Course 5, lesson 50 of 100, Ages 11+

Know your dataset

Data labels for honest AI

Like I’m 5

Food packets have labels that tell you what's inside. Good datasets have labels too, telling you where the data came from and what it's good for.

The big idea

A datasheet or data card describes a dataset: who made it, how it was collected, what it contains, who is missing, and what it should and shouldn't be used for.

This helps builders avoid nasty surprises, like using a dataset of adult faces for a children's app. Model cards do the same for trained models, listing what they were tested on and where they perform worse.

Examples

  • Ingredients: A data card lists sources, dates and how labels were made.
  • Warnings: 'Mostly English text. Not suitable for medical advice.'
  • Model cards: Show accuracy for different groups of people.

How it works

  1. Record where and how the data was collected.
  2. Describe what's included and who's missing.
  3. State the intended uses and the uses to avoid.

Check your understanding

What does a data card tell you?
Options: Where a dataset came from and what it's suitable for; The price of the computer; Your password.
Answer: Where a dataset came from and what it's suitable for. It's like an honest label for a dataset.
Why list who is missing from a dataset?
Options: So builders know where the AI might work worse; To make it longer; It's not useful.
Answer: So builders know where the AI might work worse. Knowing the gaps helps prevent unfair or unsafe uses.

Remember

Datasheets and model cards are honest labels that describe data and models, including their limits.

Talk about it

Write a mini label for a dataset of your class's favourite books.

Go deeper

'Datasheets for Datasets' (Gebru et al.) and 'Model Cards for Model Reporting' (Mitchell et al.) proposed these documentation practices, now common in responsible AI work.