金融领域大型语言模型应用：负责任采用路径

LEVERAGINGLARGELANGUAGEMODELS INFINANCE: Pathways toResponsible Adoption This report is part of a collaboration between the European Securities and MarketsAuthority (ESMA), the FaiR (Finance and Insurance Reloaded) programme at Institut LouisBachelier and the FAIR (Framework for Responsible Adoption of Artificial Intelligence inthe Financial Services Industry) programme of the Alan Turing Institute. The viewsexpressed in this publication are privately held by the authors and should not be attributedto ESMA, Institut Louis Bachelier or the Alan Turing Institute. AUTHORS: BagattiniGiulio (ESMA)BrièreMarie(Institut Louis Bachelier)GuaglianoClaudia(ESMA)MapleCarsten(The Alan Turing Institute)SabuncuogluAlpay (The Alan Turing Institute) WORKSHOP PARTICIPANTS: BalagueChristine - Institut Mines TelecomBertucciLouis - Institut Louis BachelierBonaitaAlessandro - GeneraliChenYujia - University of EdinburghChonéAnne - ESMADymaczAleksandra - OECDElieRomuald - DeepmindFlicheOlivier - ACPRGomez TeijeiroLucia - Geneva UniversityGourierElise - ESSECHammoudaMaysara - PredictivaIonitaLaura - ESMAKrasniqiDafnis - Institut Louis BachelierLauridsenNico - EUILefortBaptiste - AI for AlphaLeote De CarvalhoRaul - BNP PAMLucasIris - AMF LyAntoine – SCORMarshallWilliam - ESMAMartinezLuis - ESMAMassonCorentin - AMFMeryRami - AmundiOhanaJean Jacques - AI for AlphaOtaeguiAlain - EBAPapiotisSotiris – ESMAPiazzaFederico - ESMAPicaultMatthieu - University of OrléansRavaMatteo – ESMARitolaTuomas - FIN-FSARouilGuillaume - AXA IMSmitsArtis - CB LatviaSzpruchLukasz - The Alan Turing InstituteTownsonSian - OliverWyman Special thanks toLouis Bertucci,Anne Choné,Dafnis Krasniqi,Corentin MassonandSian Townsonfor useful comments and contributions to the report. CONTENTS 1 . INTRODUCTION......................................................................................................................52 . CURRENT USE OF LLMS AND POTENTIAL APPLICATIONS.............................................. 73. POTENTIAL HARMS FROM THE USE OF LLMS................................................................. 124. TOWARDS RESPONSIBLE ADOPTION............................................................................... 155. DISCUSSION.........................................................................................................................266. REFERENCES....................................................................................................................... 30 ABSTRACT This report presents the summary of discussions held during a workshop on the use oflarge language models (LLMs) in the financial industry organised in June 2024 by theEuropean Securities and Markets Authority, the Alan Turing Institute and the Institut LouisBachelier. The workshop engaged 38 technology and finance experts to discuss threemain issues around (1) the current use of LLMs and their potential applications in thefinancial industry, (2) the risks and challenges associated with their use, and (3) the stepsnecessary for ensuring their responsible adoption. Generative LLMs are increasingly used in the financial industry to achieve operationalefficiencies in tasks involving text analysis and production, but they are also increasinglydeployed for public communication and customer interaction. This raises potential issues,often tied to legal, ethical and reputational harm. Against this backdrop, many financialorganisations are developing pathways to responsible LLM adoption that deal with thetopicsof model robustness,data dependency,security and privacy,fairness andaccountability. The finance sector can benefit from the establishment of appropriate evaluation metricsfor the use of LLMs, including benchmarks, and the development of industry standards.An appropriate supervisory framework, together with adequate staff training, can facilitatethis effort. Meanwhile, the downside posed by the carbon footprint of LLMs may need tobe carefully evaluated as the technology spreads and its use becomes integral to theeveryday operations of global businesses in the finance sector and beyond. 1 . Introduction Historically, the finance industry has been an early adopter of many technological advancements,from electronic systems to Big Data and artificial intelligence (AI). However, the adoption of newtechnologies is often cautious, due to risk aversion, regulatory compliance and legacy systems.Regulatory bodies set standards and guidelines to protect customer interests and create guardrails,while aiming not to stifle innovation. Financial institutions typically conduct thorough due diligenceand rigorous testing before implementing new technologies to mitigate potential damage. Theresulting configuration, while potentially conservative, aims to balance innovation and risk to meetevolving business needs while safeguarding the financial system’s stability. Large language models (LLMs), and particularly their generative versions, have found severalapplication a