GERMANY Trends and Developments Contributed by: Tanja Pfitzner and Claudia Löschhorn, Pfitzner Legal (in collaboration with aginom)
These are not the only types of bias inherent in the use of AI tools in ADR. Researchers who studied AI systems acting as evaluators (comparing two pieces of work and picking the better one, which is more or less what we ask of an arbitrator or a judge) have found additional biases arising from how an argument reaches the system, irrespective of the quality of its content. They found that the systems may favour whichever submission they read first or, depending on the system, last. In a study of fifteen AI evaluators and more than 150,000 comparisons at Dartmouth College, Shi and colleagues (“Judging the Judges”, 2025) found that the effect was systemic and most pronounced when the competing answers were close in quality. This effect puts an adjudicator in a tough spot, as the order of submissions matters most when the case is finely balanced and adjudication is difficult. Another feature of using AI tools worth considering is its inherent “coin-flip problem”, which arises from the probabilistic way in which language models “cal - culate” outputs. Generative AI models typically intro - duce an element of chance into the result. The degree of output variability is determined by the AI system’s “temperature” settings, which not all AI users are aware of or know how to adjust for their specific use case. The “coin-flip” terminology may be an exag - geration. The variability does not, however, disap - pear even where the settings are fixed: for technical reasons (in particular, the behaviour of floating-point arithmetic on GPUs), the outcome is not deterministic. If you give the same system the same case file twice, with the same prompt and settings, you may still get two different assessments. If reproducibility cannot be ensured even within a single case, it is hard to see how the foundational ADR principle that like cases should be treated alike can be honoured. The risk is that the plausibility and authoritative style of the out - put (reinforced by the reviewer’s “automation bias” the tendency to treat AI output as more reliable than one’s own judgement) may conceal the element of chance involved. The appropriate response is therefore veri - fication rather than reliance. When it comes to international disputes, there is yet another issue that is particularly important but has not been widely discussed in the legal community so far. The models underlying many widely used AI sys -
tems have historically been trained on predominantly English-language material. OpenAI’s published statis - tics for GPT-3, an early prominent model, report that English accounted for approximately 92.65% of the words in its training corpus. More recent cross-lingual research likewise finds that model performance var - ies across languages, particularly between high- and low-resource languages (Li et al. “Language Ranker”, 2025). Over time, these biases and other potential deficits are highly likely to diminish as models evolve rapidly. Researchers at Oxford and Google DeepMind (Schut, Gal and Farquhar, 2025) examined internal represen - tations in four open-source models (Llama-3.1-70B, Gemma-2-27B, Aya-23-35B and Mixtral-8x22B) using French, German, Dutch and Mandarin. Across all four models, semantically loaded words in a non-English language were often routed through an internal rep - resentation closest to English before being rendered in the target language. The extent of this effect varied by model. In each case, though, the models appeared to be, to some extent, “thinking in English”. The study was generalistic. It did not examine or simulate legal analysis or arbitration. However, when translated to the ADR world, the implication is that an arbitral tri - bunal working with a German file and conducting pro - ceedings in German may receive analysis whose inter - nal representations, at an intermediate stage, were shaped by the English language. Cross-language performance gaps do vary by model, language and task, but even a modest asymmetry can have serious implications for a case built on a mixed-language file. If the AI system reads one party’s documents with more fidelity than the other’s, this creates an invis - ible imbalance between the parties to the arbitration. A human arbitrator’s language limitations are at least known, declared and managed through translation. The system’s limitations are less visible and may be difficult to measure. It is well known that translating every document in the file into English as a mitiga - tion measure can flatten or alter tone and nuance and that legal terms across jurisdictions may not share the same underlying principle. A carefully qualified wit - ness statement may emerge sounding more confident than the witness ever was or civil law concepts may get lost in translation. Anyone who has watched Treu und Glauben rendered as “good faith,” and quietly
396 CHAMBERS.COM
Powered by FlippingBook