Data Detectives: Data, Spreadsheets and AI

You need: A sheet of A3 paper · Three coloured pens · A clock or timetable of your day

1. Draw a timeline of yesterday from waking up to going to sleep.
2. Mark every moment you created data: a bus card tap, a text, a photo, a game score, a search, a library book scan, a shop receipt.
3. Colour each one: green if you chose to share it, orange if you knew but did not think about it, red if you did not know it was collected.
4. Next to each, write who might hold that data now: a company, the school, the council, a friend.
5. Count each colour. Which is biggest?
6. Pick one red item and write one way to reduce that data in future.

How you'll know: You can name at least eight pieces of data you created and explain the difference between data you shared on purpose and data collected about you.

Go further: Find the privacy settings on one app or device at home with a parent and change one setting that shares less data.

You need: Two colours of beads (or two colours of paper squares and glue) · String or a paper strip · A binary alphabet chart (A = 01000001, B = 01000010…) · Binary cards with 1, 2, 4, 8 and 16 dots

1. Practise with the dot cards: show 9 using face-up cards (8 + 1). Try 14 and 27.
2. Look up the binary code for your initials on the alphabet chart. Each letter uses eight bits.
3. Choose one colour for 1 and another for 0.
4. Thread or glue eight beads for each initial, leaving a gap between letters.
5. Swap bracelets with a partner and decode each other's initials.
6. Count the bits. How many bytes did your initials need?

How you'll know: Your partner decodes your bracelet correctly and you can explain that each bead is one bit and eight bits make a byte.

Safety: Keep small beads away from younger children who might visit the classroom.

Go further: Encode a short word such as 'kia ora'. How many bytes does it need, including the space? Macrons need more bits: find out why.

You need: 10 × 10 grid paper · A black pen or pencil · A partner and a screen to hide your drawing

1. Colour squares on the grid to make a simple picture: a fish, a bird or a letter.
2. Describe your picture row by row using 1 for black and 0 for white. Count how many numbers you needed (it should be 100).
3. Now describe it a shorter way: for each row, say how many whites, then how many blacks, and so on (for example 3, 4, 3). This is called run-length encoding.
4. Count the numbers again. How much shorter is it?
5. Read your numbers to a partner who cannot see your grid. Can they redraw it exactly?
6. Try a picture with lots of changes (a chessboard). Does compression still help?

How you'll know: Your partner redraws your picture correctly from numbers alone, and you can explain why some pictures compress better than others.

Go further: Find out why a photo saved as a JPG is smaller than a PNG, and what quality you lose.

You need: A clipboard and tally sheet for each pair · A timer · A bird ID sheet, a gate tally sheet, or a device showing Spyfish Aotearoa clips · Pencils

1. Agree your method as a class: the exact place, the start time, the length of the count and what counts (for example, only birds that land).
2. Birds: record the highest number of each kind you see at one time, so you do not count the same bird twice. Travel: tally each person once as they pass the gate. Fish: pause the clip and count the most fish of each kind in one frame.
3. Two pairs count the same thing at the same time, without talking to each other.
4. Compare your tallies. Circle any differences.
5. Discuss why they differ: hard-to-see birds, people in groups, fish hiding.
6. Write one change to the method that would make the next count more reliable.

How you'll know: Two pairs counting the same thing get results that are close, and you can explain what caused the differences.

Safety: Outdoors: stay inside the agreed boundary and with your partner. At the school gate, stand back from the kerb and never step onto the road or driveway.

Go further: Repeat the count on a different day and compare. Is the change bigger than the difference between counters?

You need: A device with Google Sheets, Excel or Numbers · The class messy dataset · A checklist of cleaning steps

1. Open the messy dataset and make a copy, so the original is kept safe.
2. Sort by the name column. Find every different spelling of the same thing (tui, Tūī, TUI) and make them match.
3. Find the duplicates: rows that are exactly the same. Decide whether they are real repeats or entered twice.
4. Find the outlier: one number that looks far too big or small. Check the tally sheet before you change it.
5. Fill blank cells if you know the answer, or write 'missing' if you do not. Never guess.
6. Use COUNTIF to count how many of each kind, and SUM for the total. Check one by hand.
7. Write a cleaning log: every change you made and why.

How you'll know: Your cleaned totals match the checked answers, and your cleaning log lets someone else see exactly what you changed.

Go further: Use a filter to show only one place or one day, then make a chart that compares two filtered groups.

You need: Three printed misleading charts · Your clean dataset · A device with a spreadsheet · Coloured pens

1. Look at the first chart. Check where the vertical axis starts. Does it start at zero? How does that change what you see?
2. Look at the second. Are the time gaps even? Look at the third, a 3D pie. Which slice looks biggest, and is it really?
3. Redraw one misleading chart honestly by hand.
4. Make a bar chart from your own data. Add a title that asks your question, labels on both axes, and units.
5. Now make a misleading version of the same chart on purpose.
6. Show both to another group. Can they spot the trick?

How you'll know: You can name at least two tricks that make charts misleading and your honest chart has a clear title, labelled axes, units and a scale starting at zero.

Go further: Find a chart in a news story or advertisement and judge whether it is honest. Suggest one fix.

You need: A laptop with a webcam and Teachable Machine (teachablemachine.withgoogle.com) · Two kinds of object, such as harakeke and kawakawa leaves, or cans and plastic bottles · A plain sheet of paper and a patterned cloth for backgrounds

1. Open Teachable Machine and choose an image project with two classes.
2. Hold up object one against the plain paper and record at least 30 examples, turning it to different angles. Repeat for object two.
3. Train the model and test it with new objects it has not seen. Record how often it is right out of 10 tries.
4. Start a new model. This time, record object one only on the plain paper and object two only on the patterned cloth.
5. Now swap the backgrounds and test. What does the model think? What has it really learned?
6. Write a note: what went wrong, and how you would choose better training data.

How you'll know: You can explain, with your test results, that the second model learned the background instead of the object, and why biased training data makes AI unfair.

Safety: Train on objects only, never on people's faces. Wash hands after handling leaves, and do not use plants you cannot identify.

Go further: Add a third class with only five examples. How does having less data change the model's accuracy?

You need: A set of 12 scenario cards (for example: counting fish in 10,000 hours of video, choosing who gets a school award, translating a te reo Māori pepeha, spotting spam emails, marking a speech) · A sorting mat with three columns: good fit, use with care, not suitable · Pens

1. Read each scenario card in your group.
2. Ask three questions: Accuracy: how bad is it if the AI gets it wrong? Fairness: could it treat some people unfairly? Context: does it need care, culture or feelings a machine does not have?
3. Place the card on the mat and write your reason on a sticky note.
4. Compare mats with another group. Discuss any card you placed differently.
5. Write your own card about your context: should AI count birds, people or fish?
6. Turn your cards into a game for younger students, with an answer card that explains your thinking, not just a right or wrong answer.

How you'll know: For each card you can give a reason that uses accuracy, fairness or context, and your group can explain one case where a person must make the final decision.

Go further: Read about Te Hiku Media's Kaitiakitanga Licence. Add a card about who should control data about a language or a people, and explain your placement.

You need: Your group's spreadsheet and shared folder · The Privacy Act 2020 principles summary (Privacy Commissioner website) · A 'could this identify someone?' checklist

1. Check every column. Could any single value, or two values together, point to one person?
2. Remove or combine anything identifying, for example 'Room 12, bikes from Rata Street' becomes 'bikes'.
3. Check who the file is shared with. Change it to only the people who need it.
4. Make a strong passphrase from four random words and explain why it is harder to guess than 'Password1'.
5. Read three of the privacy principles and match each one to something your class did.
6. Write your line of the class data promise.

How you'll know: Nobody could be identified from your shared data, the file is shared only with the people who need it, and you can explain one privacy principle in your own words.

Go further: Find out what 'two-step verification' is and why it protects an account even if a password is stolen.