Introduction to Artificial Intelligence

Part I: The Definitional Challenge

"Artificial intelligence" is one of the most widely used — and most poorly defined — terms in contemporary discourse. Politicians promise to regulate it; companies compete to claim they have it; researchers argue about whether any current system achieves it. So: what are we actually talking about?

The difficulty is not merely semantic. How we define AI shapes what we build, what risks we take seriously, and who we hold responsible when things go wrong. A system described as "just software" invites one set of ethical questions; "an intelligent agent" or "potentially conscious" invites entirely different ones.

The Moving Goalposts Problem

One pattern recurs, sometimes called Tesler's Theorem after the computer scientist Larry Tesler: "AI is whatever hasn't been done yet." (Tesler's own wording was "Intelligence is whatever machines haven't done yet"; the snappier version, popularized by Douglas Hofstadter, is the one that stuck.) Whenever an AI system succeeds at a task once thought to require genuine intelligence, that task tends to get reclassified as "mere computation."

The pattern suggests that "AI" functions less as a stable category than as a placeholder for unsolved problems: success disqualifies. That is why debates about whether current large language models are "truly" intelligent are so heated and so often unresolvable.

A Taxonomy of AI Distinctions

Contemporary AI discourse employs several overlapping distinctions, each carrying different philosophical commitments and practical implications.

Most current AI systems are narrow, weak in Searle's sense, connectionist, and increasingly agent-like — taking sequences of actions toward goals rather than answering single queries. A narrow, connectionist system acting agentically in a constrained domain raises different ethical questions than one approaching general capability.

Part II: A History of AI Technologies

Artificial intelligence did not arrive suddenly with ChatGPT. It has a rich and humbling history — genuine breakthroughs, spectacular overpromises, funding collapses, patient rebuilding — and the pattern repeats often enough to be worth learning.

Prehistory: Dreams of Artificial Minds

The dream is ancient. Greek mythology gave Hephaestus golden robots to assist him in his workshop; medieval engineers built clockwork automata that mimed human actions; Descartes speculated that animals might be nothing more than sophisticated machines. What these visions share is the intuition that intelligence might be separable from biology.

The modern story begins with Leibniz's vision of a calculus ratiocinator — a machine that could resolve disputes by calculation rather than argument. Two centuries later Charles Babbage designed the Analytical Engine (1837), a mechanical general-purpose computer never built in his lifetime. It was in notes on this machine that the first serious philosophical argument about machine intelligence appeared.

Key Thinker

Ada Lovelace (1815–1852)

Ada Lovelace — daughter of the poet Lord Byron, educated in mathematics by a mother who feared she had inherited dangerous poetic tendencies — collaborated with Charles Babbage on his Analytical Engine in the 1840s. Her 1843 notes contain what is widely regarded as the first computer algorithm, and she foresaw that such a machine could manipulate symbols of any kind, not just numbers.

But Lovelace is equally important for what she argued a machine could not do. Her central claim — "Lady Lovelace's Objection" — is that the Analytical Engine "has no pretensions whatever to originate any thing." Machines can only do what we know how to order them to perform: they execute instructions; they do not originate ideas. It remains one of the most cited objections in the philosophy of AI.

The objection is still alive: when people say large language models are "merely predicting the next token," they are invoking her position.

Alan Turing and the Birth of Computing

Alan Turing (1912–1954) turned the dream of thinking machines into a rigorous scientific question. In 1936, before electronic computers existed, "On Computable Numbers" laid the foundations of computation, introducing the Universal Turing Machine and proving that some precisely stated mathematical questions cannot be settled by any procedure at all. During World War II he led Hut 8 at Bletchley Park, attacking German naval Enigma; historians have credited the codebreaking effort with shortening the war by two to four years. In 1950 he published "Computing Machinery and Intelligence," the paper that launched the philosophy of AI.

Turing's approach was characteristically pragmatic. Rather than define "thinking," a concept he considered too vague to be useful, he proposed a behavioral test: the Imitation Game, set out in Part III. He predicted that in about fifty years' time — so, around 2000 — an average interrogator would have no better than a 70% chance of identifying the machine correctly after five minutes of questioning. Nothing in 2000 clearly met that bar. The first controlled study to report a machine passing the standard three-party version came in 2025, from Cameron Jones and Benjamin Bergen: a leading model prompted to adopt a human persona was judged to be the human 73% of the time . Turing was roughly a quarter-century optimistic — and what such a result shows remains as contested as in 1950.

Turing's life ended tragically. In 1952 the British government prosecuted him for homosexuality, then a criminal offense, and he was subjected to chemical castration as a condition of avoiding prison. He died in 1954 — a probable suicide, though the circumstances remain disputed — and received a royal pardon in 2013. The father of the field was destroyed by his own government.

The Dartmouth Conference and Early AI (1956–1974)

In 1956, John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester convened a summer workshop at Dartmouth College. Their proposal set the agenda for an entire field:

"We propose that a 2 month, 10 man study of artificial intelligence be carried out... The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

The confidence is remarkable: a two-month effort was expected to make serious progress on machine learning, natural language processing, and artificial neural networks. The timelines proved wildly optimistic; the research agenda proved remarkably durable.

The early decades of AI were dominated by Symbolic AI — the approach the philosopher John Haugeland later nicknamed "Good Old-Fashioned AI" (GOFAI) in his 1985 book Artificial Intelligence: The Very Idea . Intelligence, on this view, consists of manipulating symbols according to explicit rules. Newell and Simon's Logic Theorist (1956) proved theorems from Principia Mathematica; Terry Winograd's SHRDLU (1970) followed natural language commands in a blocks world.

AI Winters and Expert Systems

The early optimism crashed twice. Both "AI winters" — periods of funding collapse and disillusionment — followed the same pattern: narrow successes generated overconfident predictions about general capability; the predictions failed; funders withdrew.

The Expert Systems era of the 1980s illustrated the gap between narrow and general intelligence. MYCIN, built at Stanford in the 1970s, recommended antibiotic therapies for blood infections; in a blinded 1979 evaluation its recommendations were rated acceptable at least as often as those of Stanford faculty — though it was never used on actual patients. DENDRAL identified chemical structures from mass spectrometry data; XCON saved Digital Equipment Corporation tens of millions a year configuring computer orders.

But expert systems failed for a structural reason: the "knowledge bottleneck." Every rule had to be hand-coded by extracting expertise from domain specialists, and the result was brittle — a rule base assembled from adult cases had no way of knowing it should not be applied to a six-year-old. The deeper problem is that human expertise is largely tacit: experts cannot fully articulate what they know. The bottleneck did not disappear when the marketing improved, as the next generation's most expensive medical AI project showed.

Machine Learning and Deep Learning

The solution to the knowledge bottleneck was to change the question. Instead of asking "how do we encode what humans know?" machine learning asks "how do we build systems that learn from data?"

What the three paradigms have in common is that the human contribution moves upstream. Nobody writes the rules; people choose the data, the labels, or the reward. That is why questions about training data, labeling labor, and reward design became ethical questions rather than merely technical ones.

The current deep learning revolution traces to a single moment: AlexNet's victory in the 2012 ImageNet competition. A deep neural network trained on consumer graphics cards cut the top-five error rate from about 26% to about 15%, leaving every competing approach far behind. Three ingredients made it possible: far more training data (from the internet), far more computing power (from GPUs), and better training techniques (backpropagation, rectified linear activations, dropout).

The most recent phase of this revolution is the Large Language Model (LLM). The 2017 transformer architecture, introduced in "Attention Is All You Need," made training language models at previously impossible scales practical . Its central trick, attention, lets the network decide, for every word it processes, how much weight to give every other word in the passage — so "it" can be connected to the right noun forty words earlier. Because those comparisons happen all at once, transformers train efficiently at internet scale. The model is then trained on one simple task: predict the next chunk of text; everything else it appears to do is a side effect of getting very good at that. GPT-3 (2020) was widely read as showing that scaling alone produced "emergent" capabilities — a contested reading: later work argued the apparent discontinuities are artifacts of the all-or-nothing metrics researchers chose . ChatGPT's launch in November 2022 reached roughly a hundred million monthly users within two months.

Something changes character in that table around 2025. Every earlier entry is about making the model better before it ships; the last two are about what happens afterward — how long the system thinks while you wait, and whether it answers you or goes off and does something.

Two Changes Worth Naming (2025–2026)

The first is a second scaling axis. For roughly a decade, "more capable" meant "trained on more data with more compute." From 2025, systems began to improve measurably by being given more compute at the moment of use — time to work through a problem before answering. This is inference-time compute, and it changes the economics: the enormous cost used to be paid once, during training, with each answer nearly free afterward. Now capability is bought per question, and cost and environmental footprint scale with how hard you ask the system to think.

Toby Ord pointed out in 2025 that charts showcasing inference scaling almost always put compute on a logarithmic x-axis . A line that looks reassuringly straight on such a chart is not linear growth: compute — and therefore cost, and therefore electricity — must rise exponentially to buy each constant increment of capability. Nothing in the chart is false; the axis flattens the most expensive part of the story. Chapter 9 returns to who pays.

The second change is about where capability lives. Frontier AI has been governed, to the extent it has been governed at all, on an unspoken assumption: that the most capable systems sit inside a few well-resourced laboratories, which can therefore be regulated. January 2025 put a large dent in that assumption.

Pulling in the opposite direction, 2026 was the first year in which capability began constraining release. One laboratory declined to ship a model because of what it could do without human help in offensive cybersecurity; separately, the US government asked another developer to limit a rollout to trusted partners pending federal evaluation, with the restrictions lifted after testing. Small events — but the first in which "should we release this?" turned on what the system could do rather than on the calendar. Laboratory leaders describe their situation as a prisoner's dilemma: each would prefer a slower, safer industry, and none can slow down alone.

These two developments point in opposite directions: one disperses capability beyond the reach of any regulator, the other tries to hold it in. Underneath both is the shift the table calls agents — systems that operate other software and take actions rather than only producing text. That matters less for what they understand than for what they can be blamed for: when a system books the flight or changes the configuration, the human decision point accountability attaches to may not exist. Several arguments later in this book assume there is a person in the loop.

AI in Consequential Decisions

AI systems now routinely make or inform consequential decisions about people's lives. This is where the definitional and philosophical questions of this chapter connect most directly to the ethics of the rest of the book.

These systems share recurring concerns. Bias: systems trained on historical data amplify historical discrimination. Opacity: "black box" decisions may be impossible to explain or challenge. Feedback loops: predictions become self-fulfilling (predict more crime in an area → police there more → find more crime). Due process: how do you appeal a decision made by an algorithm? These concerns apply whether or not we call the systems "intelligent" — which is exactly Floridi's point .

These are not hypothetical risks. In 2016 ProPublica analyzed COMPAS, a commercial risk tool used in US pretrial and parole decisions, and reported that among defendants who did not go on to reoffend, Black defendants were labeled high risk at roughly twice the rate of white defendants (about 45% versus 23%); controlling for prior offenses, age, and sex, Black defendants were still 45% more likely to receive a higher risk score . The vendor replied that COMPAS was equally calibrated for both groups — a given score meant the same reoffending rate whatever the defendant's race — and later work showed that, given unequal base rates, no classifier can satisfy both criteria at once. The disagreement is about which definition of fairness a court should use.

Two further cases show how ordinary the mechanism is. Amazon built an experimental résumé screening tool trained on a decade of its own hiring decisions in a male-dominated engineering workforce; it learned to penalize résumés containing the word "women's" and downgraded graduates of two all-women's colleges. The team was disbanded by 2017 and Reuters made the story public in 2018; Amazon says the tool was never the sole basis for a hiring decision. And in Science in 2019, Ziad Obermeyer and colleagues showed that a commercial algorithm used to select patients for extra care underestimated the needs of Black patients: it used past healthcare spending as a proxy for health need, and less had historically been spent on Black patients at the same level of illness. Correcting the proxy would have raised the share flagged for extra help from about 18% to about 46% . Nobody set out to discriminate; the systems learned the pattern in the data .

Part III: Can Machines Think? Classical Philosophical Debates

We have surveyed what AI systems do. Now a deeper question: what would it mean for a machine to be intelligent? The answer shapes whether current AI has achieved something remarkable or merely elaborate pattern matching, and whether AI systems could deserve moral consideration.

Three debates — Turing vs. Lovelace, Searle vs. Dennett, and Chalmers on consciousness — frame the issues with unusual clarity.

Turing's Proposal: The Imitation Game

Turing's strategy in his 1950 paper was to replace the vague question "Can machines think?" with a concrete behavioral test. In the Imitation Game, an interrogator communicates via text with two hidden parties — a human and a machine — and tries to tell which is which. If the machine consistently fools the interrogator, Turing argued, we have no principled reason to deny it intelligence.

Turing's key philosophical move is pragmatism about mental concepts: he refuses the metaphysical question of what thinking "really is" and proposes an operational criterion grounded in the evidence we already use for other people. We cannot peer inside anyone else's mind; we infer intelligence from behavior.

Lovelace's Objection: Can Machines Originate?

Lovelace's objection — formulated a century before Turing's paper — cuts to the heart of debates about machine creativity that are newly urgent in the age of generative AI.

"The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform... Its province is to assist us in making available what we are already acquainted with."

— Ada Lovelace, Notes on the Analytical Engine (1843)

The pattern is easiest to see where the machine's output surprised the people who built it. When Jason Allen's AI-generated Théâtre D'opéra Spatial won the digital art category of the Colorado State Fair in August 2022, the response was that the program had merely recombined existing artworks. When a large language model was reported in 2023 to have scored around the 90th percentile on the Uniform Bar Examination, the dismissal was that it had pattern-matched rather than reasoned. (A 2024 re-analysis by Eric Martínez found the comparison pool was dominated by February repeat takers and put the score below the 69th percentile against the general test-taking population — around the 48th on the essays ; benchmark claims are claims, not measurements.) And then there is the case philosophers keep returning to.

Searle's Chinese Room

The most influential argument against strong AI was published in 1980 by John Searle, a philosopher at the University of California, Berkeley. His Chinese Room thought experiment is elegant and has generated four decades of debate.

Dennett's Systems Reply

The most powerful response is the systems reply. Searle anticipated it in the 1980 paper itself, and it has been defended since by a number of philosophers, most prominently Daniel Dennett.

Searle's counter is blunt: imagine the person memorizes all the rules and carries them out in their head, outside the room. They are now the entire system — and still understand nothing. Dennett's functionalism holds that anything implementing the right functional organization has mental states regardless of what it is made of; Searle's response is that syntax, however complex, never becomes semantics.

Read across the rows and the disagreement turns out to be one disagreement wearing four costumes. Everything follows from the bottom row: if consciousness is essentially biological, the Chinese Room is decisive; if it is functional organization, the Chinese Room is a level confusion. Neither side has produced an argument the other accepts.

Chalmers and the Hard Problem

David Chalmers introduced a distinction that reorganized the debate about machine consciousness. The easy problems — how the brain processes information, controls behavior, and reports internal states — are "easy" only in that they are in principle solvable by cognitive science. The hard problem is why any of that processing is accompanied by subjective experience at all. Why does it feel like something to taste coffee?

Chalmers' zombie thought experiment sharpens the point: we can conceive of a being physically and functionally identical to a conscious human but with no inner experience whatsoever. If such a being is conceivable (Chalmers argues it is), consciousness is not entailed by functional organization alone. So even a system functionally indistinguishable from a conscious human might not be conscious, and we could not tell: we might build AI that behaves exactly as if it suffers without ever knowing whether it does.

Part IV: Contemporary Definitions

The classical debates remain unresolved, but AI researchers and ethicists still need working definitions. Two contemporary approaches — Russell's rational agency view and Floridi's information-theoretic reframing — have proved especially influential.

Russell and Norvig: AI as Rational Agency

Stuart Russell and Peter Norvig's textbook Artificial Intelligence: A Modern Approach has been the standard introduction to the field for thirty years. They organize possible definitions around two axes — thinking vs. acting, human-like vs. rational — and prefer the last of the four combinations: AI is the study of rational agents, systems that perceive their environment and act to maximize goal achievement.

This definition is precise enough to guide research, general enough to cover everything from chess engines to self-driving cars to recommendation systems, and it avoids the philosophically treacherous question of whether systems "really" think: it focuses on what a system does rather than on what it is internally.

Floridi: Agency Without Intelligence

Luciano Floridi, an Italian philosopher of information — for many years at Oxford, and since 2023 the founding director of the Digital Ethics Center at Yale — reframes the debate. Rather than asking whether AI systems are intelligent — a question that may be unanswerable — he asks what they are: successful agents that achieve goals by means fundamentally different from human cognition.

AI systems, Floridi argues, achieve "agency without intelligence": the capacity to perform tasks and achieve goals, without the understanding, consciousness, and semantic grasp that intelligence in the full human sense implies. A dishwasher cleans dishes without "understanding" cleanliness.

Floridi identifies two symmetrical errors. Overestimating AI — attributing understanding or moral agency to systems that have none — leads to misplaced trust and abdicated responsibility: the judge who relies on COMPAS without understanding its limitations, the doctor who defers to an AI diagnosis. Underestimating AI — dismissing it as "just a tool" — leads to complacency about genuine harms at scale.

Floridi's point is that AI's successes and failures follow from its nature: it exploits patterns in data that may not reflect causal understanding. This is why it can beat grandmasters at chess while failing at tasks a three-year-old finds trivial — like knowing that a person who leaves a room still exists. Neither kind of capability is simply "better," and knowing the difference is essential to knowing when to trust AI and when not to .

Part V: Why Definitions Matter

We began by asking why it is hard to define artificial intelligence. The answer is that the definition does several jobs at once — scientific (what counts as an achievement?), philosophical (can machines have minds?), ethical (what moral status do AI systems have?), and political (what should we regulate?) — and changing it changes the answers.

Definitions Shape What We Build

How researchers define AI shapes what they try to build: a Turing Test focus leads toward human-like conversation; a symbolic focus toward logic and knowledge representation; an agent-based focus toward decision-making; a connectionist focus toward learning architectures and large-scale data. The 1980s expert systems boom and bust shows the stakes: funding flowed to explicit knowledge representation, and when it failed to generalize the whole field contracted.

Definitions Shape Ethical Responsibility

The choice has immediate practical implications. If COMPAS is "just software," only its human designers bear responsibility for its racial disparities. If it is a "rational agent," responsibility becomes more complex. If a large language model might be conscious, we may owe it moral consideration that changes what we permit companies to do to it.

The pattern is striking: under the Turing Test and Searle's criteria almost nothing currently qualifies as AI; under agent-based and agency frameworks almost everything does. Your definition therefore decides whether the systems now shaping decisions about credit, bail, employment, and information access fall within the scope of AI ethics and regulation at all.

Key Points

This chapter has examined artificial intelligence from the contested question of definition, through the history of the technology, to the classical philosophical debates that are newly relevant in the age of LLMs. The central lesson is that how we define AI is not a neutral technical matter: it shapes what we build, how we evaluate it, and who we hold responsible.

References