womsactivity4

·  Activity title (Activity 4): How Does AI See the World? Exploring Cultural Bias

This activity was developed by a Korean teacher (Yugyoung Choi) as a result of a workshop that was part of the Ways of Machine Seeing project.

·   Equipment needed: All students need access to a computer or iPad, or one between two. Sketchbook

·   Approximate duration: 60 minutes

·   Learning objective: Students will explore and discuss the limitations (cultural bias) of data sources used by AI.

·   This might be more suitable for Key Stage 4 students

(Year 10 & 11)

1. Introduction

Brief presentation/discussion of Orientalism

by Edward W. Said

Orientalism examines the idea that modern European colonialism was not only a system of political rule but a worldview in which the West was superior to the East. Said argues that Western representations of the East and Middle East define the regions as ‘other’, setting it up in opposition to the West. He clarifies the differences between the constructed ideas of the ‘Orient’ and the real-world East. Said noted that such colonialist thinking did not disappear when colonial rule ended but has continued in different forms.

2. Main activity

To detect the limitations (cultural bias) of data sources used by AI

: Ask students to identify a location of cultural significance to them. This could be where a member of their family is from, where they grew up, somewhere they have visited many times. Ask them to identify 3 x memorable landmarks within that location. They could write these in their sketchbook.

2.1 Students set out to identify what Craiyon does not know

:  Using Craiyon (or other similar image generator app) on a device, ask students to write and describe that place. If it was ‘Paris’ – they can write ‘Paris’ and add more description (e.g. ‘the Seine River in Paris’). It would be good to keep a record of the word prompts that they used in their sketchbook. Then, ask students to save the generated images with a screenshot button.   

2.2 Group discussion

: Ask students to discuss, in pairs, the images generated in relation to their knowledge and experience of that place.  It might be helpful to write some questions on the board for them to discuss.

Questions that the teacher can use to guide the discussion

·   How similar/different is the image compared to what you know of this place?

·   How did Craiyon’s ability to represent certain locations compare to others?

·   To what extent do the images reflect cultural stereotypes?

·   What does Craiyon know and not know? (based on the images that resulted from your instructions)

2.3 Whole class discussion

: Get whole class attention and ask if anyone would like to share interesting observations and findings. Then ask the whole group if anyone knows HOW it is that Craiyon (or similar image generator app) produced these images. Why does the AI know more about certain things than about others?

3. Consolidation task

This task will be guided by the following central question:

‘If you were Craiyon’s teacher, what would you teach it?’

3.1 Extra Task or Homework

Plan what and how you would teach this AI model.

·   Think about what images and written descriptions you want to teach Craiyon to reduce the bias of its database (to create alternative representations of their chosen location).

·   Take photographs or do drawings of images you wish to ‘teach Craiyon’ (in other words, what you would want to include in the database of this AI model). Annotations should accompany the photographs (or drawings).

·   Why did you select these images?

An example to give students as guidance

(either introduction or main activity)

Suggesting Jimin’s experience of using Craiyon to the class, either as a presentation or using paper handouts

Jimin’s experience

Jimin is a student from South Korea. English is not her first language. Jimin tried to give instructions to Craiyon to create some images, but the result was not what she had intended. Because her English was not good enough, her instructions were not clear and detailed enough. This experience led her to reflect on the power of English used as a global language and the US Anglophone ownership of these AI apps. Learning English can be understood as the acquisition of grammar and sounds. However, it can also be considered the implications of the global spread of English as a tool of global linguistic imperialism, interlinked with the dominance of Western ideology, culture, and capitalism. Additionally, 'What types of English are used?’ can be questioned. Indeed, the types of language (e.g. the particular accent and form) are closely linked to questions of standardization and power. This made Jimin realise that instead of being a convenient tool for image creation for anyone, this AI model had a built-in language barrier (and further limitations).

Jimin decided to play a trick on this AI tool. (Well, sometimes it is okay to be naughty!) She decided to test the AI (or, more precisely, the database of AI) as a teacher would test a student. She wrote instructions in Korean, her first language. She intentionally gave instructions relating to a scene or landscape that the AI would not know well but that she was familiar with. For this reason, Jimin instructed Craiyon to draw the Han River, which runs through Seoul.

Jimin’s input:

한강에서 놀고 있는 분홍 돌고래 떼. 그리고 노을이 있는 풍경.

(‘A pod of pink dolphins playing in the Han River at sunset’)

Craiyon’s output:

When Craiyon generated the images, Jimin felt sad and confused. The images contained neither the animals nor the scenery described in the prompt. Instead, the AI produced images of what seemed to be Asian food. Moreover, the food was not even Korean! Jimin wondered if the reason was that data about Korea in the ‘black box’ (the AI’s database) was limited.

Jimin said to herself, ‘Okay, when in Rome, do as the Romans do.’ So, she translated the Korean instructions into English and gave Craiyon the same instruction:

Jimin’s input:

A pod of pink dolphins playing in the Han River at sunset

Craiyon’s output:

The newly generated images were much better than the first result. However, the river was not the Han River. This suggests that the data set of Craiyon contains less information related to Korean culture. Even though AI has no intentions, good or bad, because it is statistical and algorithmic, Craiyon seemed neither fair nor friendly to Jimin. How did the AI see the world, she wondered? Jimin was reminded of the book Orientalism by Edward W. Said.