AI Glossary for Local and State Health Departments

AI systems designed to autonomously reason through multi-step problems and take actions with limited human oversight. Unlike standard AI assistants that respond to individual prompts, agentic AI can manage entire workflows end-to-end, deciding what steps to take next on its own.

This is an emerging category of AI. While most health departments are focused on simpler AI applications today, agentic AI represents where the technology is heading - and raises important questions about oversight and accountability.

Future example: An agentic AI might be asked to "compile our quarterly communicable disease report" and then autonomously pull data from surveillance systems, generate charts, draft narrative sections, and format the document - with a human reviewing only the final output rather than each step.

Technology that enables computer systems to perform tasks that typically require human intelligence, such as understanding language, recognizing patterns, making predictions, or generating content. AI encompasses many approaches, including machine learning, large language models, and rule-based systems.

Local and state health departments are using AI for tasks like drafting grant applications, summarizing reports for leadership, creating community communications, supporting food establishment inspections, and analyzing surveillance data.

In practice: When you use an AI assistant to help reformat a CDC grant application, summarize a lengthy state guidance document, or draft social media posts about flu season, you're using AI.

The structures, processes, and responsibilities an organization establishes to manage AI risks and guide AI use over time. Effective governance ensures AI is used responsibly, transparently, and in alignment with organizational values and legal requirements.

Governance is the ongoing organizational framework: who makes decisions about AI, how new tools get evaluated, how you monitor and adapt as needs change. An AI policy is one key output of your governance structure, codifying the rules staff follow day to day. Governance also includes training programs, review boards, vendor evaluation criteria, and incident response procedures.

Key questions for your department: Who approves new AI uses? What data can and cannot be used with AI? How are AI outputs reviewed before publication? Who do staff ask when they're unsure? Answering these questions is the foundation of AI governance. See also: AI Policy.

A set of step-by-step rules or instructions that a computer follows to complete a task or solve a problem. Algorithms are the building blocks of all software, from simple calculations to complex AI systems. In AI contexts, algorithms are the procedures used to process data and train models.

In practice: The rules in your surveillance system that flag certain disease reports for immediate follow-up, or the logic in your permit system that routes applications to different reviewers, are algorithms - whether or not they involve AI.

A way for different software systems to communicate with each other. APIs allow applications to access AI capabilities without building the AI themselves - essentially a standardized connection point between your software and an AI service.

Many AI tools are accessed through APIs, where your application sends requests to an AI service and receives responses. This is different from using a web-based chat interface directly. Your IT staff or software vendors may mention APIs when discussing how to integrate AI into existing systems.

In practice: Your department's inspection management system might use an API to connect to an AI service that helps categorize incoming complaints or suggest violation codes - without inspectors needing to copy and paste into a separate AI tool.

Systematic errors in AI systems that produce unfair, inaccurate, or skewed results for certain groups compared to others. Bias can enter AI systems through the data used to train them, the way problems are framed, or the metrics used to evaluate performance.

If historical data reflects past inequities, AI trained on that data may perpetuate those patterns. AI trained primarily on data from certain populations may perform poorly for underrepresented groups. These issues are particularly important in public health, where equity is a core value.

Questions to ask: What populations was this tool trained on and tested with? Has it been validated for communities like those we serve? What are its known limitations for different demographic groups?

The amount of text a large language model can consider at one time, measured in tokens. Think of it as the AI model's working memory: it determines how much information the model can hold in mind when generating a response.

A larger context window allows the AI to work with longer documents or maintain longer conversations without losing track of earlier information. When you exceed the context window, you may need to start a new chat with fresh context. The size of the context window varies depending on which AI tool you're using.

Context window limitations are one of the key reasons RAG and knowledge bases matter. Rather than trying to fit all of your documents into a single context window, RAG retrieves only the most relevant passages for each question, making much better use of the available space.

In practice: If you paste your 40-page Community Health Assessment into an AI assistant and ask it to summarize, the summary may ignore or contradict information from earlier sections if the document exceeds the context window. This is one reason structured approaches like RAG produce more reliable results for large document sets.

Legal and policy considerations around what data can be entered into AI systems, how that data is stored and used, and how long records must be kept. For government agencies, these considerations are shaped by your specific jurisdiction's laws and policies.

Key considerations vary by jurisdiction:

Public records: AI prompts and outputs may be considered public records subject to disclosure under your state's open records laws. Some states (like Washington) have explicitly addressed this, stating that AI interactions relating to public business are public records - even if created using personal accounts.

HIPAA: Health departments that are HIPAA covered entities must evaluate whether AI use implicates protected health information. Consumer AI tools typically cannot be used with PHI. Commercial/enterprise AI solutions may be appropriate with proper agreements, but this requires evaluation with IT and legal counsel.

Data retention: Your jurisdiction's records retention requirements likely apply to AI interactions. Not all AI records need to be retained for the same period - it depends on the content and function, just like other records.

Action step: Work with your legal counsel to understand how your state's open records laws, HIPAA status, and records retention requirements apply to AI use. Policies developed for one jurisdiction may not apply to yours.

A way of converting text (or images, or other data) into numerical representations that capture meaning. Embeddings allow AI systems to understand that "flu shot" and "influenza vaccination" are related concepts, even though the words are different.

Embeddings are the technology that powers RAG and knowledge base search. When your department uploads documents to a RAG-enabled system, those documents are converted into embeddings. When you ask a question, your question is also converted into an embedding, and the system finds documents whose embeddings are most similar to your question's embedding. This "semantic search" is far more effective than simple keyword matching.

Why it matters: If you search a keyword-based system for "food safety complaints," you'll miss documents that use the phrase "restaurant inspection concerns." An embedding-based search understands these are related and retrieves both. This is what makes RAG systems so much more useful than simple document search.

A machine learning technique that adapts a pre-trained model to perform better on specific tasks. Instead of training a model from scratch, you start with one that already understands general patterns and adjust it to work with your data. This approach leverages transfer learning: using knowledge gained from one task to improve performance on a related task.

Fine-tuning typically happens at the AI creator or lab level, rather than something end users do. It often requires using an open source model and is based on the work of a foundation model. Fine-tuning is particularly effective when you have limited training data or want to build on existing model capabilities.

In practice: A general language model might be fine-tuned on public health documents and terminology so it better understands concepts like "critical violation," "contact tracing," or department-specific procedures.

A large AI model trained on broad, diverse data that can be adapted for many different tasks and applications. Foundation models serve as the base for more specialized applications: they provide general capabilities that can then be customized or directed through prompting or fine-tuning.

Most AI chatbots and writing assistants are built on foundation models. The same underlying model might power applications ranging from customer service chatbots to code assistants to document summarizers.

AI that creates new content - text, images, audio, video, or code - rather than just analyzing or classifying existing data. Generative AI learns patterns and structures from its training data and uses those patterns to produce new, original outputs.

Most AI chatbots and writing assistants are generative AI, powered by large language models. Image generators, audio creators, and code assistants are also generative AI. The "generative" label distinguishes these from AI that only analyzes, categorizes, or predicts.

In practice: When AI drafts a press release, creates talking points, or generates an infographic, it's producing new content based on patterns learned during training - not retrieving pre-written content.

Safety measures that limit what AI systems can do, designed to prevent harmful, inappropriate, or risky outputs. Guardrails operate at multiple levels:

Built-in guardrails are safety features included by the AI vendor, like refusing to provide medical diagnoses, generate harmful content, or assist with illegal activities.

Organizational guardrails are policies and controls your department establishes - approved use cases, prohibited uses, required review processes, and escalation procedures.

Policy examples: "All AI-generated external communications require supervisory review." "AI may not be used for individual enforcement decisions." "AI-generated health guidance must be verified against authoritative sources before use."

Platforms like PH360 have built-in guardrails designed specifically for public health use cases.

When AI generates information that appears plausible and is stated confidently but is factually incorrect or entirely fabricated. Large language models work by predicting what text is likely to come next based on patterns learned during training. They do not "know" facts or verify accuracy.

AI hallucinations can include citations to studies that don't exist, invented statistics, fabricated regulations or laws, non-existent organizations, and confident answers to questions where accurate information isn't available. The AI presents these with the same confidence as accurate information.

A system design principle where humans review, verify, and approve AI outputs before they are used or published. This ensures accountability, catches errors and hallucinations, and maintains human judgment in consequential decisions.

Human-in-the-loop is essential for public health applications where errors can affect community health, public trust, legal compliance, or individual rights. The human provides contextual judgment, accountability, and quality assurance that AI cannot. Staff need training to effectively review AI outputs.

Best practice: AI drafts content → qualified human reviews and edits → human approves for use. Never allow AI to publish external communications, make enforcement decisions, or take consequential actions without human review.

The process of an AI model generating a response or prediction from new input. When you type a question into an AI assistant and receive an answer, the model is performing inference: applying the patterns it learned during training to your specific input.

Inference is distinct from training. Training happens once (or periodically) and teaches the model its capabilities. Inference happens every time someone uses the model. Inference costs (measured in tokens) are what drive the per-use pricing of most AI services.

In practice: When you ask an AI assistant to summarize a 10-page inspection report, the model runs inference to read your input and generate the summary. The time and computing resources required for that inference are what the AI vendor charges for.

A curated, organized collection of documents that an AI system can search and reference when generating responses. A knowledge base typically contains your department's policies, procedures, guidance documents, FAQs, and other authoritative materials that you want the AI to draw from.

A knowledge base is different from simply uploading a file to a consumer AI chatbot. When you upload a PDF to a tool like ChatGPT, that file is temporarily placed into the context window for that single conversation. It disappears when the session ends, cannot scale beyond the context window's size limit, and the AI has no retrieval logic to find the most relevant sections. A true knowledge base is persistent, searchable, and governed. Documents are converted into embeddings, stored permanently, and retrieved selectively using RAG when relevant to a user's question.

For health departments, a well-maintained knowledge base ensures that AI responses are grounded in your jurisdiction's actual policies and procedures rather than generic information from the model's training data.

In practice: PH360 uses a knowledge base approach where health departments upload their own policies, procedures, and guidance documents. When staff ask a question, PH360 searches the knowledge base for relevant content, provides that content to the model, and generates a response grounded in your department's actual documents, with citations back to the source material.

A type of AI model trained on massive amounts of text to understand and generate human language. LLMs learn relationships between words and concepts from their training data, enabling them to produce remarkably human-like text, answer questions, summarize documents, and perform many language-related tasks.

LLMs are "large" because they contain billions of learned parameters and were trained on enormous text datasets. They power most AI chatbots and writing assistants. The products you interact with (like ChatGPT, Claude, or PH360) are built on top of LLMs, adding features like guardrails, knowledge bases, and user interfaces to the underlying model. Because LLMs predict likely text rather than retrieving verified facts, they can produce hallucinations.

When you hear this: If someone mentions "LLMs" or "large language models," they're referring to the technology behind AI text tools like chatbots and writing assistants.

A type of artificial intelligence where systems learn patterns from data rather than following explicitly programmed rules. Instead of a human defining every rule, the system discovers patterns through exposure to examples and improves through experience.

Machine learning has been used in public health for years - often before it was called "AI." Syndromic surveillance systems, disease forecasting models, and pattern detection tools often use machine learning approaches.

In practice: A machine learning system at your state health department might analyze years of syndromic surveillance data to identify patterns that predict disease outbreaks - patterns that epidemiologists might not notice or be able to articulate as explicit rules.

An open standard that allows AI assistants to connect directly with your existing software systems, databases, and tools. Rather than copying and pasting information into an AI chat, MCP-enabled AI can access approved data sources in real time, take actions in connected systems, and work across multiple tools within a single conversation.

Think of MCP as a universal adapter. Without it, every AI integration requires custom development to connect system A to system B. With MCP, AI tools use a single standardized protocol to connect to many different systems. This matters for health departments because it means AI can work with your existing infrastructure (case management systems, databases, document repositories) rather than requiring you to replace or heavily modify what you already have.

MCP is particularly important for understanding how modern AI platforms differ from standalone chatbots. A standalone chatbot can only work with what you type or paste into it. An MCP-enabled platform can pull information from your knowledge base, check records in a database, or interact with other software, all while maintaining the security and access controls your IT team requires.

In practice: A health department using an MCP-enabled AI platform could ask a question and have the AI automatically search internal policy documents, look up relevant data from a connected system, and draft a response that synthesizes information from multiple sources, without the staff member needing to manually gather that information first.

The core component of an AI system: a mathematical structure trained on data that can make predictions, generate text, classify information, or perform other tasks. When people refer to "GPT-4," "Claude," or "Llama," they are referring to specific AI models.

A model is the product of training. During training, the model learns patterns from data and encodes those patterns as numerical values called parameters. A large language model may have billions of parameters. Once trained, the model uses those parameters to perform inference, generating responses to new inputs.

Models are not the same as the products built on top of them. A single model can power many different applications. The model provides the core intelligence; the product wraps it with a user interface, guardrails, knowledge base connections, and other features that make it useful and safe for specific purposes.

Why it matters: When evaluating AI tools for your department, understanding the difference between the model and the product helps you ask better questions. Two products might use the same underlying model and produce very different results because of how they're configured, what guardrails they include, and what data sources they connect to.

AI systems that can work with multiple types of data - text, images, audio, video - rather than just one. Multimodal AI can accept diverse inputs and often generate diverse outputs, enabling applications that weren't possible with text-only or image-only systems.

Large multimodal models have wide application in healthcare, research, public health, and drug development. These systems can analyze medical images, process spoken language, read documents, and integrate information across formats.

In practice: PH360 uses multimodal AI to take a picture of an inspection report along with other context information and generate violation summaries - working across images and text in a single workflow.

The field of artificial intelligence focused on enabling computers to understand, interpret, and generate human language. NLP encompasses technologies from basic spell-check to sophisticated large language models.

NLP enables AI to work with unstructured text - documents, emails, reports, social media posts - rather than just structured data in databases. This capability is essential for many public health AI applications, allowing AI to summarize documents, extract key findings, categorize information, and generate human-readable content.

What an AI system produces - the text it generates, predictions it makes, recommendations it provides, or decisions it supports. Understanding outputs is important because different types of outputs carry different levels of risk for your department.

Consider both the consequences if the output is wrong and who will see it. Internal drafts that get human review are lower risk than public-facing communications. Suggestions that inform human decisions are lower risk than AI making decisions directly.

A written document that defines how your department will and will not use AI. An AI policy establishes approved tools, acceptable use cases, prohibited uses, data handling requirements, review processes, and staff responsibilities. It is the practical, enforceable document that turns your department's AI governance decisions into clear guidance for staff.

A policy is different from governance. Governance is the ongoing organizational structure: who makes decisions, how you evaluate new tools, how you adapt over time. A policy is one output of that governance structure. It codifies the rules staff follow day to day.

12 things your AI policy should address:

1. Purpose and scope (who and what the policy covers). 2. Approved AI tools and how new tools get evaluated. 3. Acceptable use cases and prohibited uses. 4. Data classification requirements (what data can and cannot be entered into AI). 5. HIPAA and privacy considerations. 6. Human review and approval requirements before AI outputs are used. 7. Public records and data retention obligations. 8. Bias and equity considerations. 9. Transparency and disclosure requirements (when to tell the public AI was used). 10. Training requirements for staff. 11. Incident reporting and escalation procedures. 12. Policy review schedule and update process.

Getting started: Your policy does not need to be perfect on day one. A simple one-page document that addresses approved tools, prohibited uses, and required human review is far better than no guidance at all. You can expand it as your department's AI use matures.

F&T Labs: 12 Things Your AI Policy Needs

The input you provide to an AI system - your question, instruction, context, or the material you want it to work with. The quality of your prompt significantly affects the quality of the AI's output.

Including relevant context is perhaps the most important thing you can do. This means providing information about your goals, your audience, your constraints, and any specific requirements. Effective prompts typically include: the specific task, relevant context, the intended audience, desired format or length, and any constraints or requirements. Vague prompts produce generic results; specific prompts produce more useful outputs.

Compare:
Vague: "Summarize this report."
Specific: "Summarize this 30-page Community Health Assessment in 5 bullet points for the Smith County Board of Health. Our commissioners value data-driven recommendations and cost-effectiveness. Focus on the top three health priorities and what resources we'd need to address them. Use plain language that non-public-health board members will understand."

The practice of writing, refining, and optimizing prompts to get better results from AI systems. Prompt engineering is about learning what information and instructions help AI produce useful, accurate, and appropriate outputs.

This is one of the most practical AI skills for public health staff. Techniques include providing clear context, giving examples of desired outputs, breaking complex tasks into steps, and specifying constraints. Hands-on training helps staff build these skills quickly.

In practice: Learning to write better prompts is often the single most valuable AI skill for health department staff. A well-crafted prompt can dramatically improve output quality without any technical changes to the AI system itself.

F&T Labs: AI Training for Health Departments

The practice of documenting how work actually gets done - step by step, decision by decision - before attempting to implement AI. Process mapping creates a visual representation of a workflow, showing inputs, outputs, decision points, handoffs, and pain points.

Process mapping is essential groundwork for successful AI implementation. Without understanding your current processes, you can't identify where AI would actually help versus where it might create new problems. It also helps you set realistic expectations and measure whether AI is actually improving outcomes.

Why it matters: Health departments that skip process mapping often discover too late that AI doesn't fit their actual workflow, or that the real bottleneck was something AI can't solve. Mapping first helps you target AI where it will have the most impact.

F&T Labs: Why Process Mapping is Your Secret Weapon for AI Success

A technique that improves AI accuracy by having the system search through specific, verified documents before generating a response. Instead of relying solely on patterns learned during training, RAG grounds responses in actual source documents, dramatically reducing hallucinations and enabling the AI to cite its sources.

RAG works in a specific sequence: the system receives your question, converts it into an embedding, searches a knowledge base for the most relevant document passages, provides those passages to the model as context, and then generates a response based on that specific information. The model can cite exactly which documents it drew from.

RAG is not the same as uploading a file to a chatbot. When you upload a PDF to a consumer AI tool, that file is placed into the context window for that single conversation. It disappears when the session ends, cannot scale beyond the context window's token limit, and the AI has no retrieval logic to find the most relevant sections of larger documents. RAG, by contrast, uses embedding-based search across a persistent, governed knowledge base. It retrieves only the most relevant passages, can work across hundreds or thousands of documents, and provides source tracing back to specific materials.

Why it matters for health departments: A RAG-enabled AI connected to your department's policies and procedures can answer "What's our process for temporary food event permits?" by finding and citing your actual documents, rather than generating a plausible-sounding answer based on how other jurisdictions handle it. See also: Knowledge Base.

An approach to AI development and deployment that emphasizes safety, fairness, transparency, accountability, and respect for rights. Responsible AI frameworks guide organizations in managing risks while realizing benefits.

For health departments, responsible AI means having clear policies, training staff, maintaining human oversight, being transparent with communities about AI use, and regularly evaluating whether AI tools are serving your population equitably.

Data that has been artificially created - through statistical modeling, simulation, or AI generation - rather than collected from real-world observations. Synthetic data is designed to match the statistical properties and patterns of real data without containing actual information about real individuals.

In public health and healthcare contexts, synthetic data can enable AI development and testing without exposing real patient information. It can also be used to augment limited datasets or create examples of rare conditions. However, synthetic data inherits limitations from the real data it's based on.

The basic unit that AI language models use to process text. Unlike humans who read letter by letter, LLMs break text into tokens, which might be single characters, parts of words, whole words, or short phrases depending on the text.

Tokens matter because AI models have limits on how many tokens they can process at once (the context window), and many AI services charge based on token usage. As a rough approximation, 1 token ≈ ¾ of a word in English, or about 4 characters.

In practice: A 10-page single-spaced report might be roughly 4,000-5,000 tokens. If your AI tool has an 8,000 token limit, you have room for the report plus your prompt and a substantive response.

The information used to teach an AI model. Large language models are trained on massive text datasets; image AI is trained on millions of images. The training data fundamentally shapes what the model learns: its capabilities, limitations, knowledge, and potential biases.

AI trained on data from one context may not perform well in another. Models trained primarily on clinical healthcare data may not understand governmental public health workflows or local health department terminology. Models trained on data from certain regions may not reflect your state's regulations or community context.

Questions when evaluating AI tools: What data was this trained on? Has it been tested with local health departments or just hospitals and clinics? Does it understand state-specific regulations, or just federal requirements? What are its known limitations for governmental public health use?

The process of assessing AI tools and vendors before your department commits to using them. For government agencies, vendor evaluation goes beyond feature comparison to include security, compliance, data handling, and long-term viability.

Key areas to evaluate for health department AI tools:

Data handling: Where does your data go? Is it stored, used for training, or shared? What happens to data when you stop using the service? Does the vendor offer a Business Associate Agreement (BAA) if you need HIPAA compliance?

Security and compliance: Does the vendor hold relevant certifications (SOC 2, FedRAMP, HIPAA)? How do they handle access controls, encryption, and audit logging?

Public health fit: Was this tool designed for or tested with governmental public health departments, or is it a general-purpose tool being marketed to government? Does it understand the difference between clinical healthcare and governmental public health?

Sustainability: What is the vendor's pricing model? How dependent will you become on this specific tool? What happens if the vendor changes terms or goes out of business?

Action step: Before evaluating any AI tool, establish your department's requirements: What data will it touch? What compliance standards must it meet? What problem are you actually trying to solve? Having clear criteria prevents you from being sold capabilities you don't need.

Using AI to automate repetitive, multi-step processes that previously required manual effort. Workflow automation sits between simple AI assistance (asking a chatbot a question) and full agentic AI (AI that autonomously manages entire projects). It applies AI to specific, well-defined sequences of tasks.

For health departments, workflow automation targets the repetitive administrative work that consumes staff time: formatting reports, routing documents, generating standard correspondence, populating templates, or transforming data between systems. The goal is to free staff to focus on the judgment-intensive work that requires human expertise.

Successful workflow automation starts with process mapping. You need to understand exactly how a workflow operates today before you can identify which steps AI can reliably handle and which require human judgment.

In practice: A health department might automate the workflow of processing incoming food establishment applications: AI extracts key information from submitted forms, pre-populates the tracking system, flags incomplete applications, and drafts the standard acknowledgment email, with a staff member reviewing and approving each step before it goes out.