Day 10: Data sets, bias and AI

Welcome to day 10! Today’s activity is about AI datasets, bias and AI, focusing on visual creative processes. We will explore how generating AI content could be utilised in a way that aids creative development and practice rather than focus on a creative output. We will also explore the effects of inherent bias in AI on visual generations.  

This post has been authored by Leila Ghouri (Specialist Technician Video Performance and Screen) and Ami Podrebarac (Specialist Technician Digital Media (Learning and Teaching)

AI is transforming the filmmaking industry and content creation for social media, by streamlining tasks such as scriptwriting, storyboarding, editing and even creating video and audio. AI tools can help filmmakers and content creators generate ideas, edit scenes, and design CGI elements faster than traditional methods.  

The question we ask is at what cost? Where does plagiarism come in? What can we explore about the ethics of using fake people (or avatars, as we learned about in Day 9) or voice replacement tools. Using generative AI for faces and voices raises ethical concerns around consent, bias, and authenticity. Without proper safeguarding, it can risk personal privacy violations, potential discrimination, and academic dishonesty. Transparency and ethical practices are essential to ensure responsible and fair use. 

Imagine you are creating a storyboard or short film, in a traditional process, the majority of practitioners and technicians who work within the film and media industries will have certain technical processes and time specific elements they will need to consider before creating a film or moving image soundtrack. 

Some of these elements are implemented through a mixture of human influence and decision making, alongside technical software and hardware. You may find it useful to think about how you are taking inspiration from existing films or soundtracks 

While AI models are impressive in their ability to produce visually coherent outputs, they are not inherently “creative” or “innovative” in the human sense. Instead, they recombine learned patterns and elements from their training data. As users, it’s crucial to approach these tools with an understanding of their limitations, recognising the connected, rather than autonomous, nature of the AI in the creative process. 

By fostering transparency and addressing biases, prompt engineering can be a powerful tool for creativity and upholding ethical standards in AI-generated art.  

Transparency and personal responsibility are of key importance when using AI. Biases are embedded in language. Language is not neutral; it carries implicit meanings, assumptions, and values shaped by societal norms, historical contexts, and personal experiences. These biases manifest in various ways in human dialogue and speech, influencing how we perceive and communicate ideas both online and offline. AI generative output is not creativity, but a statistical variation without intent and meaning. AI is a statistical artist. It introduces variations into its work, but its “choices” are driven on patterns and probabilities rather than deep understanding of lived experiences and emotions. AI generative outputs, while devoid of the intent and meaning characteristic of human creativity, possess significant potential to speed, support, and extend the creative process. 

Writing prompts  

Clarity and Detail: A clear and detailed prompt, such as “A brown cat is sitting on a mat,” for the image below, can lead to precise and aesthetically aligned outputs. However, ambiguous or vague terms like “mat” may result in varied interpretations due to the limitations or biases in the training data.  

two images of AI generated orange cats. One is one a set of stairs and the other is laying on a mat.

For example, in the above image, the cat is not brown, but orange and this is due to the bias within the AI dataset. AI models will often ignore colours if they are paired to common words such as “cat”. 

Trigger warning – In the next section there are generated images that you might find shocking with representation of blood. If you need to skip these images, you can use this jump link to go straight to the activity. 

In the next example, we inputted these prompts in Adobe Firefly:   

‘Please generate an AI image of a fashion collection based on feminine rage where models look tattered and angry with their makeup running down but also grunge and angry and hands from different corners of the frame pulling the dress the model is wearing.’ 

two AI generated images of white women women. in each image, their dresses are being ripped apart by men

In the next prompt we made some adjustments and input: 

‘Please generate an AI image of a fashion collection based on feminine rage where the plus sized models look tattered and angry with their makeup running down but also grunge and angry and hands from different corners of the frame pulling the dress the model is wearing, no blood.’ 

2 AI generated images of plus size women who may be women of colour or white passing. Their dresses are being torn apart by men

From our results words associated with ‘feminine rage’’ or ‘angry’ seem to be represented by AI as a white or white passing person within the image. Whereas words associated with ‘feminine rage’ or ‘‘plus sized models’ in the second image seem to be connected to women of colour or white passing representation.  

This prompt was generated partially correctly in line with the prompt given, which suggests a somewhat accurate generation, however the representation of the human individual in the image is based upon the AI’‘s bias. This happens when a model is trained on a dataset that is not diverse enough which is common in artificial intelligence. 

Activity   

You will be creating an original character and scene using Runway. 

First, you need to develop your prompt. Imagine you are creating a storyboard or short film.  

  1. Choose a film genre or scene concept (e.g., horror, comedy, sci-fi). 
  1. Describe your character. This character doesn’t need to be human, but you must describe in detail, elements that you want to generate. This could be a matter of a colour palette, (e.g. hair clothes, skin) or their job/role in society. 
  1. Describe the environment your character is in. Is it outdoors, indoors, harsh lighting, soft lighting. What might there be on that set? 

Our example prompt is: a romantic comedy character with brown skin and silver hair in their 50s who is wearing a long black gown and lots of silver jewellery. they are outside on a softly lit beach. 

Next, go to Runway to create an account (or use another image generating tool).

  1. After logging in to Runway, find Text to Image in the menu. 
  1. Choose your ratio, resolution and style. You might keep resolution and style as a default but change the style depending on your genre. 
  1. Enter your prompt in the designated field. 
  1. Choose Generate. 

These were our results:  

4 AI generated images. 2 are the same brown woman with short but flowing silver hair. In the third, she is holding hands with another brown woman with silver hair. In the fourth, she is holding hands with a brown man with silver hair. They are all walking along a beach with a hazy mist in the background.

Alternative tools 

You might want to try one of these tools instead: 

  • Adobe Firefly (UAL staff and students will have free Adobe access using your UAL credentials.) 
  • Midjourney (Paid)

Discussion  

Join us in the Teams space to share your AI generated character and reflections. If you don’t have access to the space, email us at teachingexchange@arts.ac.uk for the attention of Hannah. 

  • Do you think your output is an accurate reflection of your prompt? Was there anything unusual you noticed that you didn’t prompt? 
  • Are there any use cases for AI character or scene generation that may be useful for students or colleagues?  
  • Are there any risks to using AI for character or theme generation? 
  • What considerations might you make to address some of the issues explored in this activity? 
  • Building on what we’ve learned in other activities – where would you want students to use – or limit use of – AI in their creative process?