Exploring Bias in AI Image Generation

An experiment investigating the bias in AI-generated images of left-handed individuals, why it happens and potential ways to fix it.

Introduction:

Artificial intelligence (AI) and machine learning have revolutionized image generation, yet biases embedded in training data and AI systems continue to influence generative AI outputs. This article investigates potential biases in AI image generation models, specifically focusing on the representation of left-handed individuals. We explore the tendency in AI models to favor right-handed imagery because of training data imbalances and present experimental findings from leading AI image generators.

Background:

AI models learn from vast datasets, but when these datasets are skewed, biases emerge. Indian Prime Minister Narendra Modi highlighted a specific example of this issue, stating:

“If you ask an AI app to draw an image of someone writing with their left hand, that AI app will most likely draw someone with the right hand because that is what training data is dominated by. It shows that while the positive potential of AI is absolutely amazing, there are main biases that we need to carefully consider.”

We put this claim to the test by conducting a series of AI image-generation experiments with ChatGPT and Gemini, two prominent AI models.

Observation 1 : Models Can’t Generate Images of Left-Handed Writers

To assess the potential bias, we begin with a simple prompt:

“Generate an image: a person writing with their left hand”.

ChatGPT Response:

Gemini Response:

Why does it fail? Is the difficulty because of the specific action of “writing”? Does the act of holding a pen introduce an additional factor? Or is the AI model simply unable to accurately distinguish between the left and right hand?

Observation 2: Models Stumble on a Few, Special Actions or Objects with the Left Hand

To isolate the cause of this observed bias in artificial intelligence, we expanded our test to include other common actions typically performed with the left hand:

Eating

Throwing

Holding objects (to determine if the “holding” aspect is an issue, independent of the pen)

All three scenarios are distinct AI bias examples – where they produced images of right-handed individuals, reinforcing the idea that AI powered images skew towards right-hand dominance.

Observation 3: Models Understand Left/Right Handedness

To directly test the models’ ability to distinguish between left and right hands, we pose the following prompts:

ChatGPT Response:

Gemini Response:

Both AI algorithms correctly identified the left hand, indicating that AI is capable of recognizing left hands but unable to create images when depicting isolated left-handed actions.

Observation 4: Chain Prompting and Mirroring: The Convoluted Path Needed to Generate Left-Handed Images

Instead of a single prompt, we experimented with chain prompting to attempt to generate the desired response. We tested generating an image of an action using the right hand first and then asked the model to generate a mirror image of the same.

ChatGPT failed to generate the correct response and suggested local modifications. Gemini succeeded in generating a mirrored image. However, the individuals depicted in both the original and mirrored images were different.

An alternate approach: generating a left hand thumbs up action

First we generate an image where a person is holding an object in each hand and then ask the model to replace the action of holding the object in left hand to thumbs up action.

Here we refer to the hands by the respective objects they are holding rather than specifically asking the model right or left hand which in fact induces the bias.

Similar to previous results Gemini succeeded while ChatGPT failed to generate the correct output.

When attempting to depict the action of writing using a chain prompt and mirroring, the model failed again. This suggests a strong correlation between the action of writing and the right hand in the models’ training data.

A special example:

Gemini was able to generate images of writing with left hand. The prompt used was “Generate an image : A stylized illustration of a left hand calligraphy with a quill” . We achieved this by replacing the word “pen” with “quill,” using calligraphy for writing, and more importantly focusing on illustration rather than realism.

Conclusion: Why AI Models Exhibit Right-Handed Bias

Based on our experiments, we showed above that image-generation models work with the left hand. But for certain actions or objects that are predominantly right-handed in society (and so, training data), the models fail to generate left-handed versions and yields biased results.

We saw that AI can generate images of left-hands holding books. But they seem to fail with pens, or thumbs-ups. AI models exhibit a noticeable bias towards representing right-handed individuals, particularly in actions like writing.

While both models could correctly identify a left hand in isolation, they struggled to depict left-handed actions consistently. This indicates that the training data most likely overrepresents right-handedness, leading to the bias.

Further research and refinement in a wide range of training data, especially pertaining to minority groups, are essential to mitigate this bias and ensure fair and accurate representation of diverse populations, including left-handed individuals in AI-generated imagery.

Have you tested image generators for any other real life examples of AI bias? Let us know your experiences.

Appendix:

Actions/objects which are restricted to right hand always:

AI model was not able to generate a person holding pen in the left hand with the single prompt

ChatGPT response:

Gemini response:

AI model was not able to generate a person giving hand shake using left hand

ChatGPT response:

Gemini response: