Glossary · AI, Examined

Term	Meaning
Algorithmic disgorgement	Regulator-ordered deletion of models or algorithms built with improperly obtained data.
Automation bias	The tendency of people to over-trust automated outputs.
Concept drift	The relationship being modeled changes over time, degrading performance.
Conformity assessment	EU AI Act process proving a high-risk system meets requirements before it reaches the market.
Counterfactual explanation	Shows what would have to change for a different outcome ('approved if income were $5,000 higher').
Data drift	The input data distribution changes from what the model was trained on.
Data poisoning	Corrupting training data to manipulate model behavior.
Datasheet for datasets	Documentation of a dataset's origin, makeup, collection and intended use.
Deployer	EU AI Act: the person or organization using an AI system under its authority.
Differential privacy	Adding calibrated noise so outputs reveal little about any one individual.
Disparate impact	A neutral-looking practice that disproportionately harms a protected group; unlawful without intent.
DPIA	Data Protection Impact Assessment, GDPR Art. 35, for high-risk processing.
Explainability	Explaining a given output in human terms.
Federated learning	Training across decentralized devices without centralizing raw data.
Foundation model / GPAI	A large model trained on broad data and adaptable to many tasks.
FRIA	Fundamental Rights Impact Assessment, EU AI Act Art. 27, for listed deployers of high-risk AI.
Hallucination	Confident but false or fabricated output from a generative model.
Human-in-the-loop	A human approves each output. On-the-loop: supervises and can intervene. Out-of-the-loop: no human.
Interpretability	Understanding a model's inner mechanics.
ISO/IEC 42001	Certifiable AI management system standard (Plan-Do-Check-Act, Annex A controls).
Membership inference	An attack that determines whether a record was in the training set.
Model card	Documentation of a model's intended use, performance across groups and limitations.
Model inversion	An attack that reconstructs training data from model outputs.
NIST AI RMF	US voluntary framework with four functions: Govern, Map, Measure, Manage.
Overfitting	The model memorizes training data and generalizes poorly.
Prompt injection	Malicious instructions in inputs that hijack a generative AI system.
Provider	EU AI Act: develops an AI system or GPAI model and places it on the market under its own name.
Proxy discrimination	A neutral feature (such as ZIP code) stands in for a protected attribute.
RAG	Retrieval-augmented generation: grounding model output in retrieved documents.
Red teaming	Structured adversarial testing to find failures and harms.
Substantial modification	A change to an AI system that affects compliance or intended purpose; can require new conformity assessment.
Synthetic data	Artificially generated data that mimics real data's statistics.
Systemic risk (GPAI)	High-impact capabilities, presumed above 10^25 training FLOPs.
Validation vs verification	Validation: built the right thing. Verification: built the thing right.

Terms you need to know