Introduction

A shipped score can be miscalibrated, opaque, or unfair, and a prompt can be too. After this chapter the reader can point at calibration, explanation, fairness, unlearning, and federated training as the constraints on a model that already exists.

The chapter starts with the score itself. Calibration explains why predicted probabilities need calibration and how to adjust them, from classical methods to neural-net approaches such as temperature scaling. Once the numbers can be trusted, the next question is why the model produced them: Interpretable & Explainable AI (XAI) gathers the courses, libraries such as LIME, Anchor, and SHAP, and articles on explaining and interpreting ML models.

An explained model can still treat people unequally. Fairness, Accountability, and Transparency covers regulation, FAT research, bias, debiasing, fairness tools, and privacy. The same concerns reach language models through their inputs, so Fairness, Accountability, and Transparency In Prompts collects papers on debiasing with prompts and on LLM hallucinations.

The last two pages are about the data behind the model. Unlearning is the challenge of erasing a data point's influence on the input-output mapping of an ML model, and Federated Learning points to a guide on training across centralized and decentralized setups.

See also Developing A Virtual Psychologist With Gen-AI (October 2024), a walk-through of building a Gen-AI chatbot that offers therapeutic assistance and emotional support, a case where every one of these constraints applies at once.