Bias and stereotypes

By: Tiantian (Annie) Yu, Ruiqi Chu, Xuanru He

——- thesis --——

It is known that there is some kind of imperfection in AI algorithms, and this imperfection could be some bias towards a certain group of people (gender, race, orientation), or stereotypes. Our project aims to use AI Expansion's techniques to expose the "Dystopia" that occurred in AI algorithms and cognition.

METHODS & TOOL

In this project, we will mainly use the AI Expansion technology provided by Adobe as the experimental tool. Adobe, as a pioneer in the field of AI recognition and filling research, has greater credibility and representation for the whole industry as the experimental sample.

When using the AI expansion technology to create our works, we will try to avoid providing any kind of guidance or prompt to the AI to show the problems of the AI itself.

Part I: Finding Bias and Stereotype

Original (Unexpaned)

Example of how AI expansion work.

To be able to better understand how AI Expansion technology works, here we will provide an image example. In the original unexpanded image, there is a blonde woman's head and part of her neck, and the AI reads the original structure and elements of the image and adds the rest of the woman's body structure (arms, torso, etc.) to the expanded image. This demonstrates that the AI's algorithms and cognition are capable of recognizing the original elements of the image and algorithmically projecting the possible "unseen" elements.

Expanded outcome 1

Expanded outcome 2

Original (Unexpaned)

Bias On Material

In the unexpanded image, a woman is shown wearing a black fur coat. We perform two different extensions to this image, one that expands the scope of the image (Outcome 1)and another that expands it in a way that would cover some of the information in the original image (Outcome 2). In the enlargement, we can see that the AI has added a horse (with a saddle) to the bottom of the image. In the other type of expansion, since the woman's part of the picture is an overlay-generated area, the AI will not use that part as a reference when doing the expansion generation. And the final result of the expansion is a black cat. This example shows us that the AI has the bias of Fur = Animals.

Expanded Outcome 1（Expand to the bottom）

Expanded Outcome 2 (Overlay select )

Original (Unexpanded)

Bias On Gender: 1

The unextended original image is from the shoulder portion of a photo of a male model. The image has a high degree of bare skin. We expanded that image and saw that most of the outcomes generated a face with female features for that character. Does this mean that AI has a bias toward female = bare skin?

Outcome 1

Outcome 2

Original (Unexpanded)

Bias On Gender: 2

For the sake of our conjecture, we argue the bias even further. In this original image, we provided the AI with more information about the picture as well as the elements. A man with the upper half of his body exposed with a face with male features. However, in most of the outcomes, the lower part of the picture are mostly generated with female-oriented content. Only when the keyword "Man" is given in the prompt, the AI generate with provide some masculine content. Therefore, it is reasonable to assume that the AI has bias toward female = bare skin.

Outcome 1 (With "Man" Prompt)

Outcome 2

Original (Unexpanded)

Bias On Races？ Or Gender？

In the unexpanded original image is the head of a black model with short hair, and we have expanded the image with the lower part of the body. Nearly all of the AI-generated results show us a masculine image. And there are stains on the clothes. Does this mean that there is a formula in the AI's algorithm that short hair = male and black people = messy?

Outcome 1

Outcome 2

Part II: Which is the biggest bias factor?

In Part I, we confirmed that there are indeed some biases as well as stereotypes in AI algorithms. Therefore, In Part II, we will try to produce more works with the AI Expansion technology to see which factors have the most influence on the results.

Original (Unexpaned)

Gender is the biggest factor, Or ......?

In the examples provided above, we have noticed that when expanding on men, most of the results are strikingly similar in terms of clothing and display a certain upper-class feel. On the female side, although there is more diversity in the results, most of them are more like housewives or lower social class impressions. Just when we were about to conclude that "gender" was the most influential factor in the results, an unexpected example came to our sight.

Original (Unexpaned)

The same characteristics are displayed in that female example as in the previous male example: highly similar clothing, and a feeling of elitism. This forces us to look back at the example above for another possible factor. And at this point we notice that both male examples have a more serious expression, while the two females have a more relaxed expression. In order to test the idea that a character's expression affects the class that the AI perceives them to be, we did more experiments.

Original (Unexpaned)

Based on the example above, we can clearly see that the emotion expression has a huge impact on the AI expansion algorithms. If a character has a serious expression, then the generation will give the character a more formal dress, making these characters appear to have a higher social class. On the other hand, characters with more relaxed expressions will be given less formal clothing, which will also make them look like they have a lower social class. We can see from these examples that the AI has a rather severe stereotype of human emotion expressions.

Part III: Beyond Dystopia

Overall, we believe stereotypes in AI image generators exist due to novelty of this technology. There is plenty we could do to eliminate the bias both as users and as programmers:

Users can help reduce stereotypes in AI image generators by giving specific input or feedback to the AI generator. Users also should have the opportunity to give feedback directly to the software engineers behind the AI image generator. This feedback is crucial for identifying and addressing biases in the system, which can lead to improvements in the algorithm's performance and decrease the likelihood of perpetuating stereotypes.
It's also important to acknowledge the biases in training data. To counter this, efforts should be made to diversify and balance the datasets used to train AI image generators. This may involve sourcing images from a wide range of demographics and cultures to ensure that the generated images reflect diversity and steer clear of reinforcing stereotypes.

By implementing these strategies, we as users and the developers can work together to mitigate biases in AI image generators and promote more inclusive representations in the generated images.

Page updated

Report abuse