Compiled by SafeQual Health Software Team 3/18/2024

Scope: Glossary of terms for understanding and reasoning about governance, selection, architecture, and implementation of robust AI systems.

A

Accuracy

Probability of receiving a desired output.

Actionable intelligence

Information that can be used in decision making.

Adversarial example

Prompt designed to cause a model to produce undesired completion, usually generated by injecting an insignificant perturbation into a clean example.

Adversarial training

Using adversarial as well as clean examples in model training to improve adversarial robustness.

Agency

Capacity of a model to exhibit a degree of initiative or independence.

Alignment

Steering a model toward or the quality of being consistent with a group’s intended goals, preferences, standards, and ethical principles for example safety culture or high reliability organization (HRO)

B

Backward alignment

Ensures practical alignment of trained models by evaluating their performance in realistic environments and implementing regulatory guardrails. (see also: Forward Alignment)

C

Categorization

Assigning a category to a text.

Chaining

Mechanism of structuring information flow between AI components to accomplish complex tasks.

Classification

Process or technique of assigning a finite set of categories to a text that meaningfully describes its contents.

Completion

Output of a generative model in response to a prompt. (see also: Prompt)

Controlled vocabulary

A curated set of words or phrases applicable to a specific industry, application, or task.

Conversational model

A model optimized for generating completions in the context of past interactions.

Co-occurrence

The presence of distinct semantic elements in the same context.

D

Domain model

A model that is optimized for a specific industry or a task.

E

Embedding

A high-dimensional numerical vector representation of text semantics that allows use of numerical algorithms in text analysis. (see also Semantics, Semantic Search, and Similarity). 

Enrichment

Extracting meaningful or relevant information from text.

Extractive summarization

Identifies and groups together significant portions of a text to form a concise summary.

F

Few-shot

A prompting or leaning technique using a small number of in-context examples, in contrast to none or many. (see also: zero-shot, fine-tuning)

Fine-tuning

Modifying an existing foundational, pre-trained model by training on additional industry or context specific data to improve its performance. (see zero-shot, few-shot, training set)

Foundational model

A baseline model trained on unspecialized body of knowledge. (see also: fine-tuning, Domain model)

Forward alignment

Aims to produce trained models that follow alignment requirements. (see also: Backward alignment)

G

Generative AI

Techniques and algorithms that learn from existing artifacts to produce new ones.

Generative summarization

Distilling core content from long text for easier comprehension. 

Governance

Creation and enforcement of rules that ensure safe development and deployment of AI.

Grounding

Mapping completion output to available factual sources.

H

Hallucination

Fabricated data in completions that appear plausible and convincing but, in fact, wrong or inaccurate.

Hullucitation

A special case of hallucinations where fabricated citations or references to sources that are wrong or inaccurate are presented as fact. (e.g., ‘New York lawyers sanctioned for using fake ChatGPT cases in legal brief’

L

LLM

Large language models are produced from vast amounts of textual data to achieve general-purpose generation and classification capabilities. (see also: Generative AI)

P

Prompt

Text used as an input to generative model. (see also: Completion)

Prompt engineering

Designing and optimizing input to produce effective and accurate completions, usually through extensive experimentation.

R

Retrieval-augmented generation (RAG)

Including trusted knowledge sources outside of the unspecialized body of knowledge to improve quality of and confidence in output, usually in the form of citations and attributions.

RICE (Robustness, Interpretability, Controllability, and Ethicality)

Principles that characterize the objectives of alignment. 

ROAI

Return on investment in AI. A pun on ROI. 😊 S

Semantic Search

Using query intent and contextual meaning, in contrast to lexical properties. (see also Embedding, Similarity).

Semantics

Study of relations between linguistic forms and concepts and their mental representations. (see also Embedding, Similarity).

Sentiment

Overall disposition expressed in a text.

Sentiment analysis

Identifying sentiment in a text.

Similarity

A degree of semantic proximity of texts, usually expressed numerically. (see also Semantics Search, Embedding)

Summarization

Creating a short, accurate, and fluent digest of a longer text, which preserves its important information and overall meaning.

T

Temperature

A given degree of unpredictability of completions, often determining the balance between accuracy and potential usefulness.

Test set

A collection of sample prompts representative of the type of challenges used to evaluate the robustness and usefulness of a model.

Training set

A balanced collection of sample prompts and their desired completions used in fine-tuning models to recognize specific problem patterns. (see also fine-tuning, domain specific models)

Z

Zero-shot

Prompting ability or technique that does not rely on any examples. (see few-shot, fine-tuning)

Sources for this Document:

Created and maintained by SafeQual Health, www.safequal.net/gen-ai-glossary-for-healthcare-risk-management-leaders.html

Further Reading

For people with strong interest in Alignment, Ethics, and Robustness in AI.

Ji, J., Qiu, T., Chen, B., Zhang, B., Lou, H., Wang, K., … & Gao, W. (2023). AI alignment: A comprehensive survey. arXiv preprint arXiv:2310.19852.