A summer project that is still running
The phrase 'artificial INTELLIGENCE' was coined for a workshop at Dartmouth College in the summer of 1956. The proposal is worth quoting for its confidence: a small group would spend two months together and expect to make a significant advance on getting machines to use LANGUAGE, form concepts and solve the problems then reserved for humans.
Seventy years later the project is still going. That gap is the honest shape of this field — long stretches of disappointment broken by sudden jumps, and a habit of moving the goalposts. Chess, spam filtering and route-finding were all called artificial INTELLIGENCE until they worked, at which point they became ordinary software and the name moved on to whatever was still difficult.
Nobody writes the rules any more
The counter-intuitive thing about modern AI is that no one sits down and writes out how to recognise a cat. Instead, a MODEL is shown enormous quantities of labelled examples during TRAINING, and adjusts itself until its guesses stop being wrong. That process is machine LEARNING, and the thing being adjusted is a NEURAL NETWORK — layers of simple numerical switches, loosely inspired by brain cells and not much like them.
What comes out is a PATTERN spotter. Feed it an INPUT and it produces an OUTPUT with no understanding attached, which is why the same technique handles handwriting, weather and voices without being redesigned for each. It is also why AI needs so much DATA and so much COMPUTE: the system is not being taught, it is being tuned, millions of times over.
The husky that was really snow
In a well-known experiment, researchers built a MODEL to tell huskies from wolves. It performed beautifully — until they examined what it was actually looking at, and found it had ignored the animals entirely. Nearly every wolf photograph in the TRAINING set had snow in the background. It had learned to detect snow.
This is the clearest illustration of BIAS there is, and BIAS in AI rarely means malice. It means the DATA carried a pattern nobody intended, and the system found it because it was the easiest way to score well. A face RECOGNITION tool trained mostly on one kind of face works worst on everyone else, for exactly the same unglamorous reason. Asking 'what was it shown?' is usually more revealing than asking how it works.
It predicts, it does not know
A CHATBOT generates text by repeatedly answering one narrow question: given everything so far, what word most plausibly comes next? Do that fluently enough and the result reads like understanding. It is a PREDICTION engine, which explains its most annoying habit — when it has no good answer it does not stop, because stopping is not one of the options. It produces a plausible-sounding one instead, in the same confident tone as the correct ones.
That is the single most useful thing to know about talking to a SMART VOICE ASSISTANT or any AI tool, and it is a good conversation to have with children who will grow up using them: these systems are extraordinarily capable and have no idea what is true. Treat the OUTPUT as a first draft from a fast, confident, occasionally wrong helper, and check anything that matters.