Feedzai Research publishes OpenL2D: a Breakthrough Human-AI Collaboration Framework, Accessible to All Tech Teams
Available from GitHub, OpenL2D will allow for Responsible AI to widely develop in the industry
New York, NY and Lisbon, Portugal | May 13, 2025 — Feedzai, a global leader in fraud and financial crime prevention, AI-powered risk management for financial services, announces a major milestone: its Research team’s pioneering work on human-AI decision-making has been published in Nature Scientific Data, one of the world’s most prestigious scientific journals. In tandem, Feedzai Research has released both the newly developed OpenL2D framework and the accompanying FiFAR dataset as open-source resources.
New Benchmarking Framework Set to Transform Human-AI Collaboration
- Feedzai’s OpenL2D framework enables the creation of realistic, customizable synthetic experts for benchmarking AI systems that defer decisions to humans, which is crucial for high-stakes industries like fraud detection.
- The accompanying FiFAR dataset, a publicly available, synthetic dataset designed to help researchers develop and test systems, features predictions from 50 simulated fraud analysts on 30,000 real-world financial cases, setting a new standard for evaluating collaborative AI systems.
- The research demonstrates that the performance of “learning to defer” (L2D) algorithms varies widely depending on the diversity and capacity of available human experts, underscoring the need for robust, real-world-inspired evaluation.
Addressing a Critical Gap in Responsible AI
Human-AI collaboration is increasingly vital in sectors where errors can have serious consequences. Until now, the lack of large, realistic datasets with expert predictions has limited progress in the field. Feedzai’s OpenL2D and FiFAR address this gap by:
- Simulating expert decision-making processes—including biases, inconsistencies, and work capacity constraints—based on the latest behavioral science and domain-specific research.
- Allowing researchers and practitioners to rigorously test AI systems under conditions that mirror real-world challenges, such as concept drift, fairness concerns, and capacity constraints.
- Enabling the development of fairer, more reliable AI systems that can intelligently defer to human judgment, especially in sensitive applications like financial fraud prevention.
“Our work bridges the gap between academic research and real-world deployment of human-AI systems. By making these tools open-source, we empower the global research community to build safer, fairer, and more effective AI for critical decision-making,” said Pedro Bizarro corresponding author, Chief Science Officer, and co-founder at Feedzai.
Open Science and Industry Impact
- OpenL2D is now available on GitHub, providing code, datasets, and evaluation tools for the AI research community.
- The FiFAR dataset is designed specifically for fraud detection research. However, by modifying expert properties, researchers can use the OpenL2D framework to generate similar datasets for other high-stakes domains.
- The framework’s flexibility supports benchmarking for fairness, robustness, and expert diversity—key priorities for regulators and industry leaders alike.
