Aiodic

Back to the table

1 Beginner Ai Artificial Intelligence Core Concepts

Artificial Intelligence

Machines doing tasks that normally need human thinking.

What it is

Artificial intelligence, usually shortened to AI, is the name for computer systems that do things we would normally say need a human mind: recognising a face, understanding a sentence, translating a language, spotting a pattern in a pile of numbers, planning a route, writing a paragraph, or drawing a picture. If a machine does something that would look like thinking when a person did it, it is probably being described as AI.

That is a deliberately broad definition, and it's broad for a reason. AI is not one technology. It is an umbrella over dozens of them, and the umbrella has been open for about seventy years. That is why it sits in the top-left corner of this table. Every other element here is a part of AI, a method for building it, a product made with it, or a concern about it. Machine Learning is a way of building AI. A Language Model is a kind of AI. Hallucination is a problem AI has. Alignment is the effort to make AI do what we actually want.

The term itself is younger than most people expect. It was coined by the computer scientist John McCarthy in a funding proposal dated 31 August 1955, written with Marvin Minsky, Nathaniel Rochester, and Claude Shannon. They asked for money to gather ten researchers at Dartmouth College for two months the following summer, on the basis of a bold guess: that every aspect of learning or intelligence could in principle be described so precisely that a machine could be made to simulate it. Seventy years on, that guess is still being tested.

It helps to know that AI has been built in two very different ways over its history.

Almost everything you hear about AI in the news now, including chatbots, image generators, voice assistants, and self-driving research, belongs to the second family.

One more distinction matters, because the word "AI" is used for two very different things. Nearly all AI that exists today is what researchers call narrow AI: very good at one kind of task, and useless outside it. A chess program can't write a birthday card, and a translation system can't drive a car. The science-fiction kind, a machine that can learn and reason across any task as well as a person, is called artificial general intelligence, or AGI. Whether and when it will arrive is one of the most argued-over questions in technology. When a headline says "AI", it almost always means the narrow kind, even when the product feels general, such as a chatbot that can chat about anything.

Governments have also had to pin the word down, because you can't regulate what you can't define. In November 2023 the OECD, a group of mostly wealthy democracies, updated its definition of an AI system. In plain English, it says an AI system is a machine that takes in input and works out for itself how to produce outputs such as predictions, content, recommendations, or decisions that affect the world around it. The key idea is "works out for itself": a calculator follows fixed steps, while an AI system infers what to do from what it has learned.

Why it matters

You already use it every day. Long before chatbots, AI was quietly running inside ordinary tools. Your email's spam filter is AI. So are the "you might also like" suggestions on shopping and streaming sites, the face unlock on your phone, the map app that reroutes you around traffic, and the automatic captions on videos. A big part of understanding AI is noticing how much of it is already invisible.

It's changing what one person can do. For creators and small business owners, the arrival of tools that write, summarise, translate, draw, and answer questions means tasks that once needed a team, or a budget, can now be started by one person in an afternoon. A self-published author can draft marketing copy, a shop owner can answer customer questions in three languages, and a consultant can turn a messy spreadsheet into a clear summary. The tools don't do the work perfectly, but they change where the effort goes.

It's being used to make decisions about you. AI systems help decide which job applications get read, which loan applications get flagged, what content appears in your feed, and which transactions look like fraud. When those systems are wrong or unfair, real people are affected, which is why Bias and Privacy have their own places on this table. Knowing roughly how AI works makes you better at questioning a decision that doesn't look right.

It's now regulated. The European Union's AI Act entered into force on 1 August 2024. It takes a risk-based approach: everyday systems such as spam filters face no special duties, systems used in areas like hiring or medical software must meet strict requirements such as human oversight, and a small set of uses, such as government "social scoring", are banned outright. Other countries are writing their own rules. If you run a business that uses AI tools, rules like these increasingly matter, and they change over time, so check the current requirements where you operate.

The hype is real, and so is the progress. AI has been through booms and busts before. Periods of excitement in the 1960s and 1980s were followed by what researchers call "AI winters", when promises outran results and funding dried up. A 1973 report by the mathematician Sir James Lighthill for the UK's Science Research Council was so critical that it led to deep cuts in UK AI research. Today's wave is different in one important way: the tools are in the hands of hundreds of millions of ordinary people, not just labs. But the old lesson still applies. Some claims made about AI are true, some are marketing, and the skill that matters most is telling them apart.

Real-world examples

The summer that gave AI its name (1955 to 1956)

On 31 August 1955, John McCarthy of Dartmouth College, Marvin Minsky of Harvard, Nathaniel Rochester of IBM, and Claude Shannon of Bell Telephone Laboratories sent out a proposal for "a 2 month, 10 man study of artificial intelligence" to be held at Dartmouth the following summer. The proposal listed the problems they hoped to make progress on: getting computers to use language, to form abstract ideas, to improve themselves, and to solve problems then reserved for people.

The Dartmouth workshop took place in the summer of 1956, and it is usually counted as the founding moment of AI as a field. The researchers were far too optimistic about how quickly the problems would fall; several of them, such as machines that use language well, took more than sixty years. But the name stuck, and so did the ambition.

It's worth knowing that the question was already in the air. Six years earlier, in 1950, the British mathematician Alan Turing had published a paper called "Computing Machinery and Intelligence" in the journal Mind. It opened by asking whether machines can think, then proposed swapping that hard-to-define question for a practical test: could a machine hold a written conversation well enough that a judge couldn't reliably tell it from a person? That idea, now called the Turing test, shaped how people thought about machine intelligence for decades.

Deep Blue beats the world chess champion (1997)

On 11 May 1997, in New York, IBM's Deep Blue computer won the sixth and final game of a rematch against Garry Kasparov, then the world chess champion. Kasparov had won the first game, lost the second, and drawn the next three, so the final score was 3.5 to 2.5. It was the first time a computer had beaten a reigning world champion in a match under standard tournament conditions.

Deep Blue is a perfect example of how AI can look like intelligence without working the way a mind does. According to IBM, it could evaluate 200 million chess positions per second. It didn't learn chess from experience the way a person does; it relied on special hardware, rules and evaluations crafted with the help of chess experts, and an enormous amount of brute-force searching. It was superhuman at one game and could do nothing else at all. That is narrow AI in its purest form, and it is also why the victory, though historic, did not lead directly to machines that could hold a conversation.

AlphaFold and a Nobel Prize (2020 to 2024)

Proteins are the tiny machines that do most of the work inside living things, and what a protein does depends on the three-dimensional shape it folds into. Working out that shape in a laboratory can take months or years. The idea that a protein's shape should be predictable from its chemical recipe goes back to the work of Christian Anfinsen, who shared the 1972 Nobel Prize in Chemistry, and predicting it by computer had been an open challenge for around fifty years.

On 30 November 2020, DeepMind (now Google DeepMind) announced that its AI system AlphaFold had reached a median accuracy score of 92.4 out of 100 in CASP14, the long-running competition where teams predict protein shapes that have been measured in the lab but not yet published. The organisers had informally treated a score of around 90 as competitive with laboratory methods. In July 2022, DeepMind and the European Bioinformatics Institute released predicted structures for more than 200 million proteins, covering nearly every protein known to science, free for anyone to use.

On 9 October 2024, Demis Hassabis and John Jumper of Google DeepMind were awarded half of the Nobel Prize in Chemistry "for protein structure prediction". (The other half went to David Baker for computational protein design.) AlphaFold is a useful reminder that AI isn't only chatbots and image tools. Some of its biggest effects are on science, in ways most people never see directly.

ChatGPT puts AI in everyone's hands (2022 to 2023)

For most of its history, AI was something done to you, inside other companies' products. That changed on 30 November 2022, when OpenAI released ChatGPT as a free "research preview": a chatbot that could answer questions, write drafts, explain ideas, and admit mistakes in ordinary conversation. OpenAI said it was built from a model in its GPT-3.5 series and trained partly with human feedback, a method known as RLHF.

The response was unlike anything the technology industry had seen. In early February 2023, Reuters reported an estimate from the bank UBS, based on web traffic data from Similarweb, that ChatGPT had reached 100 million monthly active users in January, about two months after launch. By the same analysis, TikTok had taken about nine months to reach that number and Instagram about two and a half years.

ChatGPT didn't invent the underlying technology; it rested on years of research into language models and the Transformer design. What changed was access. For the first time, anyone with a browser could talk to a powerful AI system directly, and that is why, for many people, "AI" now simply means "a chatbot".

A breakthrough nobody outside the field noticed (2012)

On 30 September 2012, a team of three at the University of Toronto, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, entered an image-recognition contest called the ImageNet challenge. Their system, now known as AlexNet, used a deep neural network trained on graphics chips originally made for video games. It got the correct label in its top five guesses often enough to reach an error rate of 15.3 percent, against 26.2 percent for the next-best entry. In a contest where gains were usually small, that gap was enormous.

Almost nobody outside computer science heard about it. But within a few years, deep learning had taken over speech recognition, translation, and image understanding, and it is the foundation that AlphaFold and ChatGPT were later built on. It's a good lesson in how AI progress often works: the turning point happens quietly, and the public notices years later.

Try it yourself

You don't need any technical background for this. Allow about half an hour, and have a free account with any AI chatbot ready.

  1. Spot the AI you already use. Write down five things you did on your phone or computer today. Next to each one, note whether AI was likely involved: spam filtering, recommendations, autocorrect, photo search, face unlock, maps, captions. Most people find at least three, which shows how much AI has been around long before chatbots.
  2. Ask a chatbot to explain itself. Open any AI chatbot and paste in this prompt: “ Explain what kind of AI system you are, in plain language for someone with no technical background. Tell me three things you are good at, three things you are bad at or often get wrong, and why. “ Read the answer critically. Does it match what you've read here? Remember that a chatbot can be wrong about itself too.
  3. Test narrow against general. Ask the chatbot something it should handle well, such as rewriting a paragraph in a friendlier tone. Then ask something that needs facts it can't check, such as the opening hours of a small local shop. Notice how confident it sounds in both cases. That confidence, even when wrong, is the hallucination problem.
  4. Check one claim. Ask the chatbot for one surprising fact about AI history, with a date. Then search for it yourself. Whether it turns out right or wrong, the habit of checking is the most useful AI skill there is.
  5. Try the hype test. Find one advert or product page that says "AI-powered". Ask yourself three questions: what task does the AI actually do, what does it learn from, and what happens when it gets it wrong? If the page can't answer any of them, be sceptical. You can ask a chatbot to help: “ Here is a product description that says it uses AI: [paste it]. In plain language, what task is the AI most likely doing, what would it need to learn from, and what questions should a customer ask before trusting it? “
  6. Pick your next element. Use what you noticed to choose where to go next on this table. If you want to know how AI learns, start with Machine Learning. If you want to use chatbots better, start with Prompt. If you're worried about the risks, start with Alignment or Bias.

Common mistakes and misunderstandings

Thinking AI is one thing. "AI" covers a spam filter, a chess engine, a protein predictor, and a chatbot, which work in very different ways. When someone says "AI can do X" or "AI can't do X", the useful question is always: which AI, trained on what, doing which task?

Thinking AI is new. The field is about seventy years old, and much of what now feels sudden rests on decades of slow work. Knowing the history, including the winters, makes it easier to stay calm about both the boldest promises and the gloomiest predictions.

Assuming today's AI understands like a person. Modern AI systems are extraordinarily good at spotting and reproducing patterns. Whether they "understand" in a human sense is genuinely debated among researchers. In practice, it's safest to treat them as powerful pattern machines that can produce fluent, confident answers that are sometimes simply wrong.

Confusing narrow AI with AGI. A chatbot that can discuss anything feels general, but it has serious gaps: it can invent facts, lose track of long tasks, and fail at things a child finds easy. AGI and Superintelligence are real topics of debate, but they describe systems that don't exist yet, not the tools on your phone.

Believing AI is neutral because it's a machine. AI learns from data made by people, and it picks up the patterns in that data, including unfair ones. A system trained on past hiring decisions can learn past prejudices. This is why Bias and Human in the Loop matter.

Believing every "AI-powered" label. Because AI sells, the term gets stretched to cover ordinary software. Some products labelled AI use nothing more than simple rules, and some make claims no technology could meet, such as reliably predicting who will commit a crime or succeed in a job. Ask what the system actually does before paying for it.

Treating AI as either magic or useless. Both reactions are common, and both lead to bad decisions. The honest picture is in between: AI is very good at some tasks, unreliable at others, and improving unevenly. The way to find out which is which, for your own work, is to try it on real tasks and check the results, which is what Benchmarks and Evals do at a larger scale.

Assuming the details stay the same. AI products, features, prices, and terms of use change fast. Anything you read about a specific tool, including here, is a snapshot at the time of writing. Check the current details yourself before relying on them.

Related elements