Bias and fairness of artificial intelligence-based systems in financial crime

When it comes to fighting financial crime, there are challenges that go beyond simply stopping fraudsters or other bad actors.

Some of the latest, advanced technologies being launched often have their own specific issues that need to be taken into account during the adoption stages in order to successfully combat fraudsters without regulatory ramifications. In fraud detection, model fairness and data bias can occur when a system is over-weighted or under-represents certain groups or categories of data. In theory, the predictive model could falsely associate surnames from other cultures with fraudulent accounts, or falsely reduce risk in segments of the population for a particular type of financial activity.

Biased AI systems can pose a serious threat when reputations can be affected and occur when available data are not representative of the population or phenomenon under study. These data do not include variables that correctly capture the phenomenon we want to predict. Or, alternatively, the data may include human-generated content that may contain biases against groups of people inherited from cultural and personal experiences, leading to biases in decision-making. Although data may initially appear objective, it is still collected and analyzed by humans and therefore may be biased.

Although there is no silver bullet when it comes to removing the dangers of discrimination and unfairness in AI systems or lasting solutions to the problem of fairness and mitigating bias in model design and the use of machine learning, these issues must be considered both for society and business reasons.

Doing the right thing in AI

Addressing bias in AI-based systems is not only the right thing to do, but the smart thing to do for business—and the stakes for business leaders are high. Biased AI systems can steer financial institutions down the wrong path by unfairly allocating capabilities, resources, information or quality of service. They even have the potential to violate civil liberties, harm people’s safety, or affect a person’s well-being if they are perceived as disparaging or offensive.

It is important for enterprises to understand the power and risks of AI bias. Although the institution is often unknown, a biased AI-based system may use harmful patterns or data that expose racial or gender biases in the lending decision. Information such as names and gender can be a proxy for categorizing and identifying applicants in illegal ways. Even if the bias is unintentional, it still puts the organization at risk by failing to comply with regulatory requirements and may result in unfairly denying loans or lines of credit to certain groups of people.

Organizations currently lack the necessary pieces to successfully mitigate bias in AI systems. But as AI is increasingly deployed in business to inform decisions, it is vital that organizations strive to reduce bias not only for moral reasons, but also to comply with regulatory requirements and build revenue.

A culture and implementation of “equity reporting”.

Solutions that focus on equity-minded design and implementation will have the most beneficial results. Providers must have an analytics culture that considers responsible data collection, processing, and management as necessary components of algorithmic fairness, because if the results of an AI project are generated from biased, compromised, or distorted data sets, affected parties will not be adequately protected from discriminatory harm.

These are the elements of data fairness that data science teams should consider:

• Representation:Depending on the context, the under- or over-representation of disadvantaged or legally protected groups in the data sample can lead to a systematic disadvantage of vulnerable parties in the output of the trained model. To avoid such types of sampling biases, expertise in the field will be critical to assess the fit between the collected or acquired data and the population to be modeled. Technical team members should propose remedial measures to correct representative sample deficiencies.
• Fitness for purpose and sufficiency: It is important to understand whether the data collected is sufficient for the intended purpose of the project. Insufficient data sets may not fairly reflect the qualities that must be weighted to produce a justifiable result that is consistent with the desired goal of the AI ​​system. Accordingly, project team members with technical and policy competencies must collaborate to determine whether the amount of data is sufficient and fit for purpose.
• Source integrity and measurement accuracy:Effective bias mitigation begins at the very beginning of the data extraction and collection processes. Both sources and measurement tools can introduce discriminating factors into a data set. To ensure discriminatory non-harm, the data sample must have optimal source integrity. This includes ensuring or confirming that data collection processes include appropriate, reliable and unbiased measurement sources and robust collection methods.
• Timeliness and relevance: If datasets include stale data, then changes in the underlying data distribution may adversely affect the generalizability of the trained model. Provided these distributional deviations reflect changing social relations or group dynamics, this loss of accuracy with respect to the actual characteristics of the underlying population can introduce bias into the AI ​​system. In preventing discriminatory results, the timeliness and relevance of all elements of the data set must be carefully checked.
• Relevance, relevance and domain knowledge: Understanding and using the most appropriate data sources and types is critical to building a robust and unbiased AI system. A solid domain knowledge of the underlying population distribution and of the project’s predictive purpose is instrumental in selecting optimally appropriate measurement inputs that contribute to the reasonable resolution of the defined solution. Domain experts should collaborate closely with data science teams to help determine optimally appropriate measurement categories and sources.

While AI-based systems help automate decision-making processes and provide cost savings, financial institutions considering AI as a solution must be vigilant to ensure that biased decisions are not made. Compliance leaders need to be in step with their data science team to confirm that AI capabilities are accountable, effective and free of bias. Having a strategy that supports responsible AI is the right thing to do and can also provide a path to compliance with future AI regulations.