I find these discussions fascinating, yet a bit hyperbolic. Why? Because I think we have more pressing issues when it comes to AI, such as fairness and social responsibility. Machine Learning, a form of AI that powers predictions on everything from what music you’ll like to where crimes will be committed, is susceptible to biases, injected into the system from the data their predictions are trained on. FaceApp—a selfie editing app with a “hotness” filter that when activated makes your skin lighter—is just the most recent example of a company employing a biased Machine Learning system. But they’re far from the only ones. Bias has cropped up in facial recognition systems, criminal risk-assessment software, and programs that determine loan eligibility.
Keeping machines ethical
But it doesn’t have to be this way. The notion of algorithm accountability is still evolving, yet many are taking action to try to avoid Machine Learning bias. In my research, I’ve identified three things organizations can do to increase the transparency and accountability of their Machine Learning programs:
1. Designate a Machine Learning governance office
Organizations need to decide who’s accountable for validating the goals and performance of Machine Learning systems. They also need to decide who’s responsible when things go wrong. For example, if a model generates discriminatory harm, what will be the process for recourse, and who will have the authority to make changes to the system in a timely manner? Several organizations, including Google, Microsoft, IBM, and police body camera supplier Axon, are creating internal AI ethics advisory boards to help them answer these tough questions.
2. Build ethical design into all data science projects
To ensure algorithmic transparency is considered early, project managers can include an ethics review at the design stage of data science projects. During this review, project managers should work with designers and data scientists to identify potential ethical concerns by considering risks in the data, algorithms, system output, and interface. For example, does the project deal with a vulnerable population? Will the outcome limit choice for anyone? Will the output be deployed directly into a production system? Or will there be human checks in place? Is this considered a controversial use of algorithms?
3. Increase the interpretability of Machine Learning methods
This is perhaps a contentious recommendation, as Machine Learning methods, especially neural nets, are notorious for their imperviousness. Still, researchers from MIT, Carnegie Mellon, and many other institutions are developing techniques to provide rationale for the output of these systems. The Defense Advanced Research Projects Agency (DARPA) even has a dedicated Explainable AI project to fund research advances in this area. There is precedent here too: Mycin, an AI system developed in the 1970s to diagnose bacterial diseases like meningitis, was able to explain the reasoning behind its diagnosis and treatment recommendation. Designing explainable Machine Learning systems may necessitate a tradeoff between interpretability and accuracy. But what good is a system with 99% prediction accuracy if nobody trusts it?
Overall, Machine Learning has the potential to augment human decision-making, causing us to be more rational, consistent, accurate, and – yes – even fairer. But only if we govern these systems closely.