← Research archive
AI Accountability · 2026

What Neural Network Design Means for Liability

A legal reading of model design, training quality and foreseeable failure

Read this research overview here. The original coursework essay will be added to the site once its source file is available.

Research question

How should law respond when an AI system fails because of training quality, architecture, overfitting, underfitting or poor generalisation?

The argument

AI error is not a single event detached from design. Liability analysis must be able to examine training data, architecture, testing, known limitations, error rates and deployment decisions.

Approach

Technical reading of neural-network design and training connected to standards of care, accountability and legal responsibility.

Why it matters

AI assurance, model risk, product safety, procurement standards and regulatory duties for high-impact systems.

This is an on-site research overview drawn from the MA portfolio, not the full original essay. Collaborative work is marked as group research.