AIGP flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Supervised learning

A machine-learning approach that trains a model on labeled input-output pairs so it can predict the label for new, unseen inputs.

Reinforcement learning

A training paradigm in which an agent learns a policy by acting in an environment and receiving rewards or penalties, optimizing long-run cumulative reward rather than fitting labeled examples.

F1 score

The harmonic mean of precision and recall, computed as 2PR/(P+R); it balances false positives and false negatives and is useful when class distributions are imbalanced.

Overfitting

A model fits noise and idiosyncrasies in training data, producing low training error but poor performance on new data; mitigated with more data, regularization, and cross-validation.

Training versus inference

Training estimates model parameters from data using substantial compute; inference applies the trained model to new inputs and scales with the number of requests served.

Generative AI

Models that learn a data distribution and produce new content such as text, images, audio, or code, typically built on transformer or diffusion architectures.

Narrow AI

An AI system designed for a bounded task or domain, such as translation or image classification; all deployed commercial systems today fall into this category.

Automation bias

The human tendency to over-trust algorithmic outputs and reduce independent scrutiny, which can amplify system errors and weaken the effectiveness of human oversight.

EU AI Act risk tiers

Four-level classification - unacceptable (prohibited), high, limited (transparency duties), and minimal risk - that sets obligations for providers and deployers of AI systems in the EU.

NIST AI RMF functions

Govern, Map, Measure, and Manage: the four voluntary functions structuring the US NIST AI Risk Management Framework, with Govern as the cross-cutting culture and accountability function.

High-risk AI system

Under the EU AI Act, a system meeting Annex I product-safety rules or an Annex III listed use such as biometrics, employment, or credit, triggering conformity assessment and documentation duties.

OECD AI Principles

Five value-based principles covering inclusive growth, human rights, transparency, robustness and safety, and accountability; they underpin many national AI policies and the EU definition.

ISO/IEC 42001

An international standard specifying requirements for an AI management system, certifiable by third parties, covering governance, risk treatment, and continual improvement of AI processes.

GDPR Article 22

Grants individuals the right not to be subject to decisions based solely on automated processing, including profiling, that produce legal or similarly significant effects on them.

NYC Local Law 144

Requires employers and employment agencies using automated employment decision tools in New York City to conduct annual independent bias audits and notify candidates and employees.

General-purpose AI obligations

EU AI Act duties on providers of models usable for many tasks: technical documentation, copyright policy, training-data summary, and added systemic-risk evaluation above a compute threshold.

Model drift

Degradation of a deployed model's performance over time as conditions change; detected by monitoring input distributions, prediction distributions, and live outcome metrics against baselines.

Concept drift

A change in the statistical relationship between inputs and the target outcome, so a previously accurate model becomes wrong even when inputs look familiar.

Residual risk

The risk remaining after mitigation controls are applied; acceptance decisions compare residual risk against the organization's defined risk appetite and tolerance thresholds.

Model card

A short document reporting a model's intended use, training data, evaluation metrics by subgroup, ethical considerations, and known limitations for downstream users.

Data sheet for datasets

Documentation of a dataset's motivation, composition, collection process, preprocessing, recommended uses, and maintenance, enabling users to judge suitability and provenance.

Human-in-the-loop

An oversight mode in which a person reviews and approves each consequential model output before action, appropriate for high-severity or low-confidence decisions.

Red teaming

Structured adversarial testing in which specialists attempt to elicit harmful, unsafe, or incorrect model behavior to surface risks that standard benchmark evaluation misses.

Shadow deployment

Running a new model alongside the production model on live traffic without acting on its outputs, allowing risk-free comparison before a staged rollout.

Alignment problem

The difficulty of specifying objectives so an AI system's behavior reliably matches human intent and values, including risks of specification gaming and reward hacking.

Post-hoc explainability

Techniques such as SHAP, LIME, and counterfactuals that approximate why a trained model produced an output; the explanations are estimates and may not be faithful.

Prompt injection

An attack in which untrusted text, such as content from a web page or uploaded document, is interpreted by a language model as instructions and hijacks its behavior.

C2PA Content Credentials

An open technical standard for cryptographically signed provenance metadata attached to media, indicating how and by which tool content was created or edited.

Open-weight model

A model whose trained parameters are publicly downloadable, enabling local deployment and inspection but limiting the developer's ability to restrict or revoke its use.

Compute governance

Policy approaches that regulate AI development through control of computational resources, such as chip export controls, training-compute thresholds, and cloud reporting requirements.

Synthetic media

Image, audio, or video content generated or altered by AI, raising concerns about fraud, non-consensual intimate imagery, and election misinformation.

AI liability gap

The difficulty of assigning legal responsibility when harm emerges from complex multi-party AI supply chains in which no single actor controlled the full system.