Making AI Art Responsibly: A Field Guide to Ethical Creation

The digital landscape is rapidly evolving, and with it, the tools and techniques available to artists. Artificial Intelligence (AI) has emerged as a powerful force, enabling new forms of creative expression. However, this innovation comes with a responsibility. The 'Making AI Art Responsibly: A Field Guide,' developed by the Partnership on AI's AI and Media Integrity program, serves as a crucial compass for artists navigating this new frontier. This guide is designed to equip artists and makers with the knowledge and ethical considerations necessary to create AI-generated art with care and integrity. At its core, this guide focuses on machine learning techniques used for generating images, videos, and text. It defines AI art not merely as output from a machine, but as 'new works made with creative intent using techniques where computer programs access data and use that data to learn for themselves, such as machine learning.' This definition underscores the artist's role in guiding the creative process. Complementing this is the concept of Responsible AI, defined as 'the practice of developing and ensuring AI technologies are ethical, transparent, and inclusive for the benefit of people and society.' Together, these definitions lay the groundwork for a thoughtful approach to AI art creation. The guide acknowledges that the field of AI art is nascent and subject to ongoing debate. Therefore, instead of providing rigid rules, it structures its content around critical questions artists should pose to themselves throughout their creative journey. This approach empowers artists to make informed decisions, fostering a proactive and ethical mindset. The guide offers emerging best practices and checkpoints, encouraging a continuous process of reflection and refinement in the creation of AI art.

Before diving into the technical aspects of AI art creation, it's essential to pause and reflect on the fundamental 'why.' The Partnership on AI's guide emphasizes the importance of honest self-assessment regarding your goals for using AI in your artistic practice. This introspection is not about judgment but about establishing a clear foundation for your creative decisions. Ask yourself: What are my objectives for using AI in this work? Am I seeking recognition, financial gain, or simply exploring the capabilities of new technologies? Understanding your motivations will help you navigate the ethical considerations that arise. Consider the pros and cons of employing AI for your specific project. Could your objectives be achieved through alternative, non-AI methods? This comparative analysis ensures that AI is the most appropriate tool for your artistic vision. Furthermore, it's vital to understand the broader societal implications of AI technologies. What does it mean to create 'Computer Critical Computer Art'? Engaging with these questions encourages a deeper engagement with the role of AI in society and how your work might contribute to or comment on these dynamics. If your intention is to address social or political issues through AI art, be mindful of potential biases or the 'Creative Savior Complex,' ensuring your commentary is authentic and well-considered. This initial phase of self-inquiry sets the stage for a more responsible and impactful creative process.

The dataset is the bedrock upon which all AI art is built. The selection and curation of this data are inherently subjective acts, and the Partnership on AI's guide urges artists to approach this stage with utmost care to avoid exploitation or harm. The core of Checkpoint 1 revolves around understanding the origin and nature of your training data. Key questions to consider include: Is there content in my dataset that might infringe on copyright? Is it in the public domain or available under Creative Commons licenses for non-commercial use? If you are using an existing dataset, do you understand how and why it was created? When scraping data from public forums or social platforms, how do you relate to these communities? Do you possess more or less power than other community members? It's crucial to understand the historical and social context of the media you are using as training data. Consider how you might contact the creators for permission to use their work and how to include them in the process. The guide presents compelling examples: Estebán Salgado creates his own datasets by algorithmically generating abstract vector shapes, training models on them for meditative animations. In contrast, the project 'This Furson Does Not Exist' trained a StyleGAN2 model on over 55,000 artworks scraped without permission from a furry art forum. This led to protests from original artists, citing disrespect and copyright infringement. This highlights the critical need to respect data creators and subjects. Diversity within the dataset is paramount. Is the training data diverse enough to ensure the model doesn't produce near-copies of the original data? What is represented, and what might be missing, and why? Are there ways your dataset might be skewed, reinforcing stereotypes about race, gender, or other traits? The guide also prompts artists to consider how they might collaborate with and include people featured in their datasets, and how to contact original data creators for permission. Stephanie Dinkins' work, creating custom AI systems from community-focused data, such as oral histories from Black families, exemplifies a collaborative and respectful approach to dataset creation.

Once your dataset is prepared, the next critical step involves the model code – the algorithms and frameworks that process your data and generate the AI art. Checkpoint 2 of the guide prompts artists to critically examine the origins and implications of the code they utilize. A fundamental question is: Whose code are you depending on to make your work? How was this code developed and labeled, and by whom? Are you comfortable with the processes involved? Understanding your relationship with the creators of the tools you are using is essential. Most likely, your model code will build upon existing frameworks developed by researchers at universities, government agencies, or companies like NVIDIA. Learning the history and supply chain of the AI architectures you are using is an integral part of your artistic practice. The guide also sheds light on the 'Politics of Classification.' The code you use may rely on datasets annotated by individuals on platforms like Amazon's Mechanical Turk for low wages. Kate Crawford and Trevor Paglen's work, 'Excavating AI,' details problematic and offensive labels found on ImageNet, a large visual database. This raises questions about the human labor and ethical considerations embedded within the code's development. Therefore, artists are encouraged to ask: Am I appropriately acknowledging the people and labor that went into the code used to produce the work? Am I respecting the people who contributed to the model code? If you are using someone else's code, are you thanking and crediting them? The guide also touches upon the complex issue of ownership in AI art, referencing Robbie Barrat, an AI artist who open-sourced a GAN model. His work led to the sale of an AI-generated piece for $432,500, with Barrat receiving none of the proceeds. This scenario highlights the ongoing legal ambiguity surrounding ownership relationships between AI frameworks, tools, models, and their outputs.

With your dataset and model code ready, the next phase involves the practicalities of training your AI model. Checkpoint 3 focuses on the resources required for this process and, crucially, their environmental impact. Training an AI model can be very resource-intensive. The guide highlights that training a single AI model, like the popular Transformer deep learning model, can emit over 626,000 pounds of carbon dioxide equivalent – nearly five times the lifetime emissions of the average American car. Considering that these models are often trained multiple times by various researchers, the cumulative emissions are significant. Artists are prompted to ask: What are the environmental costs of my training? How long do I need to train to reach satisfactory outputs? Can I ensure I don't overtrain with diminishing returns? Do I need to train a new model at all, or can I use existing pre-trained models? The guide suggests methods like transfer learning to reduce environmental costs, avoiding training from scratch. It also recommends using tools like the Machine Learning Emissions Calculator to compute expected GPU carbon emissions from training. Beyond environmental concerns, the guide also touches upon the practical needs for training, such as GPU machines and other training resources. The question of whether the output will be for profit or publicity is also raised, linking back to the artist's initial intent. The overarching message is to be mindful of the energy consumption and carbon footprint associated with AI model training and to explore more sustainable approaches whenever possible.

You've trained a model, and you have something ready to share with the world! Checkpoint 4, the final stage in the guide's framework, focuses on the crucial aspects of publishing and attributing your AI art. This phase is about responsibly disseminating your work and acknowledging those who contributed to it. How are you crediting and thanking the people involved with the model code and dataset? This is a paramount question. The guide emphasizes that the more you share about your process, the more others can learn. It also highlights the potential for your work to be misused for profit or political motives, urging artists to consider the threats and unintended consequences associated with publishing their work. Ask yourself: Who might benefit from this work? How might I make my work accessible to others to support discussion and learning? Have I documented my work with explainable AI fields to be transparent about the work that went into this project? Are outputs accessible to people with varying access needs using video and image descriptions? The guide provides examples of thoughtful attribution. In Everest Pipkin's book 'i've never picked a protected flower,' titles scraped from a forum were used. Pipkin credited hundreds of forum users in the back of the book, acknowledging the profound dependence on their work. They also open-sourced the code used and provided an opportunity for contributors to opt out. This exemplifies how to credit individual forum contributors and acknowledge the collective effort. Furthermore, the guide touches upon the technical aspects of securing your work, such as password protection or releasing to a closed community, especially if your model/code/dataset contains sensitive or confidential information. It also prompts consideration of long-term storage plans for your model, code, and dataset, and checking with people involved in your data set creation and model code to understand how your release decisions might affect them. The Partnership on AI's work on transparent and comprehensible AI documentation, including 'About ML' and 'Publication Norms,' is also referenced as a resource for better documentation practices.

As you ascend the mountain of AI art creation, the Partnership on AI's guide culminates in a summary of best practices, distilled from the checkpoints and considerations discussed. These practices are designed to guide artists toward ethical, transparent, and impactful work. Prioritize work that is in the public domain or directly ask for permission from those whose identity and/or work is represented in the dataset. This respects intellectual property and the individuals whose creations form the basis of AI models. When scraping work, always seek consent and ensure fair use. Save environmental training resources by utilizing transfer learning from a pre-trained model. This approach significantly reduces the carbon footprint associated with training AI models from scratch, promoting sustainability in digital art creation. Credit the work of others whenever possible. This extends to artists whose work is in your dataset, as well as people who have shared their code or models. Proper attribution acknowledges the collaborative nature of AI development and fosters a more equitable ecosystem. Tag and thank people when posting online. This simple act of recognition goes a long way in building community and showing respect for collaborators and sources. Document your work in detail to allow others to learn from and critique your process. This transparency is vital for the advancement of responsible AI practices and for fostering a deeper understanding of AI art creation. Make your work accessible, considering various access needs. The least risky approach to dataset creation is to make your own dataset through original media such as illustration, photography, text, and video. This grants you complete control over the content and ensures ethical sourcing from the outset. By adhering to these best practices, artists can navigate the complexities of AI art creation with integrity, contributing to a more responsible and innovative future for digital creativity.

You've successfully navigated the winding path of questioning and arrived at the summit of responsible AI art creation. The journey through the checkpoints – Datasets, Model Code, Training Resources, and Publishing & Attribution – has equipped you with a framework for thoughtful and ethical practice. The Partnership on AI's guide, 'Making AI Art Responsibly,' has provided not just information, but a call to action for artists to engage critically with the tools and processes they employ. As you move forward, remember that the field of AI art is dynamic. Continuous learning, reflection, and adaptation are key. The best practices outlined are not rigid rules but guiding principles to help you create art that is not only innovative but also respectful, transparent, and beneficial to society. Your journey through this guide has been about more than just technical execution; it has been about understanding the profound impact of your creative choices. By embracing responsibility, you contribute to a more ethical and sustainable future for AI in the arts. The creators of this guide, Emily Saltz, Lia Coleman, and Claire Leibowicz, encourage you to reflect on your experience and share your insights. Your feedback helps refine this guide and empowers future artists. So, go forth and create! Happy AI art-making!