Loading lesson...
Loading lesson...
This is the first of 8 Foundations modules. The Foundations stage builds the conceptual vocabulary you need for the Applied and Practice & Strategy stages that follow (24 modules total, ~14 hours). No prior AI knowledge is required.
In February 2011 an IBM system called Watson played the American quiz show Jeopardy! against two of the show's champions, Ken Jennings and Brad Rutter. The contest ran to two matches. The first was broadcast in two parts, on 14 and 15 February, and ended with Watson on $35,734 against Rutter's $10,400 and Jennings's $4,800. The second match went out on 16 February, and the final result was a victory for Watson on $77,147, ahead of Jennings on $24,000 and Rutter on $21,600. Only at that point was first place settled, and with it the prize of one million US dollars.
Watson played with no internet connection. What it had instead was 200 million pages of structured and unstructured material, four terabytes in all, the 2011 edition of Wikipedia included in full. According to one published account of the hardware, all of that was held in RAM during the game, because reading it back from a hard drive would have been too slow to keep pace with human champions.
The win did not come from knowing things. What made it work was not a novel algorithm. The advance was in scale and orchestration: hundreds of language analysis techniques, every one of them already proven, all running over the same clue at once, with independent agreement between them as the signal the system trusted. The researchers who built it were explicit about what the game really demanded. Their own write-up of the project in AI Magazine, Building Watson: An Overview of the DeepQA Project, puts it plainly: "Winning at Jeopardy requires accurately computing confidence in your answers." That requirement reached every stage of the pipeline. Each stage had to report how sure it was, and those separate figures were rolled up into a single number for the answer Watson eventually offered.
The limits of that design showed at the end of the first match, with the whole of the second still to be played. The Final Jeopardy category was U.S. Cities, and the clue read "Its largest airport was named for a World War II hero; its second largest, for a World War II battle". Jennings and Rutter both gave the correct response, Chicago. Watson wrote "What is Toronto?????", the five question marks standing in for how little it trusted itself. Toronto and Chicago both sat well below its confidence threshold, at 14 per cent and 11 per cent. The machine that would go on to win the contest had put a wrong city above the right one, and had almost no faith in either.
When a system gives the right answer, what have you actually learned about what it can do?
The Watson story is a useful starting point because it surfaces a confusion that persists to this day: the gap between what AI systems do and what people assume they do. Understanding that gap is essential for anyone making decisions about AI, whether you are a developer, a product manager, or a business leader.
This module assumes no prior AI knowledge. If the terms below are already familiar, use the knowledge checks to confirm your understanding and move to Module 2: Data as fuel.
Learning outcomes gives the mechanism. What artificial intelligence actually means shows where the next decision has to be made.
The term "artificial intelligence" was coined by John McCarthy for the 1956 Dartmouth Workshop, widely considered the founding event of AI as a field. McCarthy defined it as "the science and engineering of making intelligent machines." That definition is deliberately broad, and its breadth has caused confusion ever since.
In practice, virtually all AI systems in production today fall under a narrower category called narrow AI (also known as weak AI). A narrow AI system performs a specific task, often extremely well, but cannot transfer its ability to other domains. A chess engine cannot write poetry. A spam filter cannot drive a car. Watson could answer quiz questions but could not hold a conversation.
The opposite concept, artificial general intelligence (AGI), refers to a hypothetical system that could perform any intellectual task a human can. AGI does not exist. No timeline for its arrival has scientific consensus. When this course uses the term "AI," it refers to narrow AI unless explicitly stated otherwise.
“Artificial intelligence is the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings.”
Encyclopaedia Britannica - 'Artificial intelligence' entry, britannica.com/technology/artificial-intelligence
This reference-standard definition is useful because it centres on tasks rather than inner experience. Whether a system 'thinks' is a philosophical question. Whether it performs tasks associated with intelligence is measurable.
The key phrase is "tasks commonly associated with intelligent beings." This includes recognising objects in images, translating between languages, generating text, playing games, and making predictions from data. It does not require consciousness, understanding, or general reasoning.
AI, machine learning, and deep learning: three concentric circles is the next test of whether the concept works outside a toy example.
Three terms are frequently used interchangeably in the press and in marketing materials. They are not the same thing. They form a hierarchy:
This hierarchy matters because different problems require different approaches. Not every problem needs deep learning. Many production systems use simpler ML methods (logistic regression, decision trees, gradient boosting) that are faster to train, easier to explain, and cheaper to run.
The diagram below draws the three as nested scopes rather than as rival technologies. Read the arrows on the left first: each names what you must already hold before the ring it points at is open to you. Then read the "Read it" row down the three rings, because that is what each step inward costs you in explanation.
Each ring opens only to a reader who already holds what its arrow names, a task to automate, then labelled examples, then data at scale, and the Read it row weakens at every inward step, so capability is bought with the explanation you can still give.
AI, machine learning and deep learning are nested scopes, not rival technologies. Step inward only when you already hold what the inner ring needs, because every step inward trades a rule you can read for a pattern you can only explain after the fact.
How AI differs from traditional software moves the lesson from vocabulary to evidence.
Common misconception
“AI systems understand what they are doing”
Current AI systems, including large language models, perform statistical pattern matching. They process inputs and generate outputs that appear intelligent, but they do not have understanding, consciousness, or intentions. A language model predicts the next token in a sequence based on patterns in its training data. If you design a product around the assumption that an AI understands context the way a human colleague does, you will encounter failures. Building guardrails for this gap is a core AI engineering skill.
Common misconception
“AI will replace all human jobs imminently”
AI automates specific tasks, not entire jobs. A radiology AI can flag potential anomalies in X-rays, but a radiologist also consults patient history, communicates findings, handles edge cases, and takes legal responsibility. Virtually all AI in production is narrow AI: it performs one specific task, often extremely well, but cannot transfer that ability to another domain. Automating one task therefore leaves the rest of the role standing, and full job displacement would mean automating virtually all activities in a role. Framing AI as a task augmentation tool leads to realistic planning and better outcomes.
Traditional software follows explicit rules written by a programmer. If the input matches condition A, do X. If it matches condition B, do Y. The programmer anticipates every case and writes instructions for each.
Machine learning inverts this. Instead of writing rules, the programmer provides data (examples of inputs and desired outputs) and an algorithm. The algorithm discovers the rules by finding patterns in the data. This is why data quality matters so much, a topic we examine in detail in Module 2.
This inversion has practical consequences:
The diagram below makes the contrast concrete. The top region shows the two implementation paths side by side. The bottom region shows the wider that absorbs either path's output and turns it into an action. The model is one component of the system; the system is what makes the decision and carries the accountability.
Both lanes end at an output arrow entering the same system boundary, and only that boundary, holding policy, guardrails, review and audit, emits the action, so swapping a written rule for a learned model moves none of the accountability.
A learned model's output is one input to the decision. The surrounding system, with its policy, guardrails, and human review, is what makes the decision and carries the accountability.
How AI differs from traditional software sets the boundary. The Turing Test and why it is not enough examines what changes when the system meets real use.
In 1950, Alan Turing proposed what he called the "imitation game": if a human evaluator cannot reliably distinguish between a machine's responses and a human's, the machine can be said to exhibit intelligent behaviour. This has been popularised as the Turing Test.
The Turing Test is historically important but practically limited. Modern language models can fool humans in short conversations, yet they cannot reliably perform basic arithmetic, maintain consistent beliefs across a conversation, or explain their own reasoning. Passing the Turing Test tells you about surface appearance, not about capability or understanding.
Better evaluation approaches exist and are used in practice. These include task-specific benchmarks (can the system correctly answer medical questions?), adversarial testing (can you find inputs that break the system?), and calibration testing (when the system says it is 90% confident, is it correct 90% of the time?). We cover evaluation methods in depth in Module 5.
“We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956.”
John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon - A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955)
This proposal, written in 1955 for the 1956 Dartmouth Workshop, is the document that coined the term 'artificial intelligence.' The authors assumed the problem could be substantially solved in one summer. Nearly 70 years later, the field is still working on the foundational challenges they identified.
A vendor says its refund-handling chatbot is 'true AI' because it deals with open language rather than the fixed rules of the system it replaced. A colleague asks whether that makes it artificial general intelligence. Which response is accurate?
A team is choosing between gradient boosting and a deep neural network for fraud detection on 40 well-understood numeric features. Which statement matches how this module places the three terms?
A fraud detection model trained on 2022 transaction patterns has been in production ever since. Its accuracy has fallen quarter on quarter, and the code has not been changed. What is happening?
Full proposal (2 pages)
The document that coined the term 'artificial intelligence' and founded the field. Used in Section 1.1 to establish the origin and scope of the term.
Tom Mitchell, Machine Learning (1997)
Chapter 1, Definition 1.1
The standard textbook definition of machine learning: 'A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.' Used in Section 1.2.
IBM Research, 'The DeepQA Research Team'
Overview: Watson's Jeopardy! Challenge
Technical documentation of the Watson Jeopardy system. Used in the opening story to distinguish what Watson did (statistical pattern matching across 200 million pages) from what headlines claimed it did (thinking).
NIST AI Risk Management Framework (AI RMF 1.0), January 2023
Section 1 (Framing Risk), Appendix A (AI Actor Tasks)
The US government framework for managing AI risk. Introduced here as a structural reference used throughout the course. Defines four functions: Govern, Map, Measure, Manage.
Alan Turing, 'Computing Machinery and Intelligence', Mind, Volume 59, Issue 236 (October 1950)
Section 1 (The Imitation Game)
The original paper proposing the Turing Test. Used in Section 1.4 to explain why the test is historically important but practically insufficient for evaluating modern AI systems.
You now know what AI is (and is not), how AI, ML, and deep learning relate to each other, and why the gap between appearance and capability matters. The next question is: what makes AI systems work well or fail? The answer, almost always, is data. Module 2 examines how data quality, bias, and preprocessing determine whether an AI system succeeds or causes harm.
Module 1 of 33 · AI Foundations