Loading lesson...
Loading lesson...
Capability rises with the data each class learns from, but no row in that stack says whether an output is usable. The four NIST AI RMF 1.0 gates below settle that: measure error, check deployment context, manage residual risk, document limits. A good error score is not clearance.
Capability comes from pattern recognition over data. Usability comes from verification, not from the model. NIST AI RMF 1.0 names verification as the deployment gate.
Understanding AI properly requires separating nested concepts that media coverage regularly conflates. This module establishes the vocabulary and mental models you need before building anything with AI systems.
Next: What Artificial Intelligence Actually Is.
Artificial intelligence(AI) is the field of computer science concerned with building systems that perform tasks normally requiring human judgement. The word "normally" is doing significant work in that sentence. Every decade, the definition of what counts as intelligent shifts: tasks that once seemed remarkable (chess, image recognition, translation) become routine engineering once they are solved.
This definitional drift is not a bug in the field. It reflects a genuine pattern: once we understand exactly how a system achieves a result, we tend to stop calling the result "intelligent." This phenomenon, sometimes called the AI effect, means that public expectations about AI consistently run ahead of what deployed systems actually do.
Understanding AI properly requires separating three nested concepts. AI is the broadest category; (ML) is a subset; is a subset of machine learning. Each narrows the scope further, and each requires distinct mental models for working with it effectively.
The three nested fields can be understood as concentric rings. The outer ring, AI, covers any technique that enables machines to simulate aspects of intelligent behaviour. Rule-based systems from the 1970s that checked passport validity were AI, even though they contained no learning at all.
The middle ring, machine learning, covers systems that learn patterns from data rather than following hand-written rules. Instead of writing "if the email contains the word prize then mark as spam," an ML system examines thousands of examples and infers the patterns itself.
The inner ring, deep learning, uses neural networks with many layers. These layered networks learn hierarchical representations: pixels become edges, edges become shapes, shapes become objects. The breakthrough came when graphics processing units (GPUs) made training large networks economically viable around 2012.
Next: Key Milestones in AI History.
Modern AI did not arrive suddenly in 2022. It is the product of seventy years of research, two extended periods of reduced funding known as "AI winters," and a series of empirical breakthroughs enabled by more computing power and more data. Knowing the trajectory helps you calibrate how reliable current systems actually are.
1950: The Turing Test.Alan Turing proposed a simple benchmark: if a human judge cannot reliably distinguish a machine's text responses from a human's, the machine has demonstrated intelligent behaviour. Turing deliberately side-stepped unanswerable questions about consciousness and focused on observable performance.
1974 to 1980: The first AI winter. Early optimism collapsed when rule-based systems proved brittle outside narrow topics. Funding dried up sharply across both the UK and the United States.
1997: Deep Blue defeats Kasparov.IBM's Deep Blue defeated world chess champion Garry Kasparov using search algorithms and functions, not machine learning. It demonstrated that narrow, superhuman capability was achievable in well-defined domains.
2012: AlexNet and the deep learning revolution.A neural network called AlexNet, developed at the University of Toronto by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, won the ImageNet challenge by a margin that shocked researchers. Error rates dropped by nearly 10 percentage points versus the prior year's best result.
2017 to present: The era.Google researchers published "Attention Is All You Need" in 2017, introducing the transformer architecture. By 2020, GPT-3 demonstrated that scaling transformers with vast text data produced systems with broad language capability. ChatGPT launched in November 2022 and reached 100 million users in two months.
“I propose to consider the question, "Can machines think?" ... Instead of attempting such a definition I shall replace the question by another, which is closely related to it and is expressed in relatively unambiguous words.”
Alan Turing, 1950 - Computing Machinery and Intelligence, Mind, 59(236), p. 433
Turing replaced "can machines think" with an observable test. This move, replacing unanswerable philosophical questions with measurable performance criteria, is the methodological foundation of empirical AI research. Modern benchmarks (MMLU, HumanEval, BIG-Bench) descend directly from this framing.
Next: Types of AI by Scope.
Media coverage of AI frequently conflates three distinct concepts that researchers treat as separate: narrow AI, general AI, and . Conflating them leads to misaligned expectations about what current tools can and cannot do.
(Artificial Narrow Intelligence, ANI) is AI designed and trained for a specific task. It can be superhuman at that task and completely useless outside it. Every in commercial use today is narrow AI: the model that detects fraud on your bank card, the voice assistant that sets a timer, the algorithm that recommends what to watch next.
General AI (, AGI) is a hypothetical system that can perform any intellectual task a human can, applying reasoning across domains without task-specific training. No such system exists. Researchers disagree sharply about whether AGI is decades away, centuries away, or impossible in principle.
Superintelligence is a further hypothetical: an AI system that exceeds human cognitive capability across all domains. It remains a topic of philosophical and speculative discussion, not an engineering specification.
Common misconception
“ChatGPT and similar tools are getting close to general AI.”
Every large language model in production is narrow AI. These systems are trained to predict text continuations and can appear to generalise broadly, but they lack causal reasoning, reliable factual grounding, and persistent memory. The impression of generality comes from the breadth of text in their training data, not from general intelligence.
Next: What AI Cannot Do.
Understanding limitations is as important as understanding capabilities. The systems you will build agents with have genuine and structural limitations that are not bugs to be patched. They are properties of how these systems work.
. Large language models generate plausible-sounding text based on patterns in their training data. When asked about a topic for which training data is sparse, the model fills gaps with confident-sounding fabrications. This is called hallucination. The model has no concept of truth or falsehood; it produces statistically likely continuations, some of which happen to be false.
Causal reasoning. Statistical correlations in training data are not causal relationships. A model trained on data where ice cream sales and drowning incidents both peak in summer will represent that correlation, not the fact that neither causes the other.
Consistent behaviour. The same prompt can produce different outputs on different runs, particularly at higher settings. Production systems must account for this variability, especially in safety-critical or compliance contexts.
Stable identity over time. Each conversation typically starts fresh, with no persistent memory by default. An LLM does not remember your name from a previous session unless the application explicitly stores and retrieves that information.
Common misconception
“AI understands the text it processes.”
Large language models predict probable next tokens based on patterns in training data. They do not have beliefs, intentions, or understanding in any philosophically strong sense. The model has no model of the world, only weights that make certain outputs probable in certain contexts. ISO/IEC 42001:2023 (Section 6.1.2) requires organisations to identify this limitation explicitly as part of AI risk assessment.
Each row runs from the limit, to the symptom you will actually meet, to the control that contains it, and not one of those controls lives inside the prompt, so a system that has only been prompted harder has not been contained at all.
Fabrication, correlation, variance and forgetting are properties of next-token prediction. Every control that contains them sits in the system around the model, never in the prompt.
A passport office ran a system in the 1970s that decided whether a document was valid by working through a list of checks written out by officials. Using the nested picture of AI, machine learning and deep learning set out in this module, where does that system sit?
A colleague accepts that the fraud model on her bank card is narrow AI, but argues that the assistant she uses for drafting, coding and translation must be an early general AI because it handles so many different jobs. On this module's account of scope, what is wrong with her argument?
A team asks a model to summarise an obscure piece of regulation. It returns a fluent summary that cites a clause number the regulation does not contain. Which of the limits set out in this module does that behaviour show?
Alan Turing, 'Computing Machinery and Intelligence', Mind, Volume 59, Issue 236 (October 1950)
Mind, 59(236), pp. 433-460
The foundational paper proposing the imitation game as a framework for evaluating machine intelligence. Quoted in Section 1.2 to establish the shift from unanswerable philosophical questions to observable performance criteria.
NeurIPS 2012 Proceedings
The AlexNet paper that triggered the deep learning revolution. Referenced in Section 1.2 as the empirical event that made deep learning the dominant approach to computer vision and language modelling.
Vaswani, A. et al. (2017). Attention Is All You Need
NeurIPS 2017 Proceedings
Introduced the transformer architecture that underlies GPT, Claude, Gemini, and all modern large language models. Referenced in Section 1.2 as the architectural breakthrough enabling the current era of AI capability.
NIST AI Risk Management Framework (AI RMF 1.0), January 2023
Section 2.1: AI Risks and Trustworthiness Characteristics
The US government framework for categorising and managing AI risks in practice. Referenced in Section 1.3 to ground the scope limitation claim in a recognised regulatory standard.
ISO/IEC 42001:2023, Artificial Intelligence Management Systems
Section 6.1.2: AI Risk Assessment
International standard requiring organisations to identify AI limitations as part of formal risk assessment. Referenced in Section 1.4 to ground the hallucination and understanding limitations in a compliance context.
You now have the vocabulary to distinguish AI, machine learning, and deep learning - and to identify the structural limitations that every LLM inherits from its training. The next module examines the specific architecture that makes these systems capable: the transformer. More importantly, it shows how adding access to a language model turns a text predictor into an agent that can take real actions in the world.
Module 1 of 40 in the AI Agents course, and the first of the five Foundations modules
