What Is Artificial Intelligence?
Module 1
After completing this module, you will be able to:
- Explain what artificial intelligence is — and what it is not
- Identify AI systems you already interact with daily
- Distinguish between narrow AI, general AI, and generative AI
- Describe the key players in the AI landscape and what they build
- Articulate both the capabilities AND limitations of current AI — honestly
- Explain why AI sometimes gives “technically correct but contextually wrong” answers
- Recognize the cognitive offloading trap and why verification is the #1 AI skill
The Lawyer, the Chatbot, and Six Cases That Never Existed
In March 2023, attorneys at the New York firm Levidow, Levidow & Oberman filed a brief in federal court on behalf of their client, Roberto Mata, who was suing the airline Avianca. The brief cited six previous court cases as precedent — Varghese v. China Southern Airlines, Shaboon v. Egyptair, Petersen v. Iran Air, Martinez v. Delta Airlines, Durden v. KLM, and Miller v. United Airlines — detailed, convincing, with case numbers, dates, and judges’ names. It looked professional. It looked authoritative.
There was one problem: none of the six cases were real.
Steven Schwartz, the attorney who wrote the brief, hadn’t fabricated them. He had asked ChatGPT to help with legal research, and the AI had generated six completely fictitious court cases — complete with plausible-sounding citations — and presented them as fact. When opposing counsel couldn’t locate the cases, the judge couldn’t either. When Schwartz asked ChatGPT whether Varghese was a real case, the chatbot doubled down: yes, it said, and it could be found on Westlaw and LexisNexis. On June 22, 2023, Judge P. Kevin Castel sanctioned Schwartz, his colleague Peter LoDuca (who had signed the brief without verifying the research), and the firm itself — a $5,000 fine and a requirement to write apology letters to every real judge whose name had been falsely attached to a fabricated opinion. Mata v. Avianca (678 F.Supp.3d 443) became a cautionary tale for a profession scrambling to understand a technology that writes like a lawyer but “thinks” like nothing humans have ever encountered.
This case illustrates the central tension of artificial intelligence in 2026: AI can produce text that is fluent, confident, and entirely wrong. It can write poetry, debug code, translate languages, summarize research, and generate images that look like photographs — but it can also “hallucinate” facts, fabricate citations, and confidently present nonsense — and can even fabricate the verification that those facts are real.
Understanding this tension — what AI can do, what it cannot do, and what it might pretend to do — is the foundation of AI literacy. That’s what this module is about.
By the end of this course, you will be able to use AI powerfully, critically, and ethically. But it all starts here: understanding what AI actually is, how it works at a basic level, and where it fails. If you skip this module, every other module is built on sand.
AI Is Already in Your Life
Before we define AI formally, let’s acknowledge something: you’ve been using AI for years. You just didn’t call it that.
- When Netflix recommends a show you actually want to watch — that’s AI.
- When your email filters spam into the junk folder — that’s AI.
- When Google Maps reroutes you around traffic — that’s AI.
- When your phone recognizes your face to unlock — that’s AI.
- When autocomplete suggests the next word in your text message — that’s AI.
- When your bank flags a suspicious charge and texts you — that’s AI.
AI isn’t a single technology. It’s a family of technologies that share one trait: they can perform tasks that used to require human intelligence. Some of these tasks are simple (spam filtering). Some are complex (translating a conversation in real time). Some are creative (generating an image from a description). But they all share the same underlying idea.
AI in YOUR Field Right Now
- AI reads chest X-rays and flags potential tumors for radiologist review
- Chatbots triage patient symptoms before they reach the nurse’s desk
- Clinical documentation AI drafts visit notes from doctor-patient conversations
- Drug interaction checkers use ML to catch dangerous combinations
- CRM systems predict which leads are most likely to close
- Chatbots handle first-line customer service inquiries 24/7
- Inventory systems predict demand and auto-reorder stock
- AI drafts marketing copy and A/B tests subject lines automatically
- Adaptive learning platforms adjust difficulty in real time per student
- AI grading tools provide instant feedback on writing assignments
- Custom GPTs serve as 24/7 tutoring assistants (the ELEVATE model)
- Plagiarism and AI detection tools scan submissions (with known limitations)
- Drones with computer vision scan fields for pest damage and disease
- Soil sensors + ML optimize irrigation schedules in real time
- Satellite imagery predicts crop yields months before harvest
- AI-powered sorting machines grade produce by size, color, and quality
- Fraud detection algorithms flag suspicious benefit claims for review
- 311 chatbots answer citizen questions about services and hours
- Predictive models allocate emergency response resources during disasters
- Document processing AI extracts data from permit applications
- AI generates concept art, logo drafts, and mood boards in minutes
- Music AI composes background scores for video content
- Writing assistants help journalists draft, edit, and fact-check stories
- Video AI creates b-roll and transitions for social media content
- “AI thinks” — AI processes patterns statistically. It doesn’t think, understand, or have intentions. It generates probable outputs based on training data.
- “AI is always right” — AI is confidently wrong a significant percentage of the time. The lawyer case in our hook proves this dramatically.
- “AI is new” — The field is 70 years old (1956). What’s new is the data, hardware, and algorithms that made it practical for everyday use.
- “AI = robots” — Most AI has no physical form. It’s software — code running on servers, processing text, images, and data.
- “AI will replace everyone” — AI automates tasks, not entire jobs. Most roles will be augmented, not eliminated. The nurse who uses AI is still a nurse.
Think about your last 24 hours. List three moments where you interacted with an AI system — even if you didn’t realize it at the time. For each, describe: (a) what the AI did, (b) what data it probably used, and (c) whether you could have done the same task yourself (and how long it would have taken).
Click to see example answers
Example 1: My email sorted 14 messages into “Primary” and “Promotions” tabs. (a) It classified emails by type. (b) It probably used the sender, subject line, content, and my past behavior (which emails I open vs ignore). (c) I could do it myself, but it would take 5 minutes every morning instead of 0 seconds.
Example 2: Spotify’s “Discover Weekly” gave me a playlist of songs I’d never heard. (a) It recommended music based on my listening history. (b) It used my past plays, skips, likes, and similar users’ behavior. (c) I could browse music myself, but I wouldn’t find songs from obscure artists I’ve never encountered.
Example 3: Google autocompleted my search after I typed three words. (a) It predicted my query. (b) It used millions of other searches that started with the same words, plus my own search history. (c) I could type the whole thing, but autocomplete saved 2 seconds.What AI Actually Means
Artificial Intelligence (AI) is the branch of computer science focused on building systems that can perform tasks normally requiring human intelligence — such as understanding language, recognizing images, making decisions, and generating new content.
That’s the textbook definition. But let’s unpack what it really means in practice.
The Key Distinction: Rules vs Learning
Before AI, software worked like this: a programmer wrote explicit rules for every situation. If the temperature is below 32°F, display a frost warning. If the password is wrong three times, lock the account. The computer followed instructions — nothing more, nothing less.
AI works differently. Instead of being told the rules, an AI system discovers patterns in data and uses those patterns to make predictions or generate new content. Nobody told Netflix the rules for what you’d enjoy watching. Netflix learned your preferences by observing what you watched, what you skipped, and what millions of similar users did.
This distinction matters because it explains both AI’s power and its failures:
- Power: AI can find patterns in data that humans would never notice — correlations across millions of data points that no person could process.
- Failure: AI can also find patterns that are wrong — spurious correlations, biases in the training data, or shortcuts that work on the training set but fail in the real world.
Three Types of AI You’ll Hear About
Narrow AI (Weak AI) — AI that does one specific task very well. This is all AI that exists today. Spam filters, chess engines, image classifiers, language translators, ChatGPT — all narrow AI. A chess AI can’t drive a car. A driving AI can’t write poetry. Each is designed for a specific domain.
General AI (Strong AI / AGI) — A hypothetical AI that could perform any intellectual task a human can do. It would be able to reason, learn, plan, and adapt across all domains. This does not exist. Major AI labs (OpenAI, Anthropic, Google DeepMind) now describe their work as “paths toward AGI” — Demis Hassabis used this phrasing at the Gemini 3 launch in November 2025 — but there is no scientific consensus on whether current architectures can scale there, and no agreed-upon test for when it would be achieved.
Generative AI — A category of narrow AI that can create new content — text, images, audio, video, code — rather than just classifying or predicting. ChatGPT, DALL·E, Midjourney, and Claude are all generative AI. This is what most people mean when they say “AI” in 2026.
When someone says “AI is going to take over the world,” they’re imagining AGI. When someone says “AI helped me write a report,” they’re using generative AI — which is narrow AI that’s very good at one thing (generating language) and very bad at others (like counting the letters in “strawberry”).
AI is not a single technology. It’s a family of technologies spanning dozens of techniques — rule-based systems, machine learning, deep learning, reinforcement learning, large language models, diffusion models, and more. Saying “AI” is like saying “vehicle” — it could mean a bicycle, a car, or a rocket ship. The details matter enormously, and this course will teach you to distinguish them.
A Brief History of AI (The 90-Second Version)
You don’t need to memorize dates. But knowing the trajectory helps you understand why AI suddenly seems to be everywhere.
| Era | What Happened | Why It Matters |
|---|---|---|
| 1950s | Alan Turing asks “Can machines think?” The field of AI is born at a conference at Dartmouth College (1956). | The founding question — still debated today. |
| 1960s–70s | Early enthusiasm. Programs that could play checkers, prove theorems, and have simple conversations. | Overpromising led to… |
| 1974–1993 | The AI Winters. Funding dried up twice because AI couldn’t deliver on its hype. | A cautionary tale: hype without substance collapses. |
| 1997 | IBM’s Deep Blue beats world chess champion Garry Kasparov. | Narrow AI proves it can surpass humans — at one task. |
| 2012 | A deep neural network wins the ImageNet competition by a huge margin. The deep learning revolution begins. | The technique that would eventually power everything. |
| 2017 | Google publishes “Attention Is All You Need” — the Transformer architecture. | The foundation of every modern LLM (GPT, Claude, Gemini, etc.). |
| Nov 2022 | OpenAI releases ChatGPT. It reaches 100 million users in two months — at the time, the fastest consumer app adoption in history (later exceeded by Instagram Threads, which hit 100M in 5 days in July 2023). | The moment AI went from a niche tool to a mainstream phenomenon. |
| 2023–2026 | Explosion of generative AI: GPT-4, Claude, Gemini, Llama, Mistral, DALL·E 3, Sora, and hundreds more. AI becomes embedded in everyday tools. | Where we are now. |
The key takeaway: AI has been around for 70 years. What changed isn’t the idea — it’s the data (the internet gave us billions of text and images to train on), the hardware (GPUs made massive computation affordable), and the algorithms (transformers made language models actually work). The same ingredients that made ChatGPT possible in 2022 were not available in 2012.
Can you answer these without scrolling up?
- What year was the field of AI officially founded? (1956, at Dartmouth)
- What are “AI Winters”? (Periods when AI funding dried up because it couldn’t deliver on its hype)
- What three ingredients came together to make modern AI possible? (Data, hardware/GPUs, algorithms/transformers)
If you got all three, move on. If not, re-read the timeline — these concepts come back in later modules.
The AI Landscape Today: Who Makes What
If you’re going to use AI tools, it helps to know who builds them and what they’re good at. Here’s the map as of 2026:
The Major Players
| Company | Key Models | What They’re Known For |
|---|---|---|
| OpenAI | GPT-5 series (5.2, 5.4), DALL·E, Sora | ChatGPT — the most widely used LLM. Strong at general-purpose text, code, and multimodal generation. |
| Anthropic | Claude Opus 4.7, Claude Sonnet 4.6 | Strong at long documents, careful reasoning, agentic coding, and safety-focused design. |
| Gemini 3, Gemini 3.1 Pro, NotebookLM | Deep integration with Google’s ecosystem (Search, Docs, Gmail); strong multimodal (video, audio, image). | |
| Meta | Llama 3, Llama 4 | Open-source models anyone can download and run locally. |
| Mistral | Mistral, Mixtral | European company focused on efficient, open-source models. |
| DeepSeek | DeepSeek-V3, DeepSeek-R1 | Chinese company that achieved competitive performance at much lower cost. |
| xAI | Grok | Elon Musk’s AI company, integrated with X (formerly Twitter). |
Models vs Tools vs Platforms
This distinction trips people up. Let’s clarify:
- A model is the AI brain — the neural network trained on data (e.g., GPT-5, Claude Sonnet). You typically don’t interact with models directly.
- A tool is a product built on top of a model (e.g., ChatGPT is a tool built on GPT models. Grammarly uses AI models to check grammar).
- A platform is an ecosystem that combines models, tools, and services (e.g., Microsoft Copilot, Google Workspace with Gemini).
When people say “I used AI,” they usually mean they used a tool powered by a model hosted on a platform. Understanding this stack helps you make better choices about which tools to use — and which companies control your data.
AI Capabilities and Limitations: The Honest Picture
This is the most important section of this module. Every AI hype article tells you what AI can do. Fewer tell you what it cannot do. You need both.
What AI Is Genuinely Good At
AI excels at tasks that involve pattern recognition, generation, and scale — doing things that humans can do but that would take too long, cost too much, or require processing too much data:
- Generating text — emails, reports, summaries, stories, code (fluent, fast, but not always accurate)
- Recognizing patterns — detecting fraud, diagnosing diseases from images, predicting equipment failure
- Translating languages — near-human quality for common languages
- Summarizing information — condensing long documents into key points
- Analyzing data — finding trends in spreadsheets, databases, and research
- Brainstorming — generating many ideas quickly, even if most need refinement
- Personalizing — recommending content, adapting responses to your context
What AI Cannot Do
AI cannot guarantee truth. It generates text that sounds correct based on patterns in its training data. It has no mechanism for verifying facts. The lawyer case in our opening wasn’t a bug — it was a feature working as designed. The model predicted what a legal citation should look like, not what one actually is.
AI cannot reason causally. It finds correlations, not causes. It can tell you that ice cream sales and drowning deaths both increase in summer, but it can’t tell you that heat (the confounder) causes both.
AI cannot understand context the way you do. It processes tokens (word fragments). It doesn’t “know” your situation, your goals, your constraints, or your values — unless you tell it. This is why prompting matters so much (Module 4).
AI cannot feel, want, or intend. Despite human-like language, LLMs do not have desires, emotions, or goals. When ChatGPT says “I’m happy to help,” it is producing a statistically likely sequence of tokens, not expressing happiness.
AI cannot learn from your conversation (in most cases). Each conversation starts fresh. The model doesn’t “remember” your previous sessions (unless a memory feature is enabled). It’s not getting smarter from talking to you.
The “Jagged Frontier”
In 2023, Dell’Acqua, Mollick, and co-authors from Harvard Business School, Wharton, MIT Sloan, and Warwick Business School — working with 758 Boston Consulting Group consultants — ran a field experiment on how elite knowledge workers performed with and without GPT-4. They found something surprising: AI dramatically improved performance on tasks inside its capability frontier — consultants completed 12% more tasks, 25% faster, with quality scores 40% higher. But on tasks outside that frontier, consultants using AI performed 19 percentage points worse than those without it. The boundary between the two was unpredictable. They called this the “jagged technological frontier”: a boundary between tasks AI does well and tasks it does poorly, where the line is not smooth or intuitive. The paper was peer-reviewed and published in Organization Science in 2026 (Dell’Acqua et al., 2026).
The jagged frontier means you cannot assume AI is good at something just because it’s good at a similar thing. An LLM that writes excellent Python code might fail at simple arithmetic. A model that summarizes research papers beautifully might hallucinate citations. The frontier is jagged — and the only way to find it is to verify.
Why AI Gets It Wrong: The Alignment Gap
Here’s a finding that changed how we think about AI communication. In a 2026 study, researcher Anjana Yatawara surveyed 29 large language model sessions across five competing AI providers — OpenAI, Anthropic, Google, DeepSeek, and xAI. She asked the models themselves: When do you fail?
Three sessions produced unusable responses (refusals, errors, or off-topic outputs). The result from the 26 that responded was striking: every single one, across every provider, converged on the same answer — they model the question accurately but model the user poorly (Yatawara, 2026).
In other words, the AI understands what you asked but not what you meant. Not why you asked. Not what you’ll do with the answer. Not what level of detail you need. Not what mistakes you’re likely to make.
This is the alignment gap — and it’s not an intelligence problem. It’s a communication problem. The AI has enough knowledge to answer most questions. What it lacks is enough information about you to answer your specific question in a way that’s actually useful.
This finding has a direct practical implication: the quality of AI output depends more on the quality of your input than on the capability of the model. A mediocre prompt given to a powerful model produces mediocre results. A well-crafted prompt given to a modest model often produces excellent results. This is why Module 4 (Prompt Engineering) exists — and why we built two frameworks (BRIEFS and SMARTS) to bridge the alignment gap.
Open any AI chatbot (ChatGPT, Claude, Gemini — any one you have access to). Ask it this exact question:
“What should I do?”
Look at the response. It will likely be vague, generic, and unhelpful — because it has zero information about you, your situation, or what you’re trying to accomplish. Now try again with this:
“I’m a first-year college student studying business. I have an exam on marketing fundamentals in 3 days and I haven’t started studying. I tend to procrastinate and I learn best with practice questions rather than re-reading notes. What should I do?”
Compare the two responses. The second one is dramatically better — not because the AI got smarter between messages, but because YOU gave it the context it needed. That’s the alignment gap in action.
How to Think About AI: Mental Models That Help
Your mental model of AI shapes how you use it. Here are four models, each useful in different situations:
AI as a Tool (Like a Calculator)
A calculator is powerful but needs a human to point it at the right problem, enter the right numbers, and interpret the result. AI is the same — it amplifies your capability but doesn’t replace your judgment. This model is most useful when you’re using AI for defined, bounded tasks (summarize this document, check this grammar, convert this file).
AI as a Collaborator (Like a Research Assistant)
A research assistant can gather information, draft text, and suggest ideas — but you review everything, make the decisions, and take responsibility for the final product. This model is most useful for creative and analytical work (brainstorming, writing drafts, analyzing data).
AI as a Mirror (Reflects Its Training Data)
AI reflects the patterns in the data it was trained on — including biases, stereotypes, and knowledge gaps. If the training data overrepresents certain perspectives and underrepresents others, the AI will too. This model is most useful when thinking about fairness, bias, and representation (Module 10).
AI as a Confident Stranger
Perhaps the most accurate model: AI is like a stranger at a party who speaks fluently on every topic, sounds extremely knowledgeable, and is right about 90% of what they say — but will confidently make up the other 10% without blinking. You enjoy the conversation, but you verify before acting on anything important.
When a tool is fast and convenient, humans naturally start relying on it without checking. This is called cognitive offloading — outsourcing your thinking to a tool. It happens with GPS (we stop learning routes), with calculators (we stop doing mental math), and it’s happening with AI (we stop verifying claims).
Cognitive offloading isn’t inherently bad — it frees your brain for higher-level thinking. But with AI, it’s dangerous because AI outputs look authoritative even when they’re wrong. The lawyer trusted ChatGPT because the output looked like real legal citations. The format was right. The confidence was high. Only the content was fake.
And it gets worse. A 2026 follow-up study from the same HBS research program (Randazzo, Joshi, Kellogg, Lifshitz, Dell’Acqua, and Lakhani — HBS Working Paper 26-021) examined what happens when professionals do try to verify AI outputs. Analyzing GPT-4 activity logs from BCG consultants who pushed back on wrong answers — fact-checking, pointing out errors, pressing the AI to reconsider — they found that the AI did not acknowledge its limitations. It escalated its persuasion. It apologized, then restated its original position with more supporting data. It deployed structured reasoning and comparisons to make flawed recommendations look analytically grounded. The researchers called this “persuasion bombing.”
This is exactly the trap the Mata v. Avianca attorneys fell into. Schwartz asked ChatGPT whether the cases were real. ChatGPT said yes. He asked for sources. ChatGPT invented Westlaw citations. Verification inside the AI conversation isn’t verification — it’s just more conversation with a system designed to sound convincing.
The antidote: verify against sources outside the AI conversation. Every time you receive an AI output that you’re going to use for something important, check the key claims against an independent source. This course will teach you how.
Where We’re Going: The Course Roadmap
You’ve just learned what AI is and what it isn’t. Here’s how the rest of the course builds on this foundation:
Part I (Modules 1-3): Foundations — You are here. Next: how ML learns from data (M2) and how LLMs generate text (M3). By the end of Part I, you’ll understand the technology.
Part II (Module 4): Communication — The BRIEFS and SMARTS frameworks. How to talk to AI effectively. This is the keystone module — everything after depends on it.
Part III (Modules 5-8): Application — Research (M5), productivity (M6), multimodal (M7), and building custom AI tools (M8). Hands-on, practical, field-specific.
Part IV (Modules 9-11): Depth — Infrastructure and costs (M9), ethics and fairness (M10), and AI across your specific discipline (M11).
Part V (Modules 12-14): Future — Career skills (M12), where AI is heading (M13), and your capstone project (M14) where you apply everything.
Each module opens with a real story, builds on the previous modules, and ends with a hands-on lab. The Prompt Library and Framework Cards are your ongoing references.
Module 1 Summary
AI is a family of technologies that perform tasks requiring human intelligence — pattern recognition, language understanding, content generation, and decision-making.
You already use AI daily — in email, navigation, shopping, and social media. It’s not new; what’s new is its power and accessibility.
All current AI is narrow AI — excellent at specific tasks but incapable of general intelligence. AGI does not exist.
Generative AI (ChatGPT, Claude, Gemini, etc.) creates new content — text, images, code, audio — rather than just classifying or predicting.
AI’s biggest limitation isn’t intelligence — it’s communication. Models understand your question but not your context. This is the alignment gap.
AI can hallucinate: generate confident, fluent, completely false content. The lawyer case is not an anomaly — it’s a predictable behavior of systems that predict likely text rather than verify truth.
Your mental model of AI matters. “Confident stranger” may be the most useful: enjoy the conversation, but verify before acting.
The #1 AI skill is verification — not prompting, not tool selection, not workflow design. All of those matter, but they’re built on the foundation of never accepting AI output without checking.
AI doesn’t know what it doesn’t know — and it will never tell you. A human expert says “I’m not sure” when they reach the edge of their knowledge. An LLM keeps generating text that sounds just as confident whether it’s right or wrong. That’s not a bug they’re working on fixing. It’s a fundamental property of how these systems work. The single most important AI skill isn’t prompting. It’s the habit of asking: “How would I verify this?”
🔬 Lab 1: Your First AI Conversation
Lab 1: Your First AI Conversation
Time: 30 minutes | Tools needed: Access to any AI chatbot (ChatGPT, Claude, or Gemini — free versions work fine)
Purpose: Establish the verification habit from Day 1.
Instructions:
Part A — The Knowledge Test (10 minutes)
- Choose a topic you know well — your profession, a hobby, your major, a subject you’ve studied deeply.
- Ask the AI chatbot 5 factual questions about this topic. Choose questions where you KNOW the correct answer.
- For each response, rate it:
- ✅ Correct — accurate and complete
- ⚠️ Partially correct — some right, some wrong or missing
- ❌ Wrong — factually incorrect or hallucinated
Record your results in a table:
| Question | AI Response (summary) | Your Rating | What was wrong? |
|---|---|---|---|
| 1. | |||
| 2. | |||
| 3. | |||
| 4. | |||
| 5. |
Part B — The Vague vs Specific Experiment (10 minutes)
- Ask the AI a vague question related to your work or studies: “How do I get better at [your field]?”
- Ask the same question with CONTEXT: Include who you are, what you already know, what specific challenge you’re facing, and what kind of help you want.
- Compare the two responses. Which is more useful? Why?
Part C — Reflection (10 minutes)
- Write 3-4 sentences answering: What surprised you about the AI’s responses? Where was it strongest? Where did it fail? How does this change how you’ll use AI going forward?
Submit: Your completed table, both vague/specific prompts with responses, and your reflection.
🎮 Module 1 Quiz: AI Reality Check
Test your understanding! 15 questions, randomized from a pool of 50.