Learning Objectives

After completing this lecture, you will be able to:

• Understand the key concepts of types of ai systems in a government context

• Connect types of ai systems to your agency's AI initiatives

• Identify next steps for applying these concepts in your role

Key Topics Covered

• Narrow AI, general AI, generative AI, agentic AI
• Classification systems and where each type appears in government
• Government context for types of ai systems
• Practical applications and next steps

Why This Matters for Government

Government agencies face unique challenges when it comes to AI adoption. This lecture addresses these challenges head-on by providing all government employees with the knowledge and frameworks needed to navigate AI in the public sector responsibly and effectively.

As part of the L1 (AI Aware) curriculum, this lecture builds on the foundational principle that every AI system in government ultimately serves citizens. Whether you are working with AI tools daily or setting strategy for your agency, understanding types of ai systems is essential for responsible, effective government AI adoption.

======================================================================

TRANSCRIPT: Types of AI Systems

======================================================================

What you will learn: Narrow vs. general AI. Generative systems vs. classification systems. Agentic vs. reactive systems. Where each type exists in government today.

We know now that AI is pattern matching from data. But AI systems vary enormously in scope and capability. Some can do exactly one thing—recognize tax forms. Others can write essays, debug code, and answer questions about nearly any topic. Some follow instructions. Others take initiative and set their own goals.

This lecture classifies the landscape. When someone says "we're adopting AI," you'll be able to ask: what KIND of AI? Because the type determines what it can do, what it can't do, and what governance it requires.

Purpose

The AI landscape is divided along a few key dimensions: scope (how many domains can the system operate in?), capability (what tasks can it accomplish?), and autonomy (does it respond to requests or take initiative?). Understanding these dimensions lets you evaluate systems more accurately.

Why This Matters for Government

Different AI types create different risks and opportunities. A narrow classifier that approves/denies routine documents is relatively low-risk if human-reviewed. An agentic system that autonomously restructures government processes is a governance nightmare without clear controls. Understanding the type is the first step to appropriate oversight.

Core Concepts

1. Narrow AI vs. General AI vs. Super-Intelligence

These terms describe the scope of what a system can do.

Narrow AI (Also Called "Weak AI")

A narrow AI is specialized. It does one domain well. Image recognition for documents. Translation between specific languages. Fraud detection in claims. It excels in its domain but is utterly useless outside it.

All AI systems in existence today are narrow AI. ChatGPT might seem general because it can do many language tasks, but it's still fundamentally a text-prediction system trained on internet data. Ask it about quantum mechanics, and it will generate plausible-sounding text that might be completely wrong. Ask a medical AI system to predict document types, and it will fail.

Narrow AI is actually good for government. Why? Because it's easier to understand, test, and control. A system trained to do one thing has one purpose. You can measure whether it does that purpose well.

General AI (Also Called "Strong AI" or "Artificial General Intelligence")

A general AI would understand concepts across domains the way humans do. It would learn from experience. It would transfer knowledge from one domain to another. It doesn't exist. Researchers are working on it. Some think it's decades away. Some think it's impossible.

When someone says "AI will replace all human workers," they're imagining general AI: a system that can learn any human job. That's not where we are. We're in the narrow AI era. Each system is specialized.

Super-Intelligence

A hypothetical system smarter than humans at all tasks. Science fiction territory right now. Mentioned because you'll hear it in policy discussions. "We need to prepare for super-intelligence." Fair enough—long-term thinking is good. But it doesn't affect today's government AI decisions.

Practical Implication: Don't worry about general or super-intelligence. Focus on the narrow AI systems your agency is actually deploying. Evaluate their specific capabilities and limitations.

1. Generative vs. Classification Systems

These terms describe what the system outputs.

Classification Systems

A classifier takes input (a document, an image, a transaction) and assigns it to a category. "This is fraud or not fraud." "This is a tax return or correspondence." "This is a high-risk or low-risk case."

Classifiers are relatively mature. They work well. They're good for government because:

• They're interpretable: you can often understand why the system made a decision

• They fit government workflows: human staff makes the actual decision; the system provides input

• They're falsifiable: you can test whether the classifier is accurate

Examples in government:

• A system that flags potentially fraudulent welfare claims for human review

• An email classifier that sorts incoming requests by department

• A system that identifies high-priority cases for faster processing

Generative Systems

A generative system creates new content: text, images, code, audio, video. ChatGPT writing an essay. DALL-E generating an image. A system composing a government letter.

Generative systems are newer and trickier for government because:

• They're less interpretable: you can't easily explain why an essay was written the way it was

• They can hallucinate: they generate plausible-sounding content that might be false

• They create new liabilities: who's responsible if a generated government letter is inaccurate?

• They require strong human oversight: the output needs review before going public

Generative systems are useful for government but require careful governance. A system that generates a first draft of a briefing paper, which a human then edits and approves? Good. A system that autonomously generates and sends government communications without human review? Bad.

1. Reactive vs. Agentic Systems

These terms describe how much autonomy the system has.

Reactive Systems

A reactive system responds to inputs. You give it a request; it processes it and returns a result. The system doesn't set its own goals or take initiative.

Almost all AI systems today are reactive. The LLM you ask a question responds to your question. The classifier you submit a document to classifies that document. The system can't decide "I'll go reorganize the filing system" without someone asking it to.

Reactive systems are fine for government because the human still controls the agenda. The human decides what to ask the system to do. The system executes. This maintains human oversight and responsibility.

Agentic Systems

An agentic system has autonomy. It's given a goal, and it decides how to pursue it. It might break the goal into sub-goals. It might create plans. It might execute steps without asking for permission each time.

Agentic systems are emerging. An agentic system might be: "Your goal is to reduce permit processing time. Here's your budget and authority. Execute." The system might analyze current processes, identify bottlenecks, propose new procedures, implement changes, and report back. It's taking initiative, making decisions, executing plans.

For government, agentic systems are governance challenges. Who approves the sub-goals? What authority does the system have? When must it escalate to a human? What if it makes a decision that harms someone?

These questions require clear governance frameworks. Right now, most government AI policy assumes reactive systems. Agentic systems will require new oversight mechanisms.

1. Rule-Based vs. Learning-Based

These terms describe how the system works.

Rule-Based Systems

A rule-based system executes human-defined rules. If temperature is below 32 degreesF, classify as "freezing." If income is below X, approve for benefits. If risk score is above Y, flag for review.

Rule-based systems are interpretable: you can see the rules. They're reliable: they always follow the same logic. But they're brittle: they only work if you've anticipated the rules needed. If a new situation arises that you didn't write a rule for, the system fails.

Learning-Based Systems (Machine Learning)

A learning-based system learns patterns from data. You don't write rules; you provide examples, and the system finds patterns.

These are more flexible: they can adapt to new situations you didn't anticipate. But they're less interpretable: even the designers can't explain why the system made a specific decision. And they require good training data: if your examples are biased, the system will be biased.

Most modern AI systems are learning-based because they're more powerful. But this creates explainability challenges for government.

Hybrid Approaches

Smart government implementations combine both. Rule-based systems for decisions that are well-understood and stable (this is clearly benefits-eligible or not). Learning-based systems for nuanced predictions (which applications are likely to be approved given this data?). Human judgment for edge cases and high-stakes decisions.

Practical Use Cases

Case 1: IRS Form Classification (Narrow, Reactive, Learning-Based)

An optical character recognition (OCR) system combined with a document classifier. Given an image of a tax document, output "Form 1040," "Form 1099," "Form W2," "Amended Return," etc.

This is narrow: it only classifies tax documents, not everything. Reactive: it responds to document submission. Learning-based: trained on examples of documents.

Result: Mail-sorting, initial triage. Human staff still reviews and processes. The system speeds up work; it doesn't replace judgment.

Case 2: Welfare Fraud Detection (Narrow, Reactive, Learning-Based)

A system trained on historical claims: approved claims and detected fraud. Given a new claim, predict: "likely fraudulent" or "likely legitimate."

This is narrow: just fraud detection. Reactive: it responds to new claims. Learning-based: patterns from historical data.

Result: Flags for human investigation. A human caseworker does the actual investigation and determination. The system prioritizes work.

Governance requirement: Regular testing for bias. Audit that the system isn't disproportionately flagging claims from certain demographics. Appeals process for people whose claims were flagged.

Case 3: Environmental Monitoring (Narrow, Reactive, Learning-Based)

Satellite imagery + computer vision. Identify coastal erosion, illegal dumping, deforestation, etc. from satellite images.

Narrow: specific environmental phenomena. Reactive: processes imagery as it comes. Learning-based: trained on labeled imagery.

Result: Alerts for human scientists to investigate. Monitoring at scale that would be impossible manually.

Case 4: Benefits Chatbot (Broad-ish, Reactive, Generative, Learning-Based)

A language model trained on government benefits policies and FAQs. Citizens ask questions; the system provides answers.

This is relatively broad (covering benefits across many programs). Reactive (responds to questions). Generative (creates new answers, not just retrieves stored ones). Learning-based.

Governance challenge: The system might generate incorrect information. Solution: Human review of generated answers before deployment. Regular audits of common questions and answers. Clear disclaimer that citizens should verify with official resources.

Case 5: Procurement Automation (Narrow-to-Broad, Potentially Agentic, Rule-Based and Learning-Based)

A hypothetical system: given a procurement need, it researches vendors, drafts RFPs, manages the bidding process, and recommends winners. Goals: reduce procurement time and cost.

This is broad (procurement across many categories). Potentially agentic (it takes initiative to find vendors, reach out, manage process). Hybrid (rule-based for defined processes, learning-based for recommendation).

Governance challenge: Who approves major decisions? What's the escalation process? How do we ensure fairness? Vendor concerns about AI evaluation? This would require significant oversight.

Current status: Mostly ruled out for core decisions without heavy human involvement. Possible for administrative task acceleration (scheduling meetings with potential vendors), but humans make the real decisions.

Anti-Patterns / Misuse Risks

Anti-Pattern 1: Deploying Agentic Capabilities Without Governance

Risk: A system that takes initiative making consequential decisions without clear human approval processes.

Why It Happens: It seems efficient. The system can identify problems and fix them. Why wait for human review?

What Goes Wrong: The system automatically adjusts staffing or resource allocation based on its analysis. It makes a decision that harms a specific group (e.g., reduces services in a particular neighborhood) without anyone noticing until it's happened. No one is accountable because "the system decided."

How to Avoid: For agentic systems, define clear boundaries. What decisions can the system make autonomously? (Usually: none in government.) What requires human approval? (Most decisions). What requires external approval? (Anything affecting citizens). Have escalation paths. Monitor what the system actually does.

Anti-Pattern 2: Assuming a Classifier is Objective

Risk: Using a learned classifier as if it's neutral, without checking for bias or unfairness.

Why It Happens: It's a mathematical model. Models seem objective.

What Goes Wrong: A classifier learned from historical data that contains biases. It perpetuates those biases now, cloaked in mathematical authority. People trust "the algorithm" over their own judgment.

Example: A hiring classifier trained on past hiring decisions (which reflected discrimination) learns to replicate discrimination. It systematically disadvantages certain groups.

How to Avoid: Audit classifiers for fairness. Test on different demographic groups. Compare the classifier's decisions to human expert decisions. Have humans in the loop for high-stakes choices.

Anti-Pattern 3: Using a Generative System for Official Government Statements Without Review

Risk: Publishing AI-generated letters, guidance, or official statements without human review.

Why It Happens: It's fast. The system produced something that looks professional.

What Goes Wrong: The system generated official-sounding but inaccurate guidance. Citizens followed it. Legal problems ensue.

Example: An agency automated correspondence with citizens using an LLM. The LLM generated a response about benefits eligibility that sounded authoritative but was factually wrong. Citizens relied on it.

How to Avoid: Human review for all external communications. Especially for anything that looks official or could affect a citizen's decisions. Use generative systems for drafting, not publishing.

Practice / Reflection Prompts

• System Classification: Find an AI system your agency uses or is considering. Classify it along these dimensions: narrow or broad? Generative or classification? Reactive or agentic? Learning-based or rule-based? What does each classification tell you about governance requirements?

• Governance Exercise: For each type of system (generative classifier, agentic, etc.), draft governance requirements. Who reviews? When? What decisions require human approval?

• Bias Test Design: If your agency deployed a classification system, how would you test for bias? What would "fair" performance look like? How would you know if the system was working unfairly?

• Agentic Boundary Setting: Imagine an agentic system to improve your agency's processes. What should it be allowed to decide independently? What requires human approval? What requires public notice?

• Type Evolution: Many systems start narrow and generalist pressures push them toward broader use. Identify a system in your agency. How might pressure to expand its use arise? What safeguards would you want?

Key Takeaways

• All AI in existence today is narrow AI—specialized to specific domains, not general intelligences capable of learning any task.

• Classification systems assign inputs to categories—better for government than generative systems because they're more interpretable and fit government decision-making workflows.

• Generative systems create new content—useful for drafting, but require strong human oversight and review before anything goes public.

• Reactive systems respond to requests; agentic systems take initiative—government needs clear governance around agentic systems that make consequential autonomous decisions.

• Learning-based systems are more flexible than rule-based but less interpretable—you gain capability but lose explainability.

• The type of system determines governance requirements—a narrow reactive classifier needs different oversight than a broad agentic system.

• Classification systems can still be biased—even though they're more interpretable, they learn from biased data and perpetuate those biases.

Terms / Glossary Items

Narrow AI: A specialized AI system that performs well in one domain but is useless outside it. All AI systems today are narrow.

General AI: A hypothetical AI that understands and learns across domains like humans do. Doesn't exist yet.

Classification System: An AI that assigns inputs to categories (approve/deny, urgent/non-urgent, etc.).

Generative System: An AI that creates new content (text, images, code, audio, video).

Reactive System: An AI that responds to inputs without taking initiative.

Agentic System: An AI that has autonomy, sets goals, makes decisions, and takes action without explicit human direction for each step.

Rule-Based System: An AI that executes human-defined rules (if X, then Y). Interpretable but brittle.

Learning-Based System: An AI trained on examples to find patterns. Flexible but less interpretable.

You can now classify any AI system you encounter. This matters because the classification determines what questions to ask:

• Narrow classifier: Is it accurate in its domain? Is it biased?

• Generative system: What human review process exists? Has it been tested for accuracy?

• Agentic system: What are the approval boundaries? What happens when it fails?

Next lecture: what AI does well and where it fails. You'll learn to match tasks to AI capabilities.

Think of a routine decision in your agency: approving something, denying something, prioritizing. Could an AI system help? What TYPE of system would be appropriate?

• Classification system flagging cases for human review?

• Generative system drafting correspondence?

• Something else?

What safeguards would you want in place?

You've now classified the AI landscape. Every system fits into categories, and each category has its strengths and governance challenges.

Next: What AI does well and where it fails. We'll use this knowledge to predict which government tasks are good candidates for AI and which are not.

Government AI CLUB Certification Program

Level 1: AI Aware | What AI Is and Is Not | Lecture 1.1.3

A GOVT.CLUB initiative.

<- 1.1.2 How AI Actually Works 1.1.4 What AI Does Well and Where It Fails ->

Start Your CLUB Certification

This lecture is part of L1: AI Aware—8 hours of comprehensive government AI training.

Explore CLUB Certification

Related Lectures

L1 1.1.1—What AI Is and Is Not 20 min - Video + Reading

L1 1.1.2—How AI Actually Works 20 min - Video + Interactive

L1 1.1.4—What AI Does Well and Where It Fails 20 min - Video + Case Studies

Frequently Asked Questions

What will I learn in Types of AI Systems?

In this 20 min video + reading lecture, you will Narrow AI, general AI, generative AI, agentic AI. Classification systems and where each type appears in government

What level is Types of AI Systems?

This is a Level 1 (AI Aware) lecture, part of Chapter 1.1 \u2014 AI Foundations. It is designed for all government employees.

How long is lecture 1.1.3?

Lecture 1.1.3 (Types of AI Systems) takes 20 min. It is delivered as a video + reading format.

Do I need prerequisites for Types of AI Systems?

This lecture is part of L1 (AI Aware). Prerequisites: None.

What is the CLUB Certification?

CLUB (Community Leading Unified Benchmarks) is a maturity-based AI certification for government professionals with 5 levels (L1-L5), 215 lectures, and 25 chapters aligned with NIST AI RMF, OMB, and GAO frameworks.