2 Weeks, March 2023
Coded-This-Way: Unpacking AI Image Generation Tools
Probing into AI image generation algorithms to understand the issues and biases they are coded with and trained to produce . Unearthing some of the root causes of such issues to promote ethical, informed, and responsible use of AI tools.
Guidance: Prof. Manuel Beltrán
Project Overview
Despite the lack of access to the source code, Coded-This-Way, a proposed website, provides a lens to understand how Generative AI operates.
Coded-This-Way could be used to make one aware of the basics of how Generative AI tools function. It attempts to communicate values and biases embedded in AI tools. It also allows us to reflect on how much agency we could give an AI tool. Coded this way also aims to throw light on the risks of blindly believing AI to be unbiased, creative, or intelligent.
Problem Statement
Why do we need to understand Generative AI tools?
Artists, designers & thinkers use generative AI tools. Often we use image generation tools without understanding where the images are sources from and how it impacts our thinking or perception. Restricted access to source code, inner functioning's, lack of access to training data sets increases this issue significantly.
Objective
-
Unpack functionalities of Online Image Generation AI Tools.
-
Device a method to probe into its inner workings.
-
Design a method & interactive product that communicates essential findings.
-
Design a interactive product that allows users to learn through experience.
AI Algorithms, Ethics & Law, Data Labour, Data Annotation
Process inquiry to prototype
01 Secondary Research
Secondary research image generation software, AI tools, Data law and labour
02 Inquiry
Establishing objective
03 Primary Research
Analysing existing AI image generation tools by providing prompts
04 Analysing Image Generation Tools
Analysing visual data gathered through systematic prompting and AI image generation
05 Analysing Image Annotations
Juxtaposing data annotation with AI image generation tools with respect to hidden data labor
06 Ideation
Designing methods to communicate findings to audience
07 Ideas to Prototypes
Wireframing an interactive website
Secondary Research findings
Data labour and data annotation
For an AI tool, visualization of images is made possible through data annotation.
Large amount of manual effort goes into annotation. Data annotation enables image identification based on the provided training data sets gathered from online sources.
Lens for probing into AI Image generation
-
Training data set
-
Reverse engineering
-
Ethical issues
-
Generated images are based on images provided data
-
Beyond the legality of the issue of ethics, what are the other issues in using AI tools?
Mapping training data sets of Gen AI
Primary Research & Analysis
Exploring the inner workings of a Generative AI tool by using various prompts and combinations to better comprehend how artificial intelligence assembles images. Delving into the tedious process of data labour and image acquisition essential for the functioning of AI.
Prototyping Ideas to prototype
An interactive website wire-frame was made through iterative prototyping. It provides demo of actions and interactions it would perform: simulation to show fundamental functioning of AI tools; Annotation & Image generation from provided data set; Information about the site
