Title page
Contents
Foreword 3
Abbreviations and acronyms 6
Executive summary 7
1. Generative AI: Opportunities for enhancing anti-corruption and integrity in government 9
1.1. The OECD's questionnaire on generative AI for integrity and anti-corruption 10
1.2. Overview of the maturity of generative AI initiatives 11
1.3. Opportunities and benefits of LLMs for integrity actors 15
Annex 1.A. Key dimensions for assessing institutional digital maturity 21
2. Generative AI: Challenges, risks and other considerations for integrity actors in government 25
2.1. Overview of main challenges for integrity actors to adopt generative AI and LLMs 26
2.2. Building a generative AI and LLM capacity within institutions responsible for integrity and anti-corruption 31
2.3. Ensuring the responsible development and use of generative AI and LLMs by integrity actors 34
2.4. Mitigating the risk of generative AI as a tool to undermine integrity 43
References 46
Notes 49
Figure 1.1. Number of respondents to the OECD's questionnaire by type of organisation 10
Figure 1.2. Stage of generative AI and LLM use by type of organisation 12
Figure 1.3. Maturity of generative AI and LLM use by region 13
Figure 1.4. Perceived benefits of generative AI and LLMs for integrity actors' internal operations 17
Figure 1.5. Perceived benefits of generative AI and LLMs for anti-corruption activities by type of organisation (top two choices) 20
Figure 2.1. Main challenges for deploying generative AI and LLMs 26
Figure 2.2. Main challenges for deploying generative AI and LLMs by type of organisation 27
Figure 2.3. Primary data sources for building LLMs among questionnaire respondents 30
Figure 2.4. Integrity actors' approach for using LLMs 31
Figure 2.5. Safeguards to ensure responsible use of AI and LLMs 35
Figure 2.6. GAO's Artificial Intelligence Accountability Framework 40
Boxes
Box 1.1. The government-wide vision on generative AI of the Netherlands 14
Box 1.2. Lessons from Brazil's SAI and the development of ChatTCU 18
Box 2.1. The generative AI training programmes of the European Court of Auditors (ECA) 28
Box 2.2. Retrieval-Augmented Generation for LLMs 32
Box 2.3. France's LLaMandement for summarising legislative text 33
Box 2.4. The Office of the Comptroller General (CGU) of Brazil's approach to piloting LLMs 34
Box 2.5. The Corruption Prevention Commission (CPC) of Armenia's use of AI to verify asset declarations 37
Box 2.6. The OECD Principles on Artificial Intelligence 38
Box 2.7. The AI Accountability Framework of the US Government Accountability Office (GAO) 40
Box 2.8. Human-centred considerations for promoting transparency when evaluating LLMs 42
Box 2.9. Insights from the Independent Commission Against Corruption (ICAC) of New South Wales on AI's potential threats to anti-corruption work 43