Countries have been developing and deploying anti-corruption tools based on artificial intelligence with hopes of them having positive capabilities. Yet, we still lack empirical analyses of these automated systems designed to identify and curb corruption. Hence, this article explores novel data on 31 bottom-up and top-down initiatives in Brazil, presented as a case study. Methodologically, it uses a qualitative analysis and draws on secondary data and interviews to assess the most common features, usages and constraints of these tools. Data collected are scrutinised under a new conceptual framework that considers how these tools operate, who created them for what purpose, who uses and monitors these tools, what types of corruption they are targeting, and what their tangible outcomes are. Findings suggest that in Brazil, AI-based anti-corruption technology has been tailored by tech-savvy civil servants working for law enforcement agencies and by concerned citizens with tech skills to take over the key tasks of mining and crosschecking large datasets, aiming to monitor, identify, report and predict risks and flag suspicions related to clear-cut unlawful cases. The target is corruption in key governmental functions, mainly public spending. While most of the governmental tools still lack transparency, bottom-up initiatives struggle to expand their scope due to high dependence on and limited access to open data. Because this new technology is seen as supporting human action, a low level of concern related to biased codes has been observed.

各国一直在开发和部署基于人工智能的反腐败工具，希望它们具有积极的能力。然而，我们仍然缺乏对这些旨在识别和遏制腐败的自动化系统的实证分析。因此，本文以案例研究的形式探讨了巴西 31 项自下而上和自上而下举措的新数据。在方法上，它使用定性分析并利用二手数据和访谈来评估这些工具的最常见功能、用法和限制。收集的数据在新的概念框架下进行审查，该框架考虑了这些工具的运作方式、谁出于什么目的创建了它们、谁使用和监控这些工具、它们针对的腐败类型以及它们的实际结果是什么。调查结果表明，在巴西，基于人工智能的反腐败技术是由为执法机构工作的精通技术的公务员和具有技术技能的相关公民量身定制的，以接管挖掘和交叉检查大型数据集的关键任务，旨在监控，识别、报告和预测风险，并标记与明确的非法案件相关的可疑之处。目标是关键政府职能（主要是公共支出）中的腐败。尽管大多数政府工具仍然缺乏透明度，但由于高度依赖开放数据且访问权限有限，自下而上的举措难以扩大其范围。由于这项新技术被视为支持人类行为，因此人们对有偏见的代码的关注程度较低。