Panel Discussion on Ethical Data

Daniel Langkilde led a discussion of ethics in AI from an engineer’s point of view: how teams avoid drifting into harmful systems through a series of small decisions, and how to justify the extra cost of avoiding bias.

Markus Anderljung suggested openness to outside scrutiny and lessons from safety-critical fields such as nuclear power, and gave Amazon’s CV-screening model, which favoured men, as an example of the cost of getting it wrong. Danila Petrelli described where bias can enter a dataset, from the choice of sources and the definition of labels to the annotators, and argued that much of it can be handled with time and effort. Ivana Bartoletti called for diverse teams, review across functions, and standards and certification by sector, and pointed to law that already applies: anti-discrimination law and the GDPR’s rules on automated decisions and access to personal data. Mikio Braun stressed evaluation metrics that do not hide harm to minorities.

The panel also discussed whether companies or governments are better placed to regulate, and what automation means for jobs.