How to Teach Design Students AI Text-to-Image Models and Avoid Biases: Personal Experience

How AI Can Turn Your Words into Stunning Images

AI text-to-image models are machine learning models that can generate realistic images from natural language descriptions. They have been developed recently due to advances in deep neural networks, such as transformers and diffusion models. These models have many potential applications in various domains, such as art, design, education, entertainment, and communication. They can also be used as tools for inspiration and creativity, allowing users to explore different visual possibilities from their textual inputs.

However, AI text-to-image models pose challenges and risks, especially concerning disinformation, bias, and safety. For instance, AI text-to-image models can create fake or misleading images that harm people's reputations, privacy, or trust. Moreover, AI text-to-image models can reflect or amplify the biases and stereotypes in the data they are trained on or the language they are conditioned on. These biases can affect the quality and diversity of the generated images, as well as the perception and interpretation of the users. Therefore, it is essential to teach university design students, especially architecture students, how to use AI text-to-image models responsibly and critically.

The Dark Side of AI Image Generation: How It Can Affect Your Design and Your Life

AI text-to-image models have significant ethical and social implications influencing design practice and education. In this part, I will review some of the main issues and challenges that emerge from applying AI text-to-image models in various contexts and scenarios.

One significant ethical and social implication of AI text-to-image models is the possibility of disinformation and manipulation. AI text-to-image models can produce realistic and persuasive images that can be utilized to deceive, misinform, or sway people's opinions, beliefs, or behaviors. For instance, AI text-to-image models can generate fake news, propaganda, or deep fakes that can damage people's reputations, privacy, or trust, as seen in recent fake Trump arrests. Furthermore, AI text-to-image models can generate images incompatible with reality, such as images violating physical laws, ethical norms, or human rights. These images can impair people's perception of reality and cause confusion, misunderstanding, or harm. Another ethical and social implication of AI text-to-image models is the possibility of bias and discrimination. AI text-to-image models can reflect or amplify the prejudices and stereotypes in the data they are trained on or the language they are conditioned on. These biases can affect the quality and diversity of the generated images, as well as the perception and interpretation of the users. For example, AI text-to-image models can favor specific social groups when prompted with neutral text descriptions (e.g., a photo of a lawyer) or generate images that are offensive, inappropriate, or harmful to specific groups of people (e.g., a picture of a woman wearing a hijab'). Moreover, AI text-to-image models can generate images not aligned with users' preferences, expectations, or values (e.g., a beautiful house photo).

A third ethical and social implication of AI text-to-image models is the possibility of losing human agency and creativity. AI text-to-image models can create images beyond human imagination or capability, inspiring and empowering users. However, AI text-to-image models can create images that are different from existing pictures or are dependent on textual inputs, limiting users' creativity and originality. Additionally, AI text-to-image models can create images that could be more transparent or explainable, which can reduce users' control and understanding of the generation process and the resulting images. These issues can affect users' sense of ownership, responsibility, and accountability for the generated images.

AI text-to-image models' ethical and social implications pose challenges and opportunities for design practice and education. On the one hand, AI text-to-image models can be used as powerful tools for design exploration, experimentation, communication, and collaboration. They can assist designers in generating diverse and novel visual ideas from textual inputs, communicate their design concepts more effectively to clients or stakeholders, and collaborate with other designers or disciplines using a common language. On the other hand, AI text-to-image models can also raise some ethical and social questions and responsibilities for designers. They can challenge designers to critically evaluate the quality and impact of the generated images, to consider the moral and social implications of their design choices and actions, and to respect the rights and interests of other people involved or affected by their design outcomes.

Why AI Image Generation Is Biased

AI text-to-image models can be biased in different ways and for various reasons. One of the sources of bias in AI text-to-image models is the data they are trained on. The data is the collection of images and texts used to teach the AI text-to-image models how to generate ideas from readers. The data can be biased if it is not representative, diverse, or balanced enough to capture the variety and complexity of the real world. For example, the data can be biased if it contains more images or texts from specific regions, cultures, languages, genders, races, ages, or professions than others. This can make the AI text-to-image models learn biased relationships between inputs and outputs and generate images that favor or exclude certain groups of people.Another source of bias in AI text-to-image models is the language they are conditioned on. The language is the natural description given to the AI text-to-image models as input. The language can be biased if it contains implicit or explicit assumptions, norms, or values that are not universal or neutral. For example, language can be limited if it uses words or phrases that are ambiguous, vague, subjective, or loaded with meanings or emotions. This can make the AI text-to-image models generate images influenced by or aligned with the language's perspective or intention.

A third source of bias in AI text-to-image models is the user they interact with. The user is the person who uses the AI text-to-image models to generate images from texts. The user can be biased if they have preconceived notions, expectations, or preferences not shared or agreed upon by others. For example, users can be biased if they use words or phrases specific to their context, background, or culture that others do not understand or appreciate. This can make the AI text-to-image models generate images that are not satisfying or appropriate for the user or other people involved or affected by the pictures.

These sources of bias in AI text-to-image models can lead to different types of preferences in the generated images and the users' experience. One type of bias is quality bias, which affects how good or bad the generated images are in terms of accuracy, realism, diversity, novelty, or creativity. Another type of bias is impact bias, which affects how positive or negative the generated images are regarding ethics, social justice, human rights, or environmental sustainability. A third type of bias is perception bias, which affects how fair or unfair the generated images are regarding representation, inclusion, respect, or dignity. These biases in AI text-to-image models can have different consequences for design practice and education. They can affect the design process and outcome regarding exploration, experimentation, communication, and collaboration. They can also affect design ethics and responsibility regarding evaluation, reflection, feedback, and improvement.

How to Use AI Image Generation Responsibly and Ethically

?AI text-to-image models can be biased, but they can also be improved. Here, I will propose some strategies and guidelines for avoiding or mitigating biases in AI text-to-image models and how to evaluate their performance and quality based on my experience.

One of the strategies for avoiding or mitigating biases in AI text-to-image models is to use better data. The data is the foundation of AI text-to-image models, so ensuring that the data is representative, diverse, and balanced enough to capture the variety and complexity of the natural world is crucial. To achieve this, some possible steps are: collecting more data from different sources and domains; annotating or labeling the data with relevant metadata and attributes; cleaning or filtering the data to remove noise or errors; augmenting or synthesizing the data to increase its size or diversity; and sampling or splitting the data to avoid overfitting or underfitting. Another strategy for avoiding or mitigating biases in AI text-to-image models is to use better language. The language is the input of AI text-to-image models, so it is vital to ensure that the language is clear, precise, and neutral enough to avoid ambiguity, vagueness, or subjectivity. To achieve this, some possible steps are: using words or phrases that are specific, concrete, and objective; using words or phrases that are inclusive, respectful, and appropriate; using words or phrases that are consistent, coherent, and logical; using words or phrases that are relevant, informative, and accurate; and using words or phrases that are creative, expressive, and engaging.

A third strategy for avoiding or mitigating biases in AI text-to-image models is to use better user feedback. The user feedback is the output of AI text-to-image models, so it is essential to ensure that it is transparent, explainable, and actionable enough to enable user control, understanding, and improvement. To achieve this, some possible steps are: providing user feedback that shows how the generated images are related to the input texts; providing user feedback that shows how the data or the language influences the generated images; providing user feedback that shows how the generated images can be modified or customized by the user; providing user feedback that shows how the generated images can be evaluated or verified by the user; and providing user feedback that shows how the generated images can be shared or used by the user.

These strategies for avoiding or mitigating biases in AI text-to-image models can help improve their performance and quality in terms of accuracy, realism, diversity, novelty, creativity, ethics, social justice, human rights, environmental sustainability, representation, inclusion, respect, and dignity. However, these strategies could be better and more complete. They require constant monitoring and testing to ensure their effectiveness and validity. They also need continuous learning and updating to adapt to new situations and challenges. Therefore, using some methods and tools for evaluating AI text-to-image models systematically and rigorously is essential.

Some of the methods and tools for evaluating AI text-to-image models are quantitative metrics that measure how well the generated images match the input texts or the ground truth images; qualitative metrics that measure how well the generated images satisfy the user needs or expectations; comparative metrics that measure how well the generated images perform against other AI text-to-image models or human-generated pictures; participatory metrics that measure how well the generated images involve or affect other people involved or affected by them; and ethical metrics that measure how well the generated images comply with ethical principles or standards.

The Future of AI Image Generation: What You Need to Know and Do

AI text-to-image models are fascinating and challenging. They can offer many opportunities and benefits for design practice and education but pose many risks and responsibilities. In this article, I have discussed some key aspects of AI text-to-image models and how to avoid biases, especially from educators, in this process.

?This text has provided valuable insights and information for university design students, especially architecture students interested in or using AI text-to-image models for their design projects or assignments. I also hope this text has encouraged critical and creative thinking and discussion among the students and their educators about the potential and challenges of AI text-to-image models for design practice and education.?However, I acknowledge that this text needs to be more comprehensive and definitive. It is based on our knowledge and perspective, which may be limited or biased in some ways. It is also based on the current state of AI text-to-image models, which may change or improve. Therefore, we invite the readers to further explore and learn about AI text-to-image models from other sources and perspectives, such as books, articles, websites, podcasts, videos, courses, workshops, events, or experts.

?Finally,?I invite the readers to experiment and play with AI text-to-image models using different data, languages, user feedback, and evaluation methods. The best way to learn about AI text-to-image models is to use them in practice and reflect on their outcomes and impacts. I encourage the readers to share their experiences and findings with others, such as peers, educators, clients, stakeholders, or communities. I also encourage the readers to give feedback and suggestions to the developers or researchers of AI text-to-image models to help them improve their models and address their biases. ?AI text-to-image models are not just tools or toys. They are powerful and complex systems that can create unique and meaningful images from texts. They can also create problematic and harmful photos for readers. They can inspire us and challenge us as designers. They can also influence us and affect us as humans. Therefore, we should use them wisely and carefully for design practice and education. I hope you found this newsletter valuable and informative; please subscribe now, share it on your social media platforms, and tag me as Iman Sheikhansari. I would love to hear your feedback and comments!