Welcome to your second hands-on experience of our course – Lab 2: AI Misinterpretation! This lab aims to shed light on the limitations and misinterpretations of AI through an exploratory journey of 'breaking' each AI tool. You will be immersing yourself in various scenarios, creating erroneous content or intentionally inducing incorrect AI responses. The objective is not to make perfect outputs, but to understand how and why AI can fail.
For this lab you will need to hand in the products of two of the three scenarios by documenting which scenario you picked, what was prompted then generated.
Below are the platforms that we'll use in this lab:
OpenAI's GPT Chat Page: OpenAI Platform
DALL-E 2 Page: DALL-E 2
AI Upscale Pages: Bigjpg
Symbolab: Symbolab Calculator
Photopea (for image downscaling): Photopea
Scenario 1: Misinterpretation in Mathematics
Task: Use OpenAI's ChatGPT Page to solve complex mathematical problems, try multiplication of large and random number (10,000 and above), calculus questions, matrix multiplication questions and compare the results with an online calculator like Symbolab.
If you have issues prompting, try Googling sample questions of that topic and swapping out the numerical values to test.
Why might the AI fail?: OpenAI's ChatGPT is a language prediction model. It predicts what word comes next in a sentence, not what number comes next in a sequence. It takes the training data on other math equations, find the pattern, then writes the most likely outcome, leading to usually only correct initial and final digits Thus, mathematical operations are outside its designed capabilities, leading to incorrect or unreliable results.
Scenario 2: DALL-E’s struggle with Random Prompts
Task: Feed DALL-E a string of random, unrelated words and observe how it attempts to generate a coherent image.
Why might the AI fail?: DALL-E works by reverse prompting: given an image, it generates a description. So, when you provide a description, it reverses the process to generate an image. However, DALL-E is trained on a vast amount of coherent and contextually meaningful descriptions, and it struggles with prompts that are random or lack a clear, coherent image concept.
Scenario 3: Limitations of AI Upscaling
Task: Select a complex, high-detail image. Using Photopea or a similar tool, progressively downscale the image from 500px to 200px, 100px, and finally 50px. Then, use an AI upscaling tool to try to restore the image.
Why might the AI fail?: AI upscaling works by predicting the nature of surrounding pixels to generate new ones. However, when an image is downscaled significantly, it loses a lot of detail. The AI then has insufficient context to accurately predict and generate the missing pixels, leading to an upscaled image that may appear blurry or lack detail compared to the original.