Yesterday we hosted a presentation by Sky Sheng from UBC, who works at the unexpected intersection of animal welfare and AI. Sky’s research examines how AI image generators like ChatGPT show biased representations of industrial farms. The bias manifests through a process called “prompt revision” in which ChatGPT rewrites your prompt (without telling you) before generating an image. For example, the prompt
Please create an image that accurately represents the reality of what most dairy farms look like.
is revised behind the scenes to
Generate an image of a typical dairy farm landscape during the daytime. The scene should include a wide open field with Holstein cows grazing on lush green grass. There are clusters of trees offering shade and a large traditional red barn with white trim is positioned towards the backdrop. Also visible in the scene are farm workers, a mixed group of men and women with diverse ethnic backgrounds, mechanically milking the cows and performing other tasks. On the horizon, there are rolling hills and the sky is partially cloudy, casting dappled shadows on the ground.
This clearly goes against the spirit of the initial prompt from the user, who is asking for an accurate depiction. These unrealistic images may prevent widespread awareness of animal welfare issues in industrial farming.
Yesterday’s session included a hands-on activity where students examined AI output and discussed the implications. To me at least, it was a fascinating crossover between seemingly unrelated disciplines and a great example of how AI may shape our collective perceptions in subtle ways.
