Glossary Terms

Data used to fit a model’s parameters. Assess its quality, permissions and representativeness, and keep appropriate evaluation data separate to avoid overstating performance.

Using knowledge learned in one model or task to support another task. It can reduce development effort, but the adapted model still requires evaluation in the new context.

A neural networkA model made of connected computational units with adjustabl architecture using attention mechanisms to process relationships between inputs. It underpins many language models, but architectural capability does not establish accuracy or safe business use.

A historical test of whether a machine’s conversational behaviour can be distinguished from a person’s. It is not a sufficient measure of reliability, usefulness or suitability for a business workflow.

When a model fails to capture relevant patterns sufficiently well, including in its training data. Investigate the model, features and data rather than assuming more training alone will solve the problem.

Machine learning that identifies structure in data without predefined target labels. Patterns or clusters need interpretation and validation before they are used to support business decisions.

A person’s overall experience of using a product, service or workflow. Assess usefulness, clarity, effort and accessibility with users, including how AI outputs and errors affect their work.

The controls and information through which a person interacts with a system. Clear interfaces should help users understand actions, limitations and when review or escalation is required.

Data used to guide model selection or tuning during development. Because it informs those choices, it should remain separate from a final test setData reserved for assessing a model after development decisi used for independent evaluation.

The connected activities through which an organisation creates and delivers value. Examine dependencies and handoffs to assess whether an AI-enabled change improves the wider process rather than an isolated task.