Interview, Other
How to make computers less biased
The EconomistPa Manjang, Parr, Rochelle Farool, Meredith Broussard, Thomas Adams, Tamara Gilkes-Boer
- Technology is advancing rapidly, creating risks of automating structural inequalities and racism across sectors such as ride-sharing, facial recognition, medical diagnostics, and housing finance due to training data that lacks sufficient representation from ethnic minorities.
- Specific instances of embedded bias include Uber's algorithm potentially deactivating drivers without human oversight, pulse oximeters providing flawed readings for Black patients leading to inadequate care, and legacy credit-scoring algorithms rejecting Black home loan applicants at rates 80% higher than similarly situated white applicants.
- Experts anticipate that until a standard is established for identifying bias, the true extent of algorithmic discrimination will remain unknown, though new software tools are being developed to detect and counteract these biases.
- Industry reliance on self-policing is viewed with skepticism, suggesting that reducing bias will only occur organically when aligned with financial incentives, necessitating government intervention when it is not.
- Future outlooks call for significantly tighter regulation, including government agencies with increased resources and regulatory sandboxes like O'Neill Risk Consulting's system for testing algorithms before deployment.
- The European Union is positioned as a pioneer in this space, proposing risk-based oversight rules where stricter regulations apply to AI systems posing greater risks to fundamental rights, though concerns persist regarding the shared definition of risk required for effective implementation.
- Without effective regulation and concerted self-policing, there is a projection that systemic racism will become deeply entrenched in the digital future, potentially causing significant social division.