Technology & Computing · AI & machine learning
What artificial intelligence is
AI is one label for many machines, from spam filters to chatbots. See what they share, how learned systems differ from hand-written rules, and how to test a claim about what AI can do.

What you will learn
- Describe what systems called AI have in common, using the OECD's definition
- Explain why tasks often stop being called AI once computers can do them
- Describe how a rule-based expert system works and where it gets stuck
- Explain how a learning system gets its behavior from examples
Lesson 1 of 4 · 4 min
One word for a spam filter and a chatbot
Your email's spam filter, the route your map app suggests and the chatbot that drafts your messages are all sold as artificial intelligence. They work in very different ways. So what makes them the same kind of thing?
A 2025 report from AAAI, a large society of AI researchers, borrows a comparison from a 2024 book by Narayanan and Kapoor. Calling all of this AI is like having one noun, vehicle, for bicycles, rockets and container ships. Almost anything you say about vehicles is true of only a few.
The OECD, a group of governments, approved a definition in 2023. An AI system is a machine-based system that infers from its input how to generate outputs such as predictions, content, recommendations or decisions. It works toward objectives that people may state outright or leave implied.
That gives you a test you can apply. A spam filter takes an email and infers a decision, spam or not. A map app takes your position and infers a recommendation, a route. A chatbot takes your message and infers content, a reply.
In 1955 John McCarthy, a young mathematician at Dartmouth College, picked the name artificial intelligence for a new field. His proposal rested on a bold guess. Every feature of intelligence, it said, could in principle be described so precisely that a machine could simulate it.
Playing chess was once a standard sign of intelligence. So were recognizing speech and reading images. After computers managed each one, people tended to file it under routine computation instead. This habit is called the AI effect.
A saying known as Tesler's theorem sums it up. AI is whatever hasn't been done yet. By 2000, tools from AI research were in wide use, yet in the 1990s they were rarely called AI. So the word tends to point at whatever is newest.
Try to recallWhy do tasks such as playing chess often stop being called artificial intelligence once computers can do them?
Once a machine can do a task, people tend to see it as routine computation instead of real intelligence. This is the AI effect, summed up as "AI is whatever hasn't been done yet." So the label keeps moving to tasks machines can't do yet.
AI names a family of machine-based systems that infer decisions, recommendations or content from their input. Which tasks count keeps shifting, because people stop calling a task AI once it's done.
Lessons in this topic
One word for a spam filter and a chatbot
See what the many systems called AI have in common, and why the line around the word keeps moving.
Hand-written rules or lessons from examples
Compare a 1970s medical program built from experts' rules with a spam filter that learned from a pile of email.
Why a Go champion program can't make coffee
See why today's systems are built for narrow jobs, what artificial general intelligence would mean, and why researchers disagree on whether it's near.
How to judge a claim about what AI can do
Learn what demos, benchmark scores and marketing can hide, and the questions that tell a real ability from a staged one.
Three quick stories from this topic
Bright red marks spam
A spam filter that learned from email found bright red text was a top sign of spam
Paul Graham's 2002 filter learned from about 4,000 spam and 4,000 good emails. One of its strongest clues was ff0000, the web code for bright red, which ranked with any pornographic term. He had never thought to look for it.
Six months of rules
Graham first spent about six months hand-writing spam detectors. The last few percent of spam got very hard, and stricter rules caught more good mail.
Learned from examples
The learned filter scored each word by how strongly it pointed to spam. On his own mail it missed fewer than 5 in 1,000 spams and caught no good email.
The chess machine hoax
A chess machine beat Napoleon and Benjamin Franklin with a person hidden inside it
The Turk was unveiled in 1770 for Empress Maria Theresa. Its presenter opened the doors to show clockwork, but a chess master hidden inside moved the pieces with levers and magnets. It toured for 84 years.
Built to be seen
The cabinet was designed so every door could be opened to the audience. That kept up the illusion of a purely clockwork machine.
Demos still hide things
In 2023 Google's Hands on with Gemini video looked like real-time conversation but was edited. Ask whether a demo ran live.
Doubts about scaling
Most respondents to a 2025 AI research survey doubted that scaling today's AI reaches AGI
In a survey of the AAAI research community with 475 respondents, 76% said scaling up current approaches was unlikely or very unlikely to produce artificial general intelligence. Several large companies say AGI is their goal.
What AGI would be
AI that can complete nearly any cognitive task at least as well as a human. No system does that yet, and there's no single agreed definition.
Narrow by default
Most AI systems today are narrow, built for one task. AlphaGo beat Lee Sedol at Go in 2016, and Go was all it could play.
Sources
- AAAI 2025 Presidential Panel on the Future of AI Research, Association for the Advancement of Artificial Intelligence
- Updates to the OECD's definition of an AI system explained, OECD.AI
- Dartmouth workshop, Wikipedia
- AI effect, Wikipedia
- Artificial intelligence, Wikipedia
- A Plan for Spam, Paul Graham
- Mechanical Turk, Wikipedia





