Resources / Lesson pack

Signal City: how AI learns

A 30-minute lesson. Pupils train a small AI with pictures, watch it fail outside the classroom, and work out why.

Draft. This pack has not yet been checked by a teacher or tried with a class.

At a glance

The lesson at a glance
ItemDetails
Ages9 to 16, with the questions adjusted to your class
Time30 minutes
You needOne device for each pair, and an internet connection to open the game
Sign-upNone. Pupils give no names or personal details
LanguageEnglish only
ResultsThe free version on this website does not send you results. Pupils can show you their screens

Learning goals

By the end of the lesson, pupils can:

  • explain that this kind of AI learns from examples labelled by people
  • explain why an AI must be tested on examples it has not seen
  • give one way that poor training data can make an AI wrong

Before the lesson

  • Try it first. Play Picture Lab yourself. It takes about 15 minutes.
  • Choose the country. Open Teacher setup to match the pictures and places to your class. On this website, each device keeps its own setting, and the default is the UAE.
  • Printable cards (optional). For a lesson without devices, open Teacher setup and choose Print cards. There are six sets: Weather day cards, Message, Vehicle, Hour, Tracking and Prompt cards. In the weather set, for example, groups make a rule from six past days, then test it on six new days. Every set prints in Standard, Large or Display size, with a teacher key, and some have a work sheet.
  • Check your school's policy. Make sure the lesson fits your school's rules on teaching about AI, especially for younger pupils.

Lesson plan

Lesson plan, 30 minutes in total
MinutesWhat happensYour role
0 to 5Starter: in pairs, pupils answer "How could a computer learn to tell two animals apart?"Collect ideas on the board. Keep them for the end.
5 to 12Pairs open Picture Lab and play Label and train.Ask: "What is the AI using to decide?"
12 to 20Pairs play The scenery trap, stopping to predict before they test it outside.Ask: "Why did the score fall so far?" Don't give the answer yet.
20 to 25Class discussion, using the questions below.Link the scenery trap to real uses of AI.
25 to 30Exit question, then return to the starter ideas.Ask which ideas pupils would now change.

Pupils who finish early can play How many examples?, the level between these two.

Questions to ask

  • The AI scored well at the training site. Why wasn't that enough?
  • What clue did it really learn? How could you have spotted that sooner?
  • If an AI were trained using photos of only some kinds of people or places, what might go wrong?
  • Who should be responsible when an AI makes a mistake?

Exit question

"An AI gets 95% on its training pictures. Is it ready to use? Explain your answer."

A good answer says no, not yet: it must be tested on new examples, because it may have learned the wrong clue.

Common mix-ups

Common mix-ups and what to say
What pupils may thinkWhat you can say
"The AI just knows things."It knew nothing until it had examples. Its answers come from what it was given.
"The AI understands the picture."It compares numbers. It has no idea what an animal is.
"More examples make it faster."More varied examples can make it more accurate. Speed is a different question.
"A high score means it's always right."Even a good AI makes mistakes, and a score only tells you about the examples tested.
"If it sounds sure, it must be right."Sounding sure and being right are different things. Check it in a source you can trust.
"Tracking people needs names and faces."For many jobs, a simple count is enough. Collect only what the job needs.
"A 30% chance means it won't happen."It should still happen on about 3 of every 10 days like that. A chance isn't a promise.

Activity with no devices

Sort and test. Give each group 8 paper cards showing two kinds of object, such as cups and spoons. Put a sticker on every cup. Groups write a rule for sorting them. Then give them new cards with stickers on the spoons instead. Rules based on the sticker now fail, which is the same problem as the scenery trap.

Next lessons

Each district takes about 15 minutes, and stands alone.

The other districts in Signal City
DistrictBig ideaAsk afterwards
Signal LabAI can learn from sounds, numbers and words.What did the AI really compare?
Traffic TowerPatterns over time help to predict what comes next.What might break the pattern?
Weather DeskA forecast gives a chance, not a promise.If there's a 30% chance and it happens, was the forecast wrong?
Crowd WatchAI can find groups by itself, and tracking raises privacy questions.Is this tracking needed? Who can see it? How long is it kept?
Prompt WorkshopClear prompts help, but people must still check the answers.What should stay out of a prompt?

Good to know

  • Real and simulated: the small AIs in the training levels really do learn on the device. Prompt answers are pre-written, so no real AI tool is used.
  • Invented data: all traffic, weather and crowd numbers are made up for the game.
  • Privacy: the game uses no camera. People appear only as anonymous dots, and typed prompts stay on the device.
  • Curriculum: this pack isn't mapped to curriculum standards yet.