-involvement in adjudicatory aspects must be ensured in any cases. Singapore’s Model AI Governance Framework (2020) requires that consequential decisions must be human verifiable as well. This strategy is reflected in the Responsible AI Strategy (2023) in India, which includes the principles of fairness, inclusivity, and auditability in the model of its governance. These frameworks combined lead to one normative consensus, i.e. AI can help, but can never replace, human adjudicative reasoning.
From an ethical perspective, there is a common duty of disclosure between arbitral institutions and practitioners. Parties should be notified whenever AI tools are used to affect the evaluation of evidence, the orders of procedures or the draft awards. Without transparency, consent – the pillar of arbitral autonomy – is a myth. Institutions also need to have internal codes of ethics around AI that mandate that any digital tool be regularly audited for bias and validated by humans before it is deployed. Transparency reports, inspired by the principles of the Organisation for Economic Cooperation and Development (OEC), might be one possible and practical compliance mechanism, balancing accountability and innovation.
Ultimately, algorithmic fairness in arbitration is not an issue of programming finesse, but of maintaining legitimacy. If intelligent systems are going to improve decision-making, they must be operated under human oversight and generate intelligible, reviewable and justifiable reasoning. Absent these safeguards, the risk is that arbitration will lose its moral moorings of equitable justice as it turns into a mechanistic process unconnected to justice. Explainability, human oversight and disclosure are the three pillars of ethical algorithmic adjudication – the only way in which technology ensures that it is an asset to, rather than undermining the credibility of international arbitration.
CONCLUSION AND WAY FORWARD
The application of artificial intelligence (AI) in arbitration is both an opportunity of unprecedented magnitude and a responsibility of the utmost importance. Throughout the discussion above, it becomes clear that although AI has the ability to process information, give precedents, and even write drafts of analytical decisions at an unbelievable pace, it is never able to reproduce the deliberative logic, empathy, or even moral judgment that is characteristic of the human adjudication process. The effectiveness and predictability which algorithmic systems provide thus has to be weighed against the procedural protection which supports trust in arbitration.
The comparative study of the structures, including the European Union Artificial Intelligence Act (2024) and the Singapore Model AI Governance Framework (2020), as well as the Responsible AI Strategy (2023) by India, helps see that there is a common point: even with automation in decision-making, human control should be preserved. These regulatory frameworks come to an agreement where transparency, accountability, and fairness are principles that cannot be compromised in the deployment of AI in adjudicatory settings.
Moving ahead, three recommendations seem to be essential. To begin with, arbitral institutions ought to impose the mandatory disclosure in case AI tools affect the procedural or substantive elements of an award. Second, all final determinations should be maintained in human hands, it is necessary to leave the arbitral signature, both literal, and intellectual, to a human arbitrator. Third, ethical codes on the arbitral AI should be embraced in institutions, which requires bias audits, transparency of the data sources, and independent certification of the algorithmic systems. Together, these steps can be taken to avoid having efficiency devolve into opacity.
The challenge is, therefore, not against innovation but to steer it in good directions. Should human discretion and machine accuracy be allowed to co-exist within reasonable boundaries, it is possible that arbitration will be more powerful, more readily available, and more uniform. With the ever-changing nature of technology, the maxim that governs the case must still be the same: the future of arbitration will not be man vs machine -but man with machine.
