GERMANY Trends and Developments Contributed by: Tanja Pfitzner and Claudia Löschhorn, Pfitzner Legal (in collaboration with aginom)
absorbed into its common-law namesake, is familiar with the problem. The Harness In light of the issues set out above, how can arbitra - tion realise the benefits of modern AI systems while accounting for its limitations and managing the asso - ciated risks? When integrating AI into the decision-making process, it is worth considering how the law addresses human decision-makers. After all, human decision-makers can also be biased. Due to their potential individual bias, jurisdictions have implemented safeguards such as: • disclosure requirements; • the right to be heard; • reasoned decisions; and • mechanisms for scrutiny and review. To ensure that AI tools are used properly to support decision-making, similar principles (but aligned with the particular deficiencies and risks of using genera - tive AI tools to assist in decision-making) should be applied and the underlying concepts translated into a “harness”. Just as with the individual biases of human decision- makers and other sources of “noise” in human judge - ment, each of the risks associated with the use of AI (some of which we mentioned above) calls for spe - cific procedural controls to mitigate them. To name just one example: names, gender, nationalities and other identifying details should be removed in cases where identity is immaterial. It is imperative that the AI tool provide source-linked issue-by-issue justifica - tion that enables step-by-step a human check against the record and an independent reconstruction of the analysis. A bare conclusion is not sufficient. If material output is unstable, cannot be traced to the record or cannot be independently verified, the arbitral tribunal should not rely on it. Conveniently, re-running a query is itself a cost-effective test for the “coin-flip” problem: where answers diverge across runs, the question is one the system cannot reliably answer.
It is also indispensable to ensure testability and docu - mentability. Unlike a human arbitrator, an AI system can be asked the same question repeatedly, both in different sequences and in different languages. For the comparison to be meaningful, the AI system, model version, prompt, context and relevant settings should be kept constant and recorded to the extent possi - ble. Its tendencies can therefore be measured in the context of the specific case. In addition, the protocol can be recorded, disclosed and agreed upon by the parties or ordered by the arbitral tribunal. The outlined approach translates generally recognised AI governance principles to arbitration use cases. The most widely referenced voluntary and sector-neutral framework in the field, the “NIST AI Risk Management Framework”, published in 2023 by the U.S. National Institute of Standards and Technology (NIST) organ - ises governance into four functions: govern, map, measure and manage. The outlined harness reflects these four steps. The procedural order governs the process; the blinding and the language decisions map, the repeated and reversed runs measure and the human decision manages what remains. Still, this approach cannot guarantee fairness and even the NIST framework’s definition of trustworthy AI does not promise fairness outright. It promises fairness “with harmful bias managed”. The ISO’s certifiable AI man - agement standard (ISO/IEC 42001:2023), published in late 2023, has essentially arrived at the same place but is more prescriptive. Current Rules and Guidance Arbitral institutions Like some commercial courts, arbitration institutions have responded to these developments by issu - ing guidelines and holding consultations. There is a general trend of institutions emphasising the need to ensure that arbitrators continue to analyse the law and the evidence relevant to a dispute independently. Already, the first guidelines published in April 2024 by the Silicon Valley Arbitration & Mediation Center identify the amplification of bias as a significant risk and state clearly that an arbitrator must not delegate any part of the arbitrator’s personal mandate to an AI tool, in particular not the arbitrator’s decision-making process.
397 CHAMBERS.COM
Powered by FlippingBook