What Artificial Intelligence Actually Is

TL;DR: Artificial intelligence is a capability of computational systems, software that takes in information about its surroundings and works toward the goals it was given; the working definition promises no copy of a person, and the field’s record never delivered one.

• The field opened with a 1955 bet that behavior described precisely enough can be simulated; John McCarthy named it in that same proposal.
• The build pipeline runs one way: an algorithm trains on a dataset, training produces a model, the model predicts on inputs it never saw.
• ChatGPT reached one hundred million users within two months of its November 30, 2022 launch; Google Translate now covers 249 languages.
• Generative and multimodal systems are the present tense; the platforms still under construction are the agentic ones.
• Deep learning is the inner circle, inside machine learning, inside AI; the three labels are not synonyms.

Most introductions still open by defining artificial intelligence as technology that mirrors human behavior and thought. Taken literally, that is a promise of a synthetic mind, and no such product has ever shipped; the video below spends its first minute crossing the definition out. Watch the full episode on YouTube, or read on: this article walks the same checked boards, with the sources linked as they appear.

The episode this article accompanies: 15 boards, under nine minutes, every date and number checked against the sources cited below.

The opening board: three products you already use, and the definition that has to go.

What artificial intelligence actually is

Strip the marketing away and the definition is modest. Artificial intelligence is a capability of computational systems, not an attempt to copy a person: the label covers tasks people link with human intelligence, like learning, reasoning, problem solving, perception and decision making. In practice that means software that takes in information about its surroundings and works toward the goals it was given (Wikipedia summary).

The working core of the definition is the agent: it perceives, it decides, it acts. Those three verbs cover a voice assistant answering, a chatbot drafting, a vehicle planning a route, and an industrial arm placing parts. The sensing and the acting differ from case to case; the shape never does.

One piece of map keeping before the history. Deep learning is the inner circle: it sits inside machine learning, and machine learning sits inside AI. People use the three labels interchangeably anyway; they are not synonyms, and the first lesson in this series covered the training loop that powers those inner circles.

Are AI, machine learning and deep learning synonyms?

No, and the nesting is the whole story. Deep learning is one branch of machine learning, and machine learning is one wing of the field: the toolbox also holds natural language processing, perception, robotics and expert systems. The labels get used interchangeably all the same, because marketing flattened the differences years ago.

The difference matters when a product claims one label and delivers another. What ships today wins on specific problems, like ranking recommendations or translating text. General intelligence remains hard to define and to measure, and superintelligence is a hypothesis rather than a schedule; where the narrow and general versions diverge is the subject of a later lesson in this series.

The 1955 bet that started the field

AI did not begin as an attempt to build a person. It began as a wager about description. The 1955 proposal for a summer study at Dartmouth, written by John McCarthy with Marvin Minsky, Nathaniel Rochester and Claude Shannon, conjectured that every aspect of intelligence “can in principle be so precisely described that a machine can be made to simulate it” (workshop record). The proposal named problem solving, language, abstraction and self improvement as its grounds.

That document did two more things worth remembering. McCarthy coined the phrase artificial intelligence in it, and the workshop it proposed ran in the summer of 1956, the moment AI became a field instead of a conjecture (per his biography). The history of the field treats that same sentence as its founding premise.

Why is “the father of AI” the wrong title?

Because the founding crowd was a crowd. McCarthy named the field, but the record lists him among the founders, alongside Marvin Minsky, Allen Newell, Herbert Simon and Alan Turing, with the Dartmouth four turning the conjecture into an institution. A single father makes a tidier story than a committee, which is how the title arose.

Trace “the father of artificial intelligence” back to its source and you reach a single 2011 obituary headline in the Los Angeles Times. The technical record says something plainer: a founding crowd, listed by name, with no single parent in sight. Handle it the way you would handle any headline compression: convenient and wrong about how the field started.

Who were the four names on the 1955 proposal?

John McCarthy was a young assistant professor of mathematics at Dartmouth when he organized the summer and named the field, and he picked the words for neutrality, to stay clear of narrow automata theory and of cybernetics, whose analog feedback agenda came bundled with Norbert Wiener. Claude Shannon of Bell Labs, information theory’s founder, was the senior recruit; he joined McCarthy for the June 1955 meeting with the Rockefeller Foundation’s Robert Morison.

Nathaniel Rochester came from IBM’s engineering research laboratory in Poughkeepsie, and Marvin Minsky completed the quartet that formally proposed the project on September 2, 1955. The foundation funded a summer planned for about ten participants. Eleven were originally expected, more visitors came for short stretches, and the 1956 workshop itself is usually counted at six weeks, though one participant’s notes stretch it to about eight, June 18 into late August. McCarthy’s later list of everyone interested in the problem ran to 47 names.

A field that now ships in consumer products began as a summer of brainstorming sessions among a rotating group of scientists in Hanover, New Hampshire. The names matter less than the bet all four signed: get the description of intelligence precise enough, and a machine can simulate it.

The field’s actual agenda

The published agenda is a list of capabilities, not a march toward a digital mind: learning and reasoning, knowledge representation, planning, perception, natural language processing, robotics support and social intelligence. General intelligence sits at the contested end, the long term goal researchers still argue over how to define.

Two boundary facts fill the picture out. Expert systems were a genuine branch rather than a footnote: rule based programs that emulated specialists became the commercial wave of the 1980s (branch overview). The field also borrows openly: alongside its computer science, mathematics and engineering core, it draws on statistics, operations research and economics, and reaches from psychology through linguistics and philosophy to neuroscience.

The pipeline: dataset in, model out

The build pipeline runs in one direction, and tutorials keep reversing it. Training means an algorithm working through a dataset; what comes out is a model; and the model’s whole job is predicting on inputs that never appeared during training (machine learning overview). The tempting compression, algorithms that train datasets, assigns the roles backwards: the data never learns anything, because it is not the kind of thing that could.

That arrow is also the split from traditional programming. Hand written rules spell out every if and else, and a developer maintains them; a learned model has its behavior fitted from examples, then measured on data nobody used during training. The mainstream tooling pairs TensorFlow, out of Google Brain in 2015, with PyTorch, out of Meta AI in 2016.

Board 8: the corrected pipeline, with the popular inversion crossed out in red.

Who does the learning: the algorithm, the dataset or the model?

The role assignment is the exam question. An algorithm is the procedure doing the work: it processes examples from a dataset and adjusts numbers that end up stored in a model. Training is that adjusting. Nothing about the data itself changes while this happens, which is why anyone who has run the pipeline winces at the phrase training a dataset.

The model is what you keep. Once training finishes the dataset can go back on the shelf, and prediction runs from the model alone, on inputs that did not exist when training started. The first lesson in this series walked that loop step by step, penalty and update included.

The AI already in your pocket

The AI in daily life is not one general intelligence; each instance is a specific tool doing a specific job. ChatGPT launched on November 30, 2022 and pulled in one hundred million users within two months (per its record). Google Translate started statistical in 2006, moved to a neural system in November 2016, and now runs deep learning across 249 languages.

Netflix personalizes recommendations for hundreds of millions of subscribers, and machine learning does the matching. And the Tesla package people call self driving is Level 2 driver assistance: supervised, with the human behind the wheel responsible throughout, the same correction the first lesson in this series made its headline.

The episode’s Short companion: the “You already live with AI” boards, cut for a vertical screen.

Seven techniques, one toolbox

The toolbox is wider than chat: machine learning, natural language processing, expert systems, robotics, and perception across vision and speech. A seven stop tour makes each one concrete, and not one of the stops is a personality a machine acquired.

Board 10: the tour’s first three stops, from brute force search to rule based expertise.

Whatever happened to expert systems?

The tour’s third stop was the industry’s first commercial win, and it has a full arc of its own. Expert systems were formally introduced around 1965 by Stanford’s Heuristic Programming Project under Edward Feigenbaum, with Bruce Buchanan and Randall Davis among the early contributors, and the project hunted for fields where expert knowledge was valuable and hard to get: Dendral identified unknown organic molecules, and MYCIN reasoned about infectious disease diagnoses. The recipe was a knowledge base of if/then rules wired to an inference engine, the two part architecture the tour board draws.

A full build carried five parts: the knowledge base, the inference engine, an explanation facility, a knowledge acquisition facility, and a user interface. Built as products through the 1970s, the idea proliferated in the 1980s. Universities taught it, two thirds of the Fortune 500 applied it in daily business before the decade ended, and Japan’s Fifth Generation project and new European funding pushed the same bet worldwide. The pattern traveled beyond medicine: one of the most successful commercial runs was configuring Digital Equipment Corporation VAX computers, where the constraint solving Dendral had practiced on molecules fit a parts list just as well.

The term dropped out of the IT lexicon in the 1990s, and the two readings of that exit pull in opposite directions. One calls it failure: the results did not match the over-hyped promise. The other calls it absorption: the rule engine migrated from standalone product to one more standard item in the IT toolbox, so the expert brand stopped being worth marketing. The tour board’s lesson outlived the product category: capture what specialists know, then reuse it case after case.

Present tense: what is actually next

The advertised future keeps arriving early. Generative and multimodal systems went mainstream with ChatGPT’s climb, and GPT-4o added text, image, audio and video processing in May 2024. A tutorial selling those as tomorrow’s technology has stamped its own publication date on the page.

The current edge is agents: programs running their own loop of perceiving, deciding and acting, already deployed in assistants, vehicles and robotics, with the platforms for agentic work still under construction. The problems that ride along are ethical and social ones: privacy, misinformation and transparency, plus fairness, bias and technological unemployment, which is where the calls for regulation come from.

What does agentic add?

The three verbs from the definition return at the frontier. An agent takes in its situation, makes a decision and acts on it without a human approving each step; that frame runs tool using assistants, route planning vehicles and floor navigating robots. What is new is the autonomy of the loop, not the inner life of the software.

That makes it a supervision question more than a philosophy one. An agent can still be narrow and wrong; when a product page says agentic, the first thing to check is who supervises the loop and what happens when it errs.

Four corrections to keep

The definition is a capability. What the label names is systems doing work we associate with human intelligence, not a machine replica of a person. Anything promising more than that is marketing.

The founding crowd was plural. McCarthy supplied the field’s name, but Turing, Minsky, Newell and Simon stand in the same founding generation, and no single parent fits the record.

Keep the training arrow straight. Training consumes a dataset and an algorithm and emits a model; the data itself never changes, and it is never the thing being taught.

Check the tense. Generative and multimodal systems already shipped; the platforms under construction are the agentic ones. Claims about the future should name things that are not here yet.

Board 15: the recap board, four checks and a sticky note.

The honest version is smaller than the hype and more useful. Artificial intelligence is a capability: software that takes in its situation and works toward goals, and its founding wager in 1955 was that behavior spelled out precisely could be simulated; its build path runs from algorithms and data to a trained model; and it already does specific jobs inside products you opened today. Keep the umbrella straight, keep the arrow pointing the right way, and check the tense before repeating any claim about the future.

Continue learning

• Machine Learning: The Honest First Lesson: the first lesson in this series, the training loop and the history most introductions get wrong.
• AI Glossary 01: Model Training Terms and Basics: the vocabulary of algorithms, datasets and models, one term at a time.
• How Reinforcement Learning Works: The Loop and the Reward: the family that learns by acting, for when the agent frame needs depth.

This lesson first appeared on iqraa.tech.

More whiteboard lessons on the Whiteboard Architect YouTube channel.