AI Decision-Making
Algorithmic Bias
Often, algorithms get trained on historical data that contain undesirable biases. For instance, when Amazon tried to design an algorithm that could help it filter job applicants, the algorithm had a tendency to hire men. Amazon's computer models were trained with résumés submitted to the company over the past ten years. Given that there is a dominance of men in the tech industry, the algorithm learned to identify men as the most suitable job candidates. How can we make sure we avoid algorithmic bias?
- Futuremakers discussion on algorithmic bias and opacity, with Sandra Wachter, Helena Webb, and Brent Mittelstadt, chaired by Peter Millican
- Futuremakers discussion on AI and gender, with Gina Neff, Carissa Véliz, and Sian Brooke, chaired by Peter Millican
Algorithmic Opacity
One of the ethical challenges for machine learning algorithms is how opaque they are, in the sense that human beings are not able to inspect how the algorithm came to the conclusion it did. More often than not, human beings (even experts) cannot know what patterns the algorithm is picking up, or how. Opacity is a problem because it can easily lead to unknown bias. Suppose, for example, that you avoid giving a certain piece of data to an algorithm (e.g. people's race) in an effort to remove the risk of bias. If, unbeknownst to you, the algorithm is using post codes to come up with a decision, and post codes act as a proxy for race (because different neighbourhoods are racially very different), then that algorithm is going to be implicitly racist, but you might not realise this because the workings of the algorithm are opaque. How can we introduce more transparency into algorithms, and what kind of transparency is desirable? In which contexts is it legitimate to require, on ethical grounds, that AI should be explainable?
- Futuremakers discussion on algorithmic bias and opacity, with Sandra Wachter, Helena Webb, and Brent Mittelstadt, chaired by Peter Millican
AI in Business and Finance
AI is transforming the financial industry by performing increasingly important tasks such as granting and denying loans, detecting fraud, and even buying and selling stocks. Algorithms have been known to have caused 'flash crashes' in stock markets, which exemplifies how dangerous they can be. Given that AI can potentially be sexist, racist, unfair, and dangerous, how can we make sure that AI gets used in ethical ways in the context of corporations and finance?
- AI and Finance, Oxford talk by Nir Vulkan
- AI and Business, Oxford talk by Alan Morrison
- Futuremakers discussion on AI and banking, with Steve Roberts, Nir Vulkan, and Jannes Klaas, chaired by Peter Millican