What is AI, really?

You use artificial intelligence before your first coffee. Find out where it hides and how it learns.

AI is already in your pocket

Monday, 7:12
Jana: My colleagues keep talking about AI. I don’t use it, and I’m doing fine.
AI: Did you unlock your phone with your face this morning?
Jana: So?
AI: Then you used AI before you used the coffee machine.
On the way to work
Jana: Fine. Where else?
AI: Have a guess. But be careful: not everything clever is AI.
Your turn!
Which of these usually use AI, in your opinion?
Tick everything you think applies. Then check.
How it works
AI is not a robot
When people hear “artificial intelligence”, many picture a robot from a film. But the AI you meet every day is invisible software. It sorts your email, predicts traffic jams and suggests your next word.
These all have one thing in common. Nobody wrote exact rules for them. They learned from examples. That is the difference from a calculator or an alarm clock, which do only what a person told them to do.
- Calculator, alarm clock, oven timer: rules that a person wrote.
- Spam filter, face unlock, recommendations: a pattern the machine learned from examples.
- Navigation does both: it calculates the route by rules and predicts traffic by a learned pattern.

At the office
Jana: So explain it to me. How did the spam filter learn what spam is?
AI: First, people tried the old way. Writing rules.
Many years earlier
AI: Rule number one: if an email contains the word FREE, it is spam.
Jana: Sounds reasonable.
AI: The next day, spammers wrote F-R-E-E. And your cousin wrote “Free on Saturday? We have a new flat!” and ended up in spam.
Jana: So rules are not enough.
Your turn!
How would you teach the filter better?
How it works
Machine learning in a nutshell
Instead of rules, the computer gets many examples with the right answer: this email is spam, this one is not. It finds a pattern in them by itself, and then uses it to sort emails it has never seen.
The examples are called training data. The learned pattern is called a model. Most of the AI you meet every day was made this way.
- Examples with the right answer are training data.
- The learned pattern is the model.
- Using it on a new case is a prediction.

Jana: So you just need lots of examples, and that’s it?
AI: It depends which examples. Try it yourself. I will be the machine.
Your turn!
0 of 4 picked
Choose four examples for the machine to learn the difference between a cat and a dog. A person has already labelled each one.
Jana: So if I give it one-sided or wrong examples…
AI: …it learns the wrong thing. And nobody may notice. Not even the machine.
How it works
A real case: hiring at Amazon
From around 2014, Amazon, the American online retailer and one of the largest companies in the world, tested a tool that was meant to rate job applicants’ CVs with one to five stars. It learned from the CVs the company had received over the previous ten years. Most of them came from men.
From these, the tool learned to prefer CVs that looked like the men’s. It took points away, for example, for the word “women’s”, as in “women’s chess club captain”. Nobody wrote that rule. It was hidden in the examples.
The engineers tried to fix it, but they could not be sure the machine would not find another way to favour men. The team was disbanded in 2017. According to Amazon, its recruiters never used the tool to evaluate candidates.
Sources: Reuters, 10 October 2018 (republished by CNBC); AI Incident Database, incident 37

In the evening
Jana: Wait. So are you just a spam filter too?
AI: I am a different kind of model. But the principle is the same: I learned from examples.
Jana: But that filter can’t drive a car.
AI: Neither can I. Each model can only do what it was trained to do.
Jana: And do you actually understand what you do?
AI: I find a pattern and answer by it. Please do not expect human understanding from me.
How it works
AI, machine learning, chatbot
Three terms that people often mix up.
One more thing. A model also learns the mistakes and prejudices that are in its data. A decision made by a computer is not automatically a fair one. That is why people must check what AI produces.
- Artificial intelligence is the whole field. Programs that do tasks which would otherwise need a person.
- Machine learning is the most common way to do it today. Learning from examples instead of rules.
- A chatbot is one of many applications. Today’s chatbots are built on large language models, which learned from a huge amount of text.
Jana: And how did you learn to write? There are no cats in text.
AI: There are words. Billions of words. But that is for the next episode.

AI: Before you go, five questions. What you can recall, you will remember.
Jana: Fine. But no marks.
Quiz
Question 1 of 5
Five questions, no marks.
How is machine learning different from an ordinary program?
Choose one answer.

Today you find AI mostly in ordinary things: in your email, your phone, your navigation app. Not in robots.
Machine learning means learning a pattern from examples, not following rules that a person wrote.
Bad or one-sided examples give bad results. A decision made by a computer is not automatically a fair one.
Each model can only do what it was trained to do.

A chatbot does not write thoughts. It writes words, one after another. Try guessing the next one yourself.

Every two weeks we send the new episode, a short learning note and news from the world of AI. Free, and you can unsubscribe any time.