ALGORITHMIC DECISIONS (AND PITFALLS) ARE EVERYWHERE
Machine learning algorithms with access to large data sets are making myriads of critical decisions, such as medical diagnoses, welfare eligibility, and job recruitment, traditionally made by humans. However, high-profile incidents of biases and discrimination perpetuated by such algorithms1 are eroding people’s confidence in these decisions. In response, policy and law makers are proposing2 stringent regulations on complex algorithms to protect those affected by the decisions.
Anticipating and preparing for regulatory risks is a significant executive concern3; however, the more immediate need is to address the public mistrust that motivates the widespread call for regulating algorithms. This growing lack of trust4 can not only lead to harsher policies that impede innovation but can also result in substantial potential revenue losses for the organization5. Four factors drive public distrust of algorithmic decisions.
- Amplification of Biases: Machine learning algorithms amplify biases – systemic or unintentional – in the training data.
- Opacity of Algorithms: Machine learning algorithms are black boxes for end users. This lack of transparency – irrespective of whether it’s intentional or intrinsic6 – heightens concerns about the basis on which decisions are made.
- Dehumanization of Processes: Machine learning algorithms increasingly require minimal-to-no human intervention to make decisions. The idea of autonomous machines making critical, life-changing decisions evokes highly polarized emotions.
- Accountability of Decisions: Most organizations struggle to report and justify the decisions algorithms produce and fail to provide mitigation steps to address unfairness or other adverse outcomes. Consequently, end-users are powerless to improve their probability of success in the future.
WHAT CHIEF DATA AND ANALYTICS OFFICERS SHOULD DO
These are hard challenges that don’t have clear or easy solutions. Yet, organizations must act now to improve fairness, transparency, and accountability of their algorithms and get ahead of regulations. Here are three key areas to start with:
- Increase awareness of AI: Educate business leaders, data scientists, employees, and customers about AI opportunities, limitations, and ethical concerns. Train employees to identify biases in data sets and models and encourage open discussions. Guide executives on when AI truly makes a difference and when traditional decision algorithms will do.
- Create an ecosystem for self-regulation: Build interdisciplinary teams to review potential biases and ethical issues in algorithmic models. Institute multi-tier checks with human interventions7 for algorithmic decisions. Mandate review and certification from external entities for critical algorithms. Embed transparency in data models and provide recourse to end-users to petition the results of the algorithm.
- Influence global regulations: Collaborate with government, private and public entities, think tanks, and industry associations to build policies that balance regulation with innovation.
While regulations are necessary, organizations that proactively improve people’s confidence in algorithmic decisions can avoid revenue loss, shape fair regulations and policies, and future-proof AI investments against adverse regulatory impact.