AI Ethics & Policy Research

Algorithmic bias occurs when an AI system produces systematically different outcomes for different demographic groups. It arises from multiple sources: training data that reflects historical discrimination, feature selection that inadvertently encodes protected characteristics through proxy variables, and objective functions that optimise for aggregate accuracy at the expense of minority-group performance. Eliminating bias entirely is probably impossible — data is always historical, and history is always imperfect. The practical goal is to detect, measure, and mitigate bias to acceptable levels, which requires diverse development teams, representative datasets, and ongoing post-deployment monitoring. Bias audits, now mandated in several jurisdictions, are the emerging institutional mechanism for this process.

The EU has adopted the most prescriptive approach, with binding rules, prohibited practices, and pre-market conformity assessments for high-risk systems. The US has favoured a sectoral strategy — existing regulators like the FDA, FTC, and EEOC apply their existing authority to AI within their domains, supplemented by executive orders and voluntary frameworks. China has moved fastest on specific applications, regulating algorithmic recommendation systems, deep synthesis, and generative AI through targeted administrative rules issued by the Cyberspace Administration of China. Each approach reflects different political traditions: rights-based regulation in Europe, market-driven flexibility in the US, and state-directed control in China. Interoperability between these regimes is an active area of diplomatic negotiation.

It depends on what you mean by transparency and for whom. Technical transparency — publishing model architectures, training data, and performance metrics — is straightforward but insufficient for non-expert audiences. Explainability for end-users means providing reasons for specific decisions that are understandable and actionable: why was my loan denied, why did the diagnostic system flag this scan? For complex models, generating explanations that are simultaneously faithful to the model's reasoning and comprehensible to lay users is an open research problem. Post-hoc explanation methods like SHAP and LIME are useful but can be misleading — they approximate rather than reveal actual model behaviour. The field is converging on the view that transparency is a spectrum, not a checkbox.

Liability frameworks are still catching up with the technology. Under existing product liability law in most jurisdictions, the manufacturer of an AI system can be held liable for defects that cause physical harm. But the picture is murkier for non-physical harms — discriminatory lending decisions, wrongful arrests based on facial recognition, or misdiagnoses that fall below but do not clearly breach the standard of care. The EU AI Act introduces obligations on deployers as well as providers, creating a chain of responsibility. Several legal scholars have proposed strict liability regimes for high-risk AI, analogous to those governing pharmaceuticals. The practical challenge is that AI systems often involve multiple actors — data providers, model developers, fine-tuners, and deployers — making it difficult to attribute causation cleanly.