Lesson 1, Topic 1
In Progress
Unit 1: Teacher Resources
Instructor Resource — Unit 1
Teacher
Resource Guide
Everything you need to facilitate, assess, and differentiate Unit 1 of the AIAS certification — answer keys, rubrics, pacing guides, discussion prompts, misconceptions, and full standards alignment.
Unit 1 at a Glance
Unit 1 establishes the technical and conceptual foundation of the entire AIAS certification. Students move from "what is AI" to designing a complete AI strategy proposal — building the vocabulary, frameworks, and analytical thinking that carry through every subsequent unit.
4
Interactive Lessons
~5 hrs
Total Seat Time
40 pts
Project Value
20
Total Quiz Questions
| Lesson | Title | Time | Key Activity | Assessment |
|---|---|---|---|---|
| 1.1 | Introduction to AI & AI Types | ~30 min | AI Tutor Chatbot + Scenario Cards | 5-question quiz |
| 1.2 | Machine Learning Fundamentals | ~35 min | Overfitting Simulator + Fraud Lab | 5-question quiz |
| 1.3 | NLP, Computer Vision & Robotics | ~40 min | NLP Lab + CV Object Detection Sim | 5-question quiz |
| 1.4 | Emerging AI Trends | ~40 min | Prompt Lab + XAI Demo + Agent Sim | 5-question quiz |
| Project | Unit 1 AI Strategy Proposal | 60–90 min | 10-task written submission + self-rubric | 40 pts instructor-graded |
📋 Delivery Note: All lessons are self-contained HTML pages embedded in LearnDash Custom HTML Blocks. Each includes an interactive AI tutor chatbot, scenario-based knowledge check, and vocabulary section. Students can work fully independently or in facilitated sessions.
Lesson-by-Lesson Instructor Guide
🎯 Lesson Goal: Students leave with a clear mental model of what AI is, what it isn't, and where they already encounter it — anchored by the Narrow AI / Generative AI / AGI framework.
Learning Objectives
- Define Artificial Intelligence and explain why it resists a simple definition
- Distinguish Narrow AI, General AI (AGI), and Generative AI with real-world examples
- Identify AI applications in daily life and classify them correctly by type
- Explain why AGI does not currently exist and what capabilities would be required
Pacing Guidance
5 min
Hero + Objectives
10 min
AI Types + Stat Cards
8 min
Scenario Cards + Chatbot
7 min
Vocab + Knowledge Check
Knowledge Check — Answer Key
Q1BNetflix recommendations are Narrow AI — highly effective at one task but unable to generalize beyond content recommendation.
Q2CCancer diagnosis is a specific, bounded task — the definition of Narrow AI. It doesn't require general reasoning or cross-domain intelligence.
Q3BAGI doesn't exist. LLMs are sophisticated pattern matchers trained on text — they have no goals, self-awareness, or genuine general reasoning.
Q4AGenerative AI creates original content. The other answers describe classification, prediction, and detection — fundamentally different paradigms.
Q5CReal-time language translation is well-defined enough to be Narrow AI today — it solves one task extremely well with no general reasoning required.
Discussion Questions
"If an AI writes a better essay than most humans — is it smarter than humans?"
Push toward: intelligence vs. pattern matching. LLMs predict likely next tokens — they don't understand meaning the way humans do.
"Where does AI already make decisions that affect your life — even if you didn't know it?"
Expected: social feeds, loan approvals, music recs. Stretch: hiring screens, insurance pricing, bail risk scores. Great setup for the ethics unit.
"Why might a company claim 'AI' when it's just rule-based logic? What's the risk?"
Connects to AI washing. Risk: over-trust in systems that aren't actually learning or adapting. Sets up the explainability lesson in 1.4.
Common Misconceptions
| What Students Believe | What to Say Instead |
|---|---|
| "ChatGPT is AGI" | LLMs have no goals, no self-awareness, and fail on novel reasoning outside their training. They're the most impressive Narrow AI ever built — not AGI. |
| "AI always learns and improves" | Most deployed AI systems are static. Learning = the training phase. After deployment, the model is typically frozen unless explicitly retrained. |
| "AI will replace all jobs soon" | Evidence points to role transformation and augmentation. The bigger risk is being replaced by someone who uses AI well — not by the AI itself. |
Differentiation
🤝 Support Strategies
- Focus on Narrow vs. General first — the calculator (narrow) vs. human (adaptable) analogy works well
- Use scenario cards as a sorting activity before the quiz
- Reduce chatbot exploration time; use chip prompts rather than open-ended questions
🚀 Extension Strategies
- Find a current AI news story and evaluate whether the coverage is technically accurate
- Challenge: find a real example of "AI washing" in a product's marketing copy
- Debate: Should AGI development be paused? Research both sides using current sources
🎯 Lesson Goal: Students understand how machines learn from data, can classify ML scenarios correctly, and grasp overfitting as a central challenge — not just a vocabulary term.
Learning Objectives
- Explain how ML differs from traditional rule-based programming
- Distinguish supervised, unsupervised, and reinforcement learning by scenario
- Match key algorithms (regression, decision trees, neural networks, clustering, random forest, Q-learning) to real-world use cases
- Define overfitting and explain the bias-variance tradeoff using the simulator
- Describe algorithmic bias, its causes, and real-world consequences at scale
Pacing Guidance
5 min
What Is ML?
8 min
Three ML Types
7 min
Algorithm Cards
8 min
Overfitting Simulator
7 min
Bias + Fraud Lab + Quiz
⚠️ Simulator Note: Budget at least 5 minutes for students to drag the slider across the full range and find the sweet spot themselves before you explain it. Discovery first, explanation second — the concept sticks far better when students feel it before they're told it.
Knowledge Check — Answer Key
Q1BFinding 50 unlabeled personality clusters = unsupervised learning. No predefined categories, no human labels — the algorithm discovered natural groupings on its own.
Q2B99.8% training vs. 61% test accuracy is the textbook definition of overfitting — the model memorized training quirks rather than learning generalizable patterns.
Q3CAlphaGo learned by playing millions of games with win/loss feedback — reinforcement learning. No human ever labeled an individual move as good or bad.
Q4CThis is algorithmic bias from biased training data. The AI learned historical hiring discrimination and amplified it at scale. More parameters would not fix this — the data is the root cause.
Discussion Questions
"In the simulator — what complexity level gave you the best balance? What was your strategy?"
Look for students watching the gap between training and test accuracy — not just maximizing one number. That's the real insight about model evaluation.
"If a fraud model flags your legitimate transaction — what should happen next? Who decides?"
Bridges to human-in-the-loop design. Explore the tradeoffs between false positive rate and security — there's no clean answer, which is the point.
"Can you think of a real-world dataset that is almost certainly biased? What groups are underrepresented?"
Strong answers: medical imaging skewed by demographics, facial recognition trained on lighter skin tones, historical hiring records skewed by gender.
Common Misconceptions
| What Students Believe | What to Say Instead |
|---|---|
| "Everything is supervised learning" | Key diagnostic: "Was the data labeled by a human first?" No labels → unsupervised. Reward signals → reinforcement. Labels are the dividing line. |
| "Higher accuracy = better model" | Only if measured on the test set. 99% training / 60% test is a worse model than 85% / 83%. Generalization is the goal, not training performance. |
| "Bias can be fixed by adding more data" | More biased data = more bias, not less. Composition and labeling of the data matters far more than volume alone. |
Differentiation
🤝 Support Strategies
- Use the three-question diagnostic for every scenario: labeled? → supervised. No labels? → unsupervised. Reward signals? → reinforcement.
- Pair students for the simulator — verbal explanation while adjusting the slider deepens understanding
- Skip algorithm cards on first pass; return after the overfitting section
🚀 Extension Strategies
- Research the Amazon hiring AI case (2018) and write a 1-page post-mortem: what went wrong, what should have been done differently
- Find a real ML model card on Hugging Face and identify its training data, use cases, and known limitations
- Design a bias audit checklist for the fraud detection lab scenario
🎯 Lesson Goal: Students can explain how NLP, Computer Vision, and Robotics work individually — and how they integrate in real products. The pipeline is the conceptual anchor.
Learning Objectives
- Explain how NLP enables machines to understand and generate human language
- Describe how CNNs process images layer-by-layer from edges to full objects
- Walk through all five steps of the Robotics AI pipeline with real examples
- Define SLAM, sensor fusion, and transfer learning in context
- Identify products that integrate all three AI domains simultaneously
Pacing Guidance
5 min
Intro + Overview
12 min
Domain Tabs
8 min
Robotics Pipeline
8 min
NLP Lab + CV Sim
7 min
Vocab + Quiz
Knowledge Check — Answer Key
Q1BJPMorgan COiN reads and extracts meaning from legal text — Natural Language Processing. Computer Vision processes images, not text documents.
Q2CBuilding a map while simultaneously locating oneself within it is the definition of SLAM. None of the other algorithms address this dual problem.
Q3CTransfer Learning — a CNN pre-trained on millions of general images was fine-tuned on medical scans. This is why it needs far less medical training data than starting from scratch.
Q4CAct + Learn step with Reinforcement Learning — Atlas received reward signals across millions of simulation trials. No human programmed the movements.
Q5CAll three domains: CV to read handwriting, NLP to understand and translate text, Robotics to navigate and deliver medications — the convergence theme of the lesson.
Discussion Questions
"Describe a product using all three AI domains. What breaks if you remove just one?"
Best: self-driving cars, hospital robots, home assistants. Removing one domain typically makes the product nonfunctional — illustrates interdependence.
"In Step 4 of the Robotics pipeline, the AI must choose between two bad outcomes. Who is legally responsible?"
No clean answer — push through: manufacturer, user, regulator, shared? This is the live legal question in autonomous vehicle law right now.
"The NLP lab analyzed your text in real time. Name two beneficial uses — and two dangerous ones."
Beneficial: patient mental health monitoring, student engagement tracking. Dangerous: mass political surveillance, employee monitoring without consent.
Common Misconceptions
| What Students Believe | What to Say Instead |
|---|---|
| "NLP = chatbots" | Chatbots are one application. NLP also powers search engines, medical record analysis, legal document review, translation, and sentiment analysis at enterprise scale. |
| "Self-driving cars just use cameras" | Most systems fuse cameras, LIDAR, radar, and GPS. Cameras fail in fog; LIDAR works in fog. Sensor fusion is the key architectural decision. |
| "Robots are programmed step-by-step" | Modern robots learn movement through RL. Atlas was never programmed to do a backflip — it learned through millions of simulated failures. Programmers set goals, not steps. |
Differentiation
🤝 Support Strategies
- Use domain tabs sequentially — NLP → CV → Robotics builds naturally on prior knowledge
- Have students open each Robotics pipeline step and read the real-world example aloud before moving on
- Pair the NLP lab with a simple personal sentence to keep it low-stakes and relatable
🚀 Extension Strategies
- Research Google's BERT and explain — in a short video or blog post — how attention mechanisms work for a non-technical audience
- Compare Tesla's camera-only FSD vs. Waymo's LIDAR approach and argue for one in a structured debate
- Find a current SLAM algorithm paper and identify one real-world deployment challenge it describes
🎯 Lesson Goal: Students connect emerging trends to the technical foundation built in 1.1–1.3 — and begin thinking strategically about where AI is going, not just where it's been.
Learning Objectives
- Explain Generative AI and LLMs — and distinguish them from prior AI categories
- Describe Edge AI and three model compression techniques (quantization, pruning, distillation)
- Explain XAI and why it is now a legal requirement for high-risk AI deployments
- Classify an AI system under the EU AI Act's four risk tiers with justification
- Describe what an AI agent is and how the ReAct loop enables autonomous action
Pacing Guidance
5 min
Intro + Trend Cards
12 min
5 Trend Panels
8 min
Prompt Lab
8 min
XAI Demo + Agent Sim
7 min
Vocab + Quiz
⚠️ Prompt Lab Note: Students often try to "game" the scorer by keyword-stuffing rather than writing a genuinely useful prompt. Redirect with: "Would a real AI produce something useful from this? That's the real test — not the score."
Knowledge Check — Answer Key
Q1BFace ID processes biometrics entirely on-device with no server transmission — textbook Edge AI. Federated learning involves sharing model updates, which Face ID does not do.
Q2CAttributing a loan denial to specific input features is the core purpose of XAI (SHAP/LIME). GDPR Article 22 legally requires this for automated decisions affecting EU citizens.
Q3CEmotion recognition in workplaces is explicitly listed as an Unacceptable Risk under the EU AI Act — banned outright regardless of claimed benefit.
Q4BAn AI that autonomously searches, books, updates a calendar, and emails — without further input — is an AI Agent using the ReAct loop.
Q5BAdding role, context, audience, constraints, and format to a prompt is the definition of prompt engineering. Transfer Learning involves reusing model weights — not communication with a model.
Discussion Questions
"GenAI can write code and legal briefs. Which cognitive work will remain distinctly human — and why?"
Push past "creativity" — AI is creative. Best answers land on accountability, ethical judgment, relationship trust, and novel cross-domain reasoning under uncertainty.
"The EU AI Act bans emotion recognition at work entirely. Do you agree? What principle guides your answer?"
Strong answers invoke consent, power asymmetry (employer vs. employee), and the difference between voluntary wellness tools and mandatory surveillance.
"An AI agent books a flight, sends an email, and charges your card from one instruction. What could go wrong?"
Target: prompt injection attacks, ambiguous instructions causing cascading errors, accountability gaps — connects prompt quality to real-world consequences.
Common Misconceptions
| What Students Believe | What to Say Instead |
|---|---|
| "GenAI is just a better search engine" | Search retrieves existing documents. GenAI synthesizes and creates new content — a fundamentally different paradigm that also introduces hallucination risk. |
| "Edge AI is just for phones" | Edge AI runs on factory sensors, surgical robots, satellites, and cars. Industrial and safety-critical applications are often the most consequential. |
| "AI agents are chatbots that do more" | Agents use tools to take real actions — sending emails, executing code, booking flights. An agent's output affects the real world, not just a conversation window. |
Differentiation
🤝 Support Strategies
- Start with GenAI and Edge AI only — the two most tangible trends. Add XAI and Agents after students are grounded
- Use the EU AI Act risk pyramid as a sorting activity before the quiz: give 5 AI examples and have students place each in a tier
- For the Prompt Lab, have students explain what makes the strong example better before they write their own
🚀 Extension Strategies
- Read EU AI Act Article 6 and identify one product they use that falls under High Risk — then list the compliance requirements
- Map out a complete AI agent workflow on paper: choose a goal, list tools needed, trace the ReAct loop across 5 steps
- Research one real AI hallucination incident and write a 1-page risk mitigation plan for that deployment
🎯 Project Goal: Students apply every Unit 1 concept to a realistic AI strategy proposal — demonstrating not just recall, but the ability to think like an AI strategist across design, ethics, and business.
Project Structure
| Part | Tasks | Points | Core Concepts Assessed |
|---|---|---|---|
| Part 1 — Design | 4 tasks | 16 pts | Problem definition, AI/ML classification, data strategy, algorithm selection |
| Part 2 — Risk & Ethics | 3 tasks | 14 pts | Algorithmic bias, overfitting prevention, human-in-the-loop design |
| Part 3 — Business Case | 3 tasks | 10 pts | EU AI Act compliance, business value quantification, executive communication |
Instructor Grading Rubric
| Criterion | Pts | 1 — Beginning | 2 — Developing | 3 — Proficient | 4 — Exemplary |
|---|---|---|---|---|---|
| Problem Definition | /4 | Vague; no measurable outcome | Problem identified; solution unclear | Clear specific problem + named solution | Specific, measurable problem; precisely named solution with clear scope |
| Technical Classification | /4 | Missing or incorrect | Types named; justification weak | All 4 elements correct with basic justification | All 4 correct with detailed, logically connected reasoning |
| Data Strategy | /4 | Data not described or unrealistic | Data type identified; risks absent | Data type, volume, and one risk addressed | Full strategy: type, volume, source, multiple risks, mitigation ideas |
| Algorithm Choice | /4 | Doesn't match problem type | Reasonable; no comparison made | Good choice with justification | Best-fit algorithm justified against at least one alternative |
| Bias Identification | /4 | No specific risks identified | One generic risk named | Two specific risks with audit approach | Two+ risks with audit methods, affected populations, and mitigation |
| Overfitting Plan | /4 | No generalization strategy | Train/test split mentioned only | Clear plan with monitoring approach | Comprehensive: split, drift detection, retraining triggers, metrics |
| Human Oversight | /4 | Human role not addressed | Involvement mentioned generically | Specific escalation scenario with threshold | Full design: scenario, threshold, reviewer role, feedback loop |
| EU AI Act Classification | /4 | Tier missing or clearly wrong | Correct tier; no requirements listed | Correct tier with 2+ requirements | Correct tier, full requirements, impact on solution design addressed |
| Business Value | /4 | No quantification attempted | Vague estimate; no reasoning | Specific estimate with logical basis | Compelling estimate tied to real costs with comparable references |
| Executive Summary | /2 | Missing or off-topic | Present but uses jargon or is incomplete | Clear; covers all four required elements | Executive-ready; no jargon; persuasive and precise throughout |
* Executive Summary is scored out of 2 points. All other criteria are out of 4. Total = 40 points.
Grade Scale
36–40
🏆 Exemplary
Portfolio-ready. Genuine AI strategy thinking beyond course material.
28–35
✅ Proficient
Strong command of Unit 1 concepts. A few areas need deeper reasoning.
20–27
📝 Developing
Core concepts present but surface-level. Revisit weaker sections before Unit 2.
0–19
🔁 Beginning
Lesson content review recommended before resubmission.
Grading Tips
Part 1 — Red Flag: Students writing "supervised learning because AI learns from data" — that's not a justification. Push for: what are the labeled inputs? What are the target outputs? Why would unsupervised not work?
Part 2 — Proficient vs. Exemplary: Proficient students name bias risks. Exemplary students name which population is underrepresented, describe what a false positive vs. false negative means for a real person, and explain which error type is more harmful in this specific context.
Part 3 — Executive Summary Test: Read it aloud. If it would confuse a non-technical VP, it needs revision. Strong summaries: opportunity → primary risk → mitigation → recommendation. No acronyms without explanation.
Differentiation
🤝 Support Strategies
- Allow Healthcare or Financial Services — both have abundant lesson examples to reference
- Provide sentence starters for Part 1 Task 2: "This uses supervised learning because the training data includes labeled [X] paired with [Y]…"
- Allow the executive summary as a bulleted brief for students who struggle with professional writing
🚀 Extension Strategies
- Add a Failure Mode Analysis: name the three most likely production failures and their early warning signals
- Present the executive summary as a 3-minute verbal pitch to a peer playing a skeptical CFO
- Research a real competitor who deployed a similar solution and use it to strengthen or challenge the business value estimate
Standards Alignment
Unit 1 aligns with the Knowledge Pillars AIIAS Exam Objectives as the primary framework, and maps to AP Computer Science Principles (AI strand) and ISTE Standards for Students as supporting frameworks.
Knowledge Pillars AIIAS Exam Objectives — Primary Alignment
| Exam Domain | Unit 1 Coverage | Lessons |
|---|---|---|
| Domain 1 Understanding AI Technologies & Trends |
Full coverage. Students learn fundamental AI concepts and algorithms (supervised learning, neural networks, NLP, CNNs, decision trees, linear regression), differentiate AI types (narrow, general, generative), identify emerging trends (generative AI, edge AI, autonomous systems), and analyze AI impact across healthcare, finance, manufacturing, and retail. | 1.1, 1.2, 1.3, 1.4, Project |
| Domain 2 Strategic AI Integration |
Introduced in the Unit 1 Project. Students design AI-based solutions for specific use cases, select appropriate ML approaches (supervised, unsupervised, reinforcement, transfer learning), and identify AI applications in finance, healthcare, retail, and manufacturing. Full coverage continues in Unit 2. | 1.4, Project |
| Domain 6 Ethical & Regulatory Considerations |
Introduced throughout Unit 1. Lesson 1.2 covers algorithmic bias, regularization, overfitting, and generalization. Lesson 1.4 covers EU AI Act risk tiers, GDPR's right to explanation, XAI (explainability), and responsible AI principles. Full coverage in Unit 2. | 1.2, 1.4, Project |
| Domain 7 Client Communication & Advisory Skills |
The Unit 1 Project's executive summary task directly develops the ability to present technical AI concepts to non-technical audiences and build a persuasive business case. Full coverage in later units. | Project (Part 3) |
📌 Domain 1 AIIAS Exam Terms Directly Taught in Unit 1: Supervised Learning, Unsupervised Learning, Reinforcement Learning, Neural Networks, Convolutional Neural Networks (CNNs), Linear Regression, Decision Trees, Transfer Learning, Natural Language Processing, Computer Vision, Image Recognition, Data Clustering (K-Means), Narrow AI, General AI, Generative AI, Bias, Regularization, Overfitting, Generalization, Edge AI, Autonomous Systems, Pathfinding Algorithms (Robotics Pipeline).
AP Computer Science Principles (AP CSP) — AI Strand
Big Idea 5: Impact of Computing
AP CSP Framework
5.3Computing has impacted collaboration and communication — Lessons 1.3 (NLP, chatbots) and 1.4 (GenAI, AI agents)
5.4Programs can be used for harm — Lesson 1.2 (algorithmic bias), Lesson 1.4 (deepfakes, EU AI Act)
5.5Legal and ethical concerns — Lesson 1.2 (Amazon hiring AI bias case), Lesson 1.4 (EU AI Act risk tiers, GDPR)
Big Idea 2: Data
AP CSP Framework
2.3Programs use data — Lesson 1.2 (training vs. test data, supervised/unsupervised learning) and Project (data strategy)
2.4Using programs with data — Lesson 1.2 (fraud detection lab), Lesson 1.4 (XAI decision demo)
Big Idea 4: Algorithms & Programming
AP CSP Framework
4.1Algorithms solve problems — Lesson 1.2 (ML algorithms; decision trees, regression, neural networks)
4.2Evaluating algorithms — Lesson 1.2 (overfitting simulator: evaluating model accuracy on training vs. test data)
ISTE Standards for Students
1. Empowered Learner
ISTE Standard 1
1cStudents use technology to seek feedback — AI tutor chatbots in all four lessons provide real-time adaptive feedback on student questions
1dStudents understand fundamental technology concepts — all four lessons build explicit mental models of how AI systems work
2. Digital Citizen
ISTE Standard 2
2bPositive, safe, legal, and ethical behavior — Lessons 1.2 (bias) and 1.4 (EU AI Act, ethics of autonomous systems)
2dManage digital identity and privacy — Lesson 1.4 (Edge AI, data privacy, federated learning)
3. Knowledge Constructor
ISTE Standard 3
3aPlan and employ effective research strategies — all lessons model how AI systems gather and evaluate information from data
3cCurate information from digital resources — Lesson 1.4 (prompt engineering teaches precise information retrieval from AI systems)
4. Innovative Designer
ISTE Standard 4
4aUse a deliberate design process — Unit 1 Project (Part 1) requires designing an AI solution through defined technical and strategic steps
4dExhibit tolerance for ambiguity — Project (Parts 2–3) require reasoning through genuinely ambiguous bias, ethics, and value questions
5. Computational Thinker
ISTE Standard 5
5bCollect and use data sets — Lesson 1.2 (NLP and fraud labs simulate data collection and pattern recognition pipelines)
5cBreak problems into component parts — Project (Part 1) requires decomposing an AI solution into algorithm, data, and deployment components
5dUnderstand automation — Lesson 1.4 (AI agents, ReAct loop, autonomous decision-making)
7. Global Collaborator
ISTE Standard 7
7aConnect with learners from varied backgrounds — chatbot activities model cross-cultural AI communication challenges (bias across populations, NLP translation)
7cContribute to project teams — Unit 1 Project and discussion questions are designed for collaborative learning formats
Unit 1 Master Vocabulary List
All 28 terms introduced across the four Unit 1 lessons. Use for vocabulary quizzes, pre/post-unit assessment, or review activities.
| Term | Lesson | Definition |
|---|---|---|
| Narrow AI | 1.1 | AI designed for one specific task — the type powering virtually every AI product in use today |
| Artificial General Intelligence (AGI) | 1.1 | Hypothetical AI with human-level reasoning across all domains — does not currently exist |
| Generative AI | 1.1 / 1.4 | AI that creates original content (text, images, code, audio) by learning patterns from training data |
| Machine Learning | 1.2 | AI that learns patterns from data rather than following explicitly programmed rules |
| Supervised Learning | 1.2 | ML training on labeled data — inputs paired with correct outputs — to predict on new unseen examples |
| Unsupervised Learning | 1.2 | ML that finds hidden structure in unlabeled data — clusters, patterns, and anomalies without predefined answers |
| Reinforcement Learning | 1.2 | An agent learns by taking actions and receiving rewards or penalties, maximizing long-term reward through trial and error |
| Overfitting | 1.2 | When a model memorizes training data — including noise — and fails to generalize to new data |
| Generalization | 1.2 | A model's ability to perform well on data it was never trained on — the ultimate goal of every ML project |
| Algorithmic Bias | 1.2 | Systematic unfairness in AI outputs caused by biased or unrepresentative training data |
| Regularization | 1.2 | Techniques that reduce overfitting by penalizing excessive model complexity |
| Natural Language Processing (NLP) | 1.3 | AI that enables machines to understand and generate human language — text and speech |
| Transformer Architecture | 1.3 | Neural network design using attention mechanisms that process entire sequences simultaneously — powers GPT, BERT, and all modern LLMs |
| Named Entity Recognition (NER) | 1.3 | NLP technique that identifies and classifies people, places, organizations, and dates in unstructured text |
| Computer Vision (CV) | 1.3 | AI that extracts meaning from images and video — detection, classification, segmentation, and tracking |
| Convolutional Neural Network (CNN) | 1.3 | Deep learning architecture that processes images layer-by-layer from edges to full object recognition |
| LIDAR | 1.3 | Sensor that maps 3D environments using laser pulse timing — critical for autonomous vehicle navigation |
| SLAM | 1.3 | Simultaneous Localization and Mapping — enables a robot to build a map while tracking its own position within it |
| Sensor Fusion | 1.3 | Combining multiple sensor streams into a unified world model for reliable autonomous operation |
| Transfer Learning | 1.3 | Reusing a model pre-trained on one task as a starting point for a new related task, requiring far less training data |
| Large Language Model (LLM) | 1.4 | Transformer-based neural network trained on massive text datasets — GPT-4, Claude, and Gemini are examples |
| Prompt Engineering | 1.4 | Crafting precise inputs to guide an AI model toward desired outputs using role, context, constraints, and output format |
| Edge AI | 1.4 | AI computation on local devices rather than cloud servers — enabling privacy, offline operation, and low latency |
| Federated Learning | 1.4 | Privacy-preserving training where devices share only encrypted model updates, never raw user data |
| Explainable AI (XAI) | 1.4 | Techniques (SHAP, LIME) that attribute AI model decisions to specific input features, making outputs understandable and auditable |
| AI Agent | 1.4 | Autonomous AI that perceives its environment, plans strategies, uses tools, and takes real-world actions to achieve goals |
| Hallucination | 1.4 | When a generative AI model confidently produces false or fabricated information as if it were fact |
| EU AI Act | 1.4 | The world's first comprehensive AI regulation law (2024), using a four-tier risk-based framework from unacceptable to minimal risk |
Frequently Asked Questions
Yes — quiz retakes are available by design. The knowledge checks are formative tools, not high-stakes assessments. The Unit 1 Project is the summative assessment. We recommend allowing unlimited quiz attempts; the explanations for incorrect answers are part of the learning experience.
This is an AI certification — restricting AI use is counterproductive. Instead, set expectations: AI should be used as a thinking partner and editor, not a ghostwriter. Require students to explain and defend every claim verbally if you suspect uncritical AI generation. The rubric rewards specific reasoning and personal justification, which raw AI output typically lacks.
60–90 minutes for a student who completed all four lessons carefully. Students who rushed the lessons will need longer. We recommend positioning the project as a take-home assignment with a 3–5 day window — not an in-class timed exercise. This gives students time to choose an industry they care about and produce responses they're genuinely proud of.
Lean in — creative industry choices often produce the most engaged submissions. The rubric rewards rigorous thinking, not safe choices. A student designing AI for underwater archaeology can still demonstrate perfect technical classification, solid bias analysis, and a compelling executive summary. Redirect only if the proposal is so abstract that real technical reasoning becomes impossible.
Yes — and it's a good teaching moment. The NLP lab, CV simulator, fraud detection lab, and agent simulator are rule-based simulations designed to demonstrate concepts, not live AI API calls. This reinforces the Lesson 1.1 point that not everything labeled "AI" uses machine learning. Real AI integration is available through the chatbot tutors, which do connect to a live model.
No technical prerequisites. Unit 1 is designed as the entry point for students with no prior AI or programming experience. Lesson 1.1 builds vocabulary from scratch. The only assumed skills are high-school-level reading comprehension and willingness to engage with conceptual thinking. Students with prior CS or data science experience should be directed to the Extension Strategies to stay challenged.