Glossary

Plain-English definitions for the AWS, generative AI, and EU AI Act terminology used across the ccX site.

EU AI Act

EU AI Act - Regulation 2024/1689 on artificial intelligence. Classifies AI systems by risk (unacceptable, high, limited, minimal) with obligations on providers and deployers.
High-risk AI system - Annex III systems (employment, critical infrastructure, education, law enforcement, etc.) or safety components of regulated products. Strictest obligations: risk management, data governance, human oversight, logging, transparency, accuracy.
GPAI (General-Purpose AI) - broad-capability models (Claude, GPT, Llama) covered by the Act's GPAI chapter. Transparency, copyright, and - at systemic-risk scale - additional safety obligations.
Article 9 - Risk management - continuous, iterative risk-management process across the lifecycle.
Article 12 - Logging - automatic event logging for traceability.
Article 14 - Human oversight - design for effective natural-person oversight in use.
Conformity assessment - verification procedure before placing a high-risk AI system on the EU market.

AWS - Amazon Web Services, the public cloud platform.
Amazon Bedrock - managed foundation-model service (Anthropic, Meta, Mistral, Amazon Nova/Titan, others), plus agents, guardrails, knowledge bases, evaluation.
Bedrock Agents - Bedrock feature for agentic workflows that plan and call tools.
Bedrock Guardrails - content/PII/grounding policies layered on foundation-model calls.
AgentCore Runtime - AWS's enterprise runtime for long-running, memory-equipped agents.
Amazon EKS / ECS - managed Kubernetes / AWS-native container orchestration.
AWS IAM - identity, roles, and least-privilege policy on AWS.
VPC - isolated virtual network on AWS.
Auto Scaling Group (ASG) - dynamically-scaled group of EC2 instances.

GenAI - generative AI; models that produce new content rather than only classify or predict.
Foundation model - large pre-trained model adaptable to many tasks.
Agent - AI system that perceives, plans, and acts over multiple steps via tool calls.
Agentic AI - architectures built around autonomous or semi-autonomous agents.
RAG (Retrieval-Augmented Generation) - retrieve context, condition generation on it.
Evals - automated tests for AI quality, safety, and regression risk.
Guardrails - content filters, denied topics, PII redaction, output validation.
Hallucination - confident-but-wrong AI output.
Prompt injection - attack class where untrusted input manipulates the model.