Language technologies: assisted drafting and artificial intelligence

Read the PDF edition (2 pages)

Generative artificial intelligence tools entered the professions of the financial centre from 2023 onwards. They draft first versions of letters, summarise documents, translate, extract information from contracts and answer legal or financial questions. Their command of French is good. Their command of Swiss legal French is improving, though it remains marked by the predominance of English and of French law in the training data.¹ A Geneva practitioner who uses these tools must check every legal reference, every figure and every technical term, because the errors occur precisely at the points where Swiss law differs from French law or from Anglo-American law.

Three questions arise for institutions. The first is confidentiality: client data cannot be submitted to external tools without contractual and technical safeguards. The supervisory authorities further expect documented governance of these uses.² The second is liability: a document produced with the help of a tool binds the institution in the same way as a document drafted by an employee, so that human review remains mandatory. The third is competence: while the tools make drafting faster, they do not train drafters. A financial centre that allowed its command of its own legal language to atrophy in favour of tools whose data and biases it controls neither would be taking a long-term risk.

Used with discernment, these tools instead reinforce the linguistic advantage of the financial centre. They make it possible to serve a more diverse clientele in its own language, to process documents in Portuguese, Arabic or Chinese with sufficient understanding for sorting and analysis, to produce multilingual versions of information documents rapidly and to free up time for advisory work, where human language remains irreplaceable. The Geneva institutions that combine these tools with a demanding command of French and English will have a service capacity that their competitors will not necessarily possess.

Notes

1. Chen, Yin and Zhou, “LegalCiteBench: Evaluating Citation Reliability in Legal Language Models”, arXiv:2605.10186, May 2026: the benchmark assesses, without access to a document database, the ability of models to retrieve, complete and verify case-law references drawn from one thousand American court decisions; all twenty-one models tested, proprietary and open, score below 7 out of 100 on the citation recall and completion tasks, and the rate of misleading answers exceeds 94% for twenty of them; the detection of errors in a citation supplied to the model succeeds far better, from 40 to 96 depending on the model, and legal specialisation, like an increase in size, brings only modest gains; the authors specify that the benchmark covers American common law exclusively, civil-law systems and non-American jurisdictions not having been assessed. ↩
2. FINMA, Guidance 08/2024 of 18 December 2024, “Gouvernance et gestion des risques en lien avec l’intelligence artificielle” (Governance and risk management in connection with artificial intelligence) (Centre de droit bancaire et financier, Caballero Cuevas, 19 December 2024, cdbf.ch): seven areas of attention: governance (roles, responsibilities, competences), inventory and classification of risks, data quality, testing and ongoing monitoring, documentation, explainability of results and independent verification; supervised institutions adapt their arrangements to their risk profile. ↩

For general information only. This article does not constitute legal or tax advice, does not replace advice on an individual case and creates no lawyer-client relationship. It reflects the legal position as of September 2025.