How an AI system learns
- Collect data
- Train
- Test
- Use
- Improve
Then back to step 1: Collect data
| Everyday use | What the AI does | What it learns from |
|---|---|---|
| Face unlock on a phone | Recognises whether the face in front of the camera matches the owner's | The owner's face, recorded when face unlock is set up |
| Phone keyboard | Suggests the next word and corrects spelling | Patterns in large amounts of text, and your own typing |
| Photo gallery search | Finds photos of 'dog', 'beach' or 'birthday cake' without labels you typed | Millions of labelled images |
| Maps and traffic | Estimates travel time and suggests a faster route | Speeds reported by many phones on the road and past traffic patterns |
| Shopping and video apps | Recommends products or videos you may like | What you and people with similar tastes bought or watched |
| Bank fraud alerts | Flags a payment that looks unusual for you | Your normal pattern of spending |
| Voice assistants | Turns speech in Hindi, Tamil, Bengali or other languages into text, works out the request and replies | Recordings of speech matched with written text |
| Translation apps | Translates text or speech between languages | Large collections of the same sentences in two languages |
How the bank's AI handles Ramesh's 2 a.m. payment
- Learns his usual pattern Amounts, times, places and devices.
- Payment differs on all four
- Scores it as risky
- Blocks it or asks him to confirm
AI finds patterns in data and uses them to recognise, predict or recommend. It is only as good as the data it learned from, and it can be wrong.
Key takeaways: AI is a computer system doing tasks linked with human intelligence. Machine learning lets it learn from examples instead of fixed rules. Face unlock, keyboards, photo search, maps, recommendations, fraud alerts, voice assistants and translation are all everyday AI. Today's AI is narrow, can make mistakes and can carry bias from its data.