SCAIO Learn · Primer 09

A legislator's AI glossary.

25 key AI terms in plain language, calibrated for South Carolina legislators and legislative staff. Cross-referenced to SC bills where each concept matters.

SCAIO · scaio.org
How to use this glossary

A reference, not a reading.

This is designed for quick reference. Each term is defined plainly, with a one-line note on why it matters for SC policymakers. Browse by section: Foundations (what these systems are), How they work, How they fail, Policy vocabulary, SC-specific context.

Where a term appears in current SC legislation, the relevant bill is named.

Foundations

What these systems are.

Artificial Intelligence (AI)
An umbrella term for software that performs tasks usually requiring human cognition. Most current discussion is really about generative AI specifically.
Generative AI
AI systems that produce new content — text, images, audio, code. The category that drove the 2023+ public conversation.
Large Language Model (LLM)
A type of generative AI specialized in text. ChatGPT, Claude, Gemini, and similar systems are LLMs. See Primer 01.
Foundation model
A large, general-purpose AI model that other applications build on. GPT-4, Claude, Gemini are foundation models.
Multimodal AI
A model that handles multiple types of input/output — text, image, audio, video — within a single system.
Agent / Agentic AI
AI systems that take multi-step actions in the world (use tools, make purchases, send messages) rather than just generate text. Emerging category.
How they work

Useful technical vocabulary.

Training
The process by which an AI model learns from data. Pre-training (on broad data) is followed by fine-tuning (on narrower task-specific data).
Fine-tuning
Additional training on a narrower dataset to improve performance on a specific task or domain. Common for industry-specific deployments.
Inference
The process of using a trained model to produce output. "Running" the AI, as distinct from training it. Inference is what costs money per query.
Prompt
The instruction or input given to an AI model. The same model produces very different output depending on how it is prompted.
RAG (Retrieval-Augmented Generation)
Letting the model look things up in a curated source before answering. Reduces hallucination on factual questions. Common in government deployments.
Tokens
The units of text models read and produce. Pricing is typically per-token. Roughly 4 characters per token in English text.
How they fail

Failure modes worth naming.

Hallucination
When a model produces confident-sounding output that is factually wrong. Made-up citations, invented case law, fabricated statistics.
Bias
Systematic differences in model performance across demographic groups. Often a reflection of biases in the training data.
Drift
When a model's behavior changes over time — either because the model was updated by the vendor, or because the world has changed.
Adversarial input / prompt injection
An attack in which inputs are crafted to manipulate the model's output. Security-relevant for any public-facing AI deployment.
Deepfake / synthetic media
AI-generated images, audio, or video depicting events that did not occur. Covered in Primer 04.
Catastrophic forgetting
When a fine-tuned model loses capabilities it previously had. Relevant for evaluating long-term deployments.
Policy vocabulary

Terms that appear in legislation and guidance.

Human in the loop
A deployment pattern in which a human reviews and approves AI output before action. The core of S.443's approach to AI in coverage decisions.
Algorithmic accountability
The idea that institutions deploying AI systems should be answerable for the outcomes those systems produce, including errors.
AI use-case inventory
A public catalog of where AI is being used inside a government. Federal model: OMB M-24-10. State examples: California, Pennsylvania.
Bias audit
A formal evaluation of whether an AI system performs differently across demographic groups. Increasingly required in regulated deployments.
Parental consent
A consent framework for AI tools used with minors. The core mechanism in SC's H.5253 (AI in Education).
Disclosure
Notification to affected parties that AI was used in a decision or interaction. Common in proposed and enacted state AI legislation.
South Carolina context

SC-specific vocabulary worth knowing.

SC AI Strategy
The 2024 SC Department of Administration framework for state-government AI use. Established the AI Center of Excellence.
AI Center of Excellence (CoE)
SC's coordinating body for AI use across state agencies. Director: Rich Heimann.
SCRA AI Leadership Hub
SC Research Authority's cross-sector convening for AI in SC — academia, industry, and public sector.
ADAPT in SC
The NSF-funded EPSCoR program building distributed AI research capacity across SC institutions.
House Reg/Admin/AI/Cyber Committee
SC House standing committee with AI oversight. Chair: Rep. Jeff Bradley. Source of much SC AI-related legislation.
Bradley AI Assistant
SC state-government internal AI assistant deployed under the AI Center of Excellence. Named for the committee chair who championed it.
SCAIO Learn

Public-interest AI research for South Carolina.

This glossary will grow as the SC AI conversation develops. Suggest additions, corrections, or refinements at scaio.org/#contact. The full SCAIO Journal, the policy tracker, and the flagship report are at scaio.org.

scaio.org · jimmy@scaio.org

1 / 8 Exit