Our Approach to Generative AI in Better Images of AI 

In practice, no generative AI image generators (that we are aware of) meet all three criteria. Therefore, images generated by models, including, but not limited to, DALLE-E, Midjourney, Stable Diffusion and Adobe Firefly are not accepted in our library. This may change, and if any AI image generators do comply with our criteria, we’d love to know. Techniques that use AI, which leave the original image intact and are not image generators, such as those found in digital editing platforms (e.g., background remove/eraser or basic filters), are permitted if they only ‘enhance’ the original image instead of transforming it (see our Submission Handbook for further explanation of how we’re thinking about the boundaries here). 

One reason you might consider using generative AI to create images is if you feel you do not have the artistic skill or ability to create an aesthetic, compelling visual. 

Generative AI visual outputs often reinforce biases reflected in our inequitable society. Using digital heritage collections offers a more reflective approach to challenge existing narratives and can prompt us to explore and understand how they are relevant to current AI developments and practices.

Furthermore, archival images can expose past histories from marginalised communities, which can allow us to understand how systems of oppression are embedded in the context of AI. However, it can also introduce alternative understandings and approaches to AI – showing us how the current practices are not universal, or inevitable, and alternatives have existed in the past or in different geographies. 

The Playbook also includes a list of public domain resources (an additional list is in our Submission Handbook too), which enable creators to explore how images outside of copyright protection can be used to create visuals for submission into our library. Although an ongoing legal question in courts around the world, generative AI text-to-image models can infringe copyright through the use of the input data to train models alongside the resulting output images.

There are numerous ethical and moral objections to using generative AI models for these reasons too, including the use of creators’ work without permission, consent or remuneration to train lucrative, proprietary models. Using public domain images, such as those found within digital heritage collections, can therefore offer a more legally compliant and ethical approach. 

While we want to hold space for all creators who wish to improve the visual representations of AI, we believe we can only enable the use of AI in this process if it has been developed in ways that respect creative communities (this requires consented training data, disclosure, and remuneration). We are also deeply concerned about the environmental impact of generative AI supply chains and therefore endorse human-created artworks which do not have the same detrimental impact on our ecosystems. 

In terms of content for our blog, we’re opposed to the use of AI for writing, ideation, research, grammar for many of the same reasons as we are for generating images, such as mass non-consensual use of authors’ work, biases and plagiarism, environmental factors, exploitation, as well as cognitive decline and integrity. We understand that common digital tools might have embedded AI-based features that are triggered without consent. We discourage collaborators from using these tools, for example, by disabling the AI overview on Google by adding ‘-ai’ to the end of your searches, or just by using alternative web browsers. 

We understand that some people may use AI in the process of writing for other important reasons, such as for translation purposes. We do not wish to exclude these individuals from contributing to our blog and sharing their work with our community. 

Non-generative AI alternative approach: However, instead of relying on generative AI, our community of volunteers are a nifty bunch, and we often find ways in which we can support you. For example, this blog post features an interview with an artist in her native language (Spanish) with one of our multilingual volunteers who translated the post into English. Additionally, Zoya, who manages the day-to-day running of the library, is always happy to help write up ideas, edit content, or connect you with a volunteer or two who might be able to support you (and they can also benefit from the exchange too by learning from you!).

We’re not incentivised or structured to benefit from fast, predictable, polished outputs. Instead, we value (often slower) options which can forge new human relations and offer mutual learning opportunities (even if we sometimes miss out a semi-colon, whoops!). If you’d like to use generative AI in some way to write a post for our blog, we’d love it if you can get in touch first and explain why you’re planning to use generative AI so we can learn more, understand, and see if we can help you.

Image descriptions: Since our creators do not use generative AI to create images for our library, we also believe that we should take the same consideration, time, and respect to write descriptions and alt text for the images. The process of doing so can enable us to explore the plurality of stories told by each image, while also learning and engaging with the creators’ intentions and understanding/experiences of AI.

Where we are uncertain about creative choices made or parts of the image, we can engage with the artist to learn about the image process and creation. This has the benefit of connecting our artist community with our volunteers, allowing for human connection and learning opportunities. 

We have found that AI-generated image descriptions do not effectively communicate the artist’s process, and so cannot prompt us to critically think about what is represented in an image. Writing image descriptions is also a learning opportunity for volunteers, particularly new ones, who want to learn more about how AI can be visually communicated in more accurate and representative ways. Contextual elements of images, such as those informed by an artist’s own experience or understanding of AI, can be replaced or lost by using AI-generated descriptions.

In addition, material processes involved in creating the images can be lost. We encourage creators to detail any specific choices, materials, and processes taken to visualise AI, showing how all representations of AI come from somewhere, influenced by worldviews, contexts, and geographies. Our community blog also develops these descriptions to explore in more depth the human stories, challenges, intentions, research, and organisations behind images (see here, here, and here, for example). 

Alt text: For alt text, we also believe that preserving human-generated text increases the accessibility of our library for people who are blind or have low vision. AI-generated alt text descriptions of images may not communicate the specific choices an artist made or parts of an image necessary to understand the representation of AI portrayed in a certain image. This is particularly relevant for images which incorporate visual metaphors, whereby a literal description may not offer the same visual experience that other users are exposed to when using our library.

Being aware of how the image is presented and the context in which it sits can give people who are blind or have low vision a better experience using our library, which supports our purpose to increase public understanding of AI. Furthermore, we are cautious of uncritical uses of AI being used to ‘increase’ accessibility in a guise to minimise costs needed to invest in changing the ways the world is designed, which exclude individuals (e.g., people with disabilities), which cannot be simply overcome by technological approaches. 

Finally, as with writing descriptions, the process of writing alt text can be helpful for our volunteers to learn more about the visual communication of AI. Being able to describe images about AI often involves us undertaking research to understand the contents of the images and describe them in a way that communicates the same information from the description of an image as someone who relies on the visual experience.

While we do not currently have a resource that can help and guide creators or volunteers to write alt-text descriptions, we hope to be able to develop one in the near future, which is specific to our library and visualising AI.