Responsible AI glossary
Key terms from this course, in plain English.
- Accountability
- Identifiable people and organisations answering for an AI system's design, use and effects, with routes to redress.
- AI ethics
- The study of moral questions raised by AI: what should be built, for whom and with what trade-offs.
- Calibration
- A fairness property where a given risk score means the same likelihood for every group.
- Contestability
- The ability of affected people to challenge AI-influenced decisions and obtain human review.
- Counterfactual testing
- Changing one attribute in an input (e.g. a name) and comparing outputs to detect bias.
- Data Protection Impact Assessment (DPIA)
- An assessment of risks to individuals from processing personal data, required under GDPR for high-risk processing.
- Datasheet for datasets
- Documentation of how a dataset was collected, its composition, intended uses and known issues.
- Demographic parity
- A fairness definition requiring equal rates of positive outcomes across groups.
- Disaggregated evaluation
- Measuring a model's performance separately for different groups rather than only overall.
- Equal opportunity
- A fairness definition requiring equal true-positive rates across groups; 'equalised odds' also requires equal false-positive rates.
- Ethics washing
- Publishing ethical principles without meaningfully changing practice.
- Explainability
- The ability to give meaningful reasons for an AI output.
- Fail-safe design
- Designing systems so that failures lead to a safe state, such as deferring to a human.
- Historical bias
- Bias that arises because data accurately reflects an unfair world.
- Indirect discrimination
- Under the Equality Act 2010, a practice that puts a protected group at a particular disadvantage without objective justification.
- Interpretable model
- A model whose decision-making can be directly understood, such as a simple scorecard.
- Measurement bias
- Bias from using a poor proxy for the real quantity of interest.
- Model card
- Documentation of a model's intended use, performance across groups, limitations and ethical considerations.
- Proxy
- A feature that indirectly encodes a protected characteristic, such as postcode.
- Representation bias
- Bias from under-representation of some groups in data.
- Responsible AI
- Designing, building, buying and using AI in ways that are safe, fair, transparent, accountable and rights-respecting.
- Responsible Scaling Policy
- Anthropic's framework setting safety measures that scale with model capabilities.
- Robustness
- The ability of a system to perform well under unusual, noisy or adversarial conditions.
- System card
- A document published with a frontier model describing its capabilities, safety evaluations and mitigations.
- Transparency note
- Microsoft's documentation explaining how an AI service works, its intended uses and limitations.