Bias and fairness
Models inherit the biases in their training data. Measure per-group performance, document intended use and known limits in a model card, and treat an unexplained gap between slices as a defect, not a statistic.
Privacy and security
Minimise personal data, scrub it at ingestion, and remember that anything in a prompt may be logged. Threats include training-data extraction, membership inference, model theft via distillation, and prompt injection through retrieved content.
Evaluation and governance
Ship an automated eval suite — golden questions, regression cases, adversarial prompts — and run it on every model or prompt change. Red-team before launch. Regimes such as the EU AI Act tier obligations by risk, requiring documentation, human oversight and transparency for higher-risk uses.