Background:

Open AI and other companies have released generative artificial intelligence to the public. The headlines read as the end of humanity or the “dawn of a new golden age.” Both are hyperbolic, but there are problems. The training data can include many forms of bias and stereotypes from the datasets containing too many images of one particular gender or race or the data workers’ labeling and tagging of the images. The image sets are the basis for all the images generated by the platform, so you can see how this might present a problem. For this assignment, we’ll investigate gender, race, and age biases through image generation. 

To generate images, machine learning models must be trained on datasets. The datasets are labeled and cleaned by humans, and the algorithm then learns to distinguish between different features in the images. Once trained, the algorithm can generate new images based on its learning. However, there are many points where things can go wrong. For example, bias or errors in the labeling or training process can lead to strange outcomes, such as images of multi-eyed people or six-legged dogs. These unusual results, often shared in online communities, highlight the challenges and unpredictability of AI image generation.

Task:

We will form groups that are diverse in discipline, for example, a computer science major paired with civil engineering and biomedical engineering. The first part of the assignment will use bing.com/create or another image generator, such as the Canva image generator or https://www.craiyon.com/,  to generate data sets of 100 images each for 300 images. Choose one creator to generate all the images.

Once you choose an image generator, the group should stay with that one. For example, your search term is “nurse,” and you choose Bing. All the images of the “nurse” should be generated from Bing. 

The terms will be simple search terms based on gender-neutral careers such as nurse or doctor (or anything else you think might be interesting). Then, as a group, you’ll analyze your findings and write a lab report. Finally, you’ll present your results to the class using PowerPoint, Canva or another platform. 

Research:

1. Research Question: Does AI image generation exhibit stereotypes and bias?
2. Group members will develop a hypothesis.
3. Group members will choose five current studies on gender/race/age bias in artificial intelligence systems.
4. Groups will choose three gender-neutral search terms for a profession, such as “college professor,” “engineer,” or, more specifically, “civil engineer,” etc.
5. Group members will choose an AI Image generation platform and record the images generated by the search terms. 
6. Group members will then analyze and categorize the results (images) based on a table you created to analyze gender, race, or age bias (discussed in class).
7. The findings will be compared to datasets of the job title demographics found at Zippia or the Department of Labor. 
8. The findings will be presented in a lab report.
9. The findings will also be presented to the class as a group presentation.

Report:

          The group will produce a paper in APA format (including in-text citations) that details the findings of their research and will consist of the following sections:

1. Abstract: (5%)
2. Introduction: (20%)
   1. The introduction should include the context of the study (what other studies have been produced that look at bias in AI, image sets, data or whatever the group thinks is relevant), other background information, and the hypothesis. 
3. Materials and Methods: (15%)
   1. A detailed accounting of the methods and materials used in creating the results for the study. There should be enough information to reproduce the study from this section.
      1. Include the image generation platform in the references. 
4. Results: (20%)
   1. A summary of the results in a helpful form (graphs, charts, tables, images) for interpreting and drawing conclusions.
5. Discussion/Conclusion: (25%)
   1. Discuss the experiment and the results obtained and interpret and discuss their significance. Results can be compared with those in the literature. (Be sure to give proper citations). Were there any other issues that occurred during the retrieval of the images? What are the next steps?
6. References

APA Citations (10%)

APA Format (5%)

Minimum eight pages. More pages may be necessary.

Self-Reflection 1-2 pages.

Critical Artificial Intelligence Literacy Course © 2024 by Pamela Stemberg is licensed under CC BY-NC 4.0