Model Cards & System Documentation

A model card is a standardised document that describes what an AI model does, how it was built, what data it was trained on, how it performs across different groups, and where its known limitations lie. Think of it as a nutrition label for AI. The concept was introduced by researchers at Google in 2018 and has since been adopted, in various forms, across the industry. Good model cards include intended use cases (and explicitly unintended ones), performance metrics broken down by relevant subgroups, details about training data, and ethical considerations. System documentation goes broader, covering not just the model but the entire system it operates within - data pipelines, human review processes, update schedules, and escalation procedures. In practice, the quality of model cards and documentation varies enormously. Some organisations produce thorough, honest documents. Others treat them as compliance exercises, filling in the minimum required fields without providing genuinely useful information. The test of good documentation is whether someone who hasn't built the system can understand its capabilities, limitations, and appropriate use - and whether that understanding would change their behaviour.