Generative AI and Bias

Video credit: How AI Image Generators Make Bias Worse – posted by LIS – The London Interdisciplinary School on YouTube

Think

Which of the following statements best describes your understanding of the relationship between Generative AI (GenAI) and bias?
- GenAI is always biased.
- GenAI bias is informed by the platform's algorithm and training data.
- GenAI is bias-free.
You are right! Biases arise from selected training data and algorithm design, and GenAI further perpetuates and augments the biases within the training data and algorithm. Mitigating biases in GenAI requires careful data selection, algorithm design, and ongoing monitoring efforts.
That’s true! Think of AI like hanging out with a group of friends who all have the same opinions - it can start seeming like that's how everyone feels! Biases not only sneak in from the algorithm and the data it's trained on, but they may also be further amplified as GenAI acquires and replicates these biases.
Not exactly. While AI algorithms may start with a neutral framework, they can absorb biases present in the training data, just like a sponge soaking up water. Biased prompts can also lead to biased GenAI output.
What are the potential impacts brought by bias in GenAI output? (Choose all that apply)
- Reinforcing harmful stereotypes and perpetuating societal inequalities.
- Reducing the accuracy and reliability of generated content.
- Generating misleading or false information.
This is true! Bias in GenAI output can amplify stereotypes and contribute to unequal representations in the content it produces. This perpetuation of biases can further marginalize people based on gender, race, and other characteristics, and deepen societal divides.
You are right! Bias generally leads to less accurate and reliable content, as it can skew the outputs based on prejudiced data rather than reflecting diverse perspectives.
This is a notable problem too! If the training data contains biased or incorrect information, the AI can produce outputs that contribute to the spread of misinformation.
What can we do to mitigate issues related to bias when using GenAI? (Choose all that apply)
- Critically evaluate GenAI outputs.
- Compare outputs from different AI platforms and other sources.
- Stay objective by not focusing on the biases.
This is a very important step! AI lacks human-like cognitive abilities and operates based on algorithms and training data, and may potentially produce superficial, off-topic or incorrect content. Human judgment remains crucial for evaluating AI outputs accurately and discerning their limitations. Check out this video by UBC’s Chapman Learning Commons to learn how to evaluate GenAI output.
This can definitely help! AI imperfections stem from various factors, including diverse interpretations and responses to human prompts. Comparing outputs from multiple platforms can help with quality assessment. A good idea is to cross-check AI-generated content with reliable sources like expert publications, making sure that you take into account who the author or publisher is and the purpose of the work.
This may not be the best way to handle the issue. Ignoring biases leads to unfair results and disregards the ethical duty of AI developers and users to address pervasive issues. Instead, recognizing and addressing biases is crucial for ensuring fairness in AI.
Name
This field is for validation purposes and should be left unchanged.

Explore

Generative Artificial Intelligence (GenAI) tools have become increasingly integrated into our daily lives as well as academic and work assignments. They have provided shortcuts for completing various tasks. We can now generate organized text easily using ChatGPT and Gemini, or create vivid images using DALL-E 2 and Midjourney. However, like other tools in the digital realm, there are potential issues associated with the use of GenAI tools. In another tutorial, we explored the privacy concerns when using GenAI tools, and the discussion does not end there. Another concern regarding the use of GenAI tools is the bias in their output.

As discussed in our tutorial about algorithmic bias, AI can exhibit bias depending on the dataset it was trained on. With the latest developments in GenAI, the issue of AI bias has intensified. A UNESCO report explains how Large Language Models (LLMs) can reinforce gender stereotypes and biases through biased AI recruitment tools. Unlike traditional AI models used for classification or prediction, GenAI models create new content based on patterns from their training data. This makes bias measurement challenging since there is no single “correct” output. The report summarizes the three main sources of biases in AI:

Graph showing the three main sources of biases in AI

Click here for PDF version

A paper by Zhou et al. (2024) also investigates how AI image generators can unintentionally reinforce and exacerbate societal biases concerning gender, race, and facial expressions and appearances in the images they produce. Visual content, in particular, has a strong influence on shaping perceptions and reinforcing stereotypes among viewers. For example, in the world created by Stable Diffusion, an AI tool that turns text into images, most CEOs are white men. Women rarely appear as doctors or lawyers, and people with dark skin, especially men, are unfairly shown as criminals. This research guide by UofT has provided more examples on how bias and discrimination occur in AI-generated images. These outputs might strengthen harmful stereotypes, affecting beliefs in ways that are difficult to change.

The saying “Garbage in, garbage out” is often used to highlight potential biases in GenAI output. This means quality GenAI output relies on complete, organized, and accurate data. If the training data is flawed, the resulting output will be flawed as well. As users of GenAI tools, we must recognize their limitations and understand that biases are always embedded in GenAI output depending on the datasets they are trained on. Additionally, how prompts are written and evaluated also impacts the extent to which information is biased. It is crucial to consider these factors to ensure more reliable and fair outputs from GenAI tools.

Discuss

The Digital Tattoo Project encourages critical discussion on topics surrounding digital citizenship and online identity. There are no correct answers and every person will view these topics from a different perspective. Be sure to complete the previous sections before answering the one or both questions below.

Have you encountered biases in GenAI output? What is your observation?
What steps can be taken to ensure more equitable and representative training data for GenAI tools?
What ethical responsibilities do developers and users of GenAI tools have in addressing and mitigating biases?
How can we evaluate and refine prompts to identify and reduce bias in AI-generated content effectively?

Think

Explore

Links

Discuss

Leave a Reply Click here to cancel reply.