Definition
You say "Hey, play my favourite song." What happens before the music starts?
An AI system is a machine that takes in information from its surroundings (input), makes sense of it, decides what to do to reach a goal, and acts (output). Many AI systems also learn, so they improve over time.
In simple words
Think of a smart speaker. Its microphone hears your voice (perceive). It turns the sound into words and works out that you want a song (understand and decide). It plays the song (act). If you often skip a song, it learns to suggest it less (learn).
Why it matters
These parts appear in every AI system, from a robot vacuum to a self-driving car. When you meet a new AI product, you can ask: what does it sense, what does it decide, what does it do, and does it learn? Those four questions explain almost any AI system.
Key terms
- Agent
- Anything that perceives its surroundings and acts on them to reach a goal.
- Input
- The information a system takes in, such as a photo, a voice, text or sensor readings.
- Output
- What the system produces: an answer, a decision, an action or new content.
- Goal
- What the system is trying to achieve, such as "reach the destination quickly".
- Autonomy
- How much a system can act without a person approving each step.
- Feedback
- Information about how well an action worked, which a learning system uses to improve.
Explained in detail
Thinking of AI as an agent
A helpful way to describe any AI system is as an agent: something that perceives its environment and acts upon it to reach a goal. A thermostat, a chess program, a self-driving car and a chatbot can all be described this way. The words change from system to system, but the same questions always apply. What does it sense? What is it trying to achieve? What can it do? How does it decide?
The five characteristics
- Perception: it takes in information through sensors, cameras, microphones or typed text.
- Understanding: it turns raw input into something meaningful, such as recognising that a sound was the word "play".
- Decision: it chooses what to do, using rules, a learned model or both.
- Action: it produces an output, from a spoken reply to turning a car's wheel.
- Learning: many systems use feedback to improve over time.
Goals shape behaviour
Every AI system works towards a goal that people chose, and the choice of goal matters a great deal. A video app whose goal is "maximise the time people watch" will behave differently from one whose goal is "show people what they will be glad they watched". A route planner told to find the shortest route may send you down narrow lanes; one told to find the fastest safe route will not. When an AI system behaves badly, the cause is often a goal that was set carelessly, not a broken machine.
Levels of autonomy
Systems differ in how much they do on their own. A spelling checker only suggests; you decide. A spam filter acts by itself, but you can check the spam folder. A robot vacuum plans and acts with no help at all. More autonomy means less work for people, but it also means mistakes can happen before anyone notices. That is why high-risk systems, such as those in medicine or driving, are usually designed so that a person can watch and step in.
Describing a system in four lines
| System | Senses | Goal | Acts by |
|---|---|---|---|
| Self-driving car | Cameras, radar, map | Reach the destination safely | Steering, braking, speeding up |
| Spam filter | Text and sender of each email | Keep unwanted mail out | Moving mail to the spam folder |
| Crop-advice app | Photo of a leaf | Identify the disease | Showing the name and advice |
How it works
| Part | Job | In a smart speaker |
|---|---|---|
| Input (perceive) | Take in information through sensors | Microphone hears your voice |
| Understand | Turn raw input into meaning | Sound becomes the words "play my song" |
| Decide (reason) | Choose an action that meets the goal | Pick which song and from where |
| Output (act) | Do something in the world | Speaker plays the music |
| Learn | Improve from feedback | Notices songs you skip |
Engineers call a system like this an agent: something that senses its environment and acts in it to reach a goal.
Try it yourself
Follow your voice through a smart speaker
You say “Play some music”. Every AI system has these same parts.
What happens between your question and the answer?
Put the smart speaker's steps in the right order.
- 1The system works out what you want
- 2The sound is turned into words
- 3The microphone picks up your voice
- 4The speaker gives the answer or plays the song
- 5It learns from whether you were happy with the result
- 6It chooses the best action
Label the parts of a robot vacuum
Match each part of the robot to its job.
- Bump sensors and camera
- Map of the room in its memory
- Choosing the next spot to clean
- Wheels and brushes
Common mistakes
- Thinking AI needs a body. Many AI systems are only software, with a screen as their output.
- Forgetting the goal. An AI system always works towards a goal that people set.
- Thinking every AI system learns. Some only follow knowledge given to them.
Check yourself
Summary
- An AI system perceives, understands, decides, acts and often learns.
- Inputs come from sensors; outputs are actions or answers.
- Ask four questions of any AI product: what does it sense, decide, do and learn?
Finished reading?
Sources: Russell & Norvig, Artificial Intelligence: A Modern Approach · AI4K12: Five Big Ideas in AI