Make a Model Card for a Fictional AI

Submitter: Jill Walker Rettberg, U of Bergen

——————————————————

The experiment:

This assignment first required students to learn about model cards for actual AI models, which are a way of documenting how the model was developed, what training data was used, what uses it is recommended for, and, vitally, what biases and problems the developers are aware of. Then students wrote their own model cards for a fictional AI in one of the science fiction novels or video games they had studied.

My goal was for students to not only learn about AI ethics and AI bias but also to understand how model cards are being used to document AI. I also wanted to encourage students to think critically about the sociotechnical imaginaries of AI that are expressed in science fiction.

ASSIGNMENT: In groups, create a model card for an AI character in a science fiction novel we have read, or for one of the Androids in Detroit: Becoming Human. In addition, write at least 2-300 words reflecting on why you chose to write the model card the way you did, and discussing any problems you ran into.

Be creative, but do use headings and the general style of actual model cards for AI. You can create your own template based on examples like these:
https://github.com/fau-masters-collected-works-cgarbin/model-card-template/blob/master/model-card-template.md
https://github.com/tensorflow/model-card-toolkit

Results:

Students enjoyed the task, and came up with some very creative responses. In many cases they had to make up details that were not made explicit in the novel or game they drew upon. For example, one group made a model card for the “artificial friend” Klara in Kazuo Ishiguro’s novel Klara and the Sun. They stated the training data was “the robot’s own experiences” but also added what types of data Klara can sense (visual data, temperature, sound, kinetic). The group that made a model card for the android Connor in the video game Detroit: Become Human mimicked the style of marketing, admitting the possibility of instability in the model, but reassuring customers that customer service would be available 24/7 in case of deviance.

Over all both I and the students enjoyed the assignment. If doing this again I would spend a bit more time analysing existing model cards before starting the assignment.

I think it’s important to design assignments that not only critique AI bias but also explore ways that people are trying to remedy or mitigate problems.

Contact: jill[DOT]walker[DOT]rettberg[AT]uib[DOT]no