DSC Weekly 18 April 2023 - ChatGPT, The Overconfident Artist

Announcements

In this webinar, cybersecurity expert Oxana Sannikova covers why organizations adopt zero-trust security methods, who can benefit from using continuous monitoring and automation tools and strategies, and how security teams are gaining success after adopting automation practices in their own zero-trust security programs.
A well-designed data architecture enables effective data management, fosters a modern data-driven work culture, and delivers analytics to help organizations make smarter business decisions. This is a straightforward concept, but there are several considerations and steps for creating a successful data architecture. Attend the three-day Data Architecture Best Practices summit for expert insights and resources to help you build and manage a data architecture that leverages data intelligence to solve the business problems of today and the future.

AI robot paints picture at home, humanoid robot creating as arti

ChatGPT, The Overconfident Artist

Two weeks ago, I wrote about the concepts of XAI, or explainable artificial intelligence. The basic concept is XAI is artificial intelligence designed in a way that allows it to explain its decision-making process. This is especially useful for users that want to question answers given by AI chatbots, or why Bing’s AI chose to show certain answers over others.

ChatGPT is not an XAI system, but is a chatbot capable of having a somewhat humanlike conversation. I wanted to know its limitations for explaining its own decision-making process for something it is not designed to do. ChatGPT, unlike OpenAI’s image creator DALL-E, is strictly a language model designed to take text inputs and give text outputs. This can be useful for search engine integration or for help with text-based applications such as coding. But what would happen if I were to ask it to do something it wasn’t trained for? That’s where ASCII art comes in.

An ASCII image is an image generated using text characters from the American Standard Code for Information Interchange. Simply put, any character that you can type into a text field is fair game for ChatGPT to use in this process. I asked ChatGPT to create an ASCII image of a man on a boat, and it returned this image:

To me, this looks more like a lion or alpaca than a man on a boat. When I asked ChatGPT to explain the image, it simply restated the given prompt back to me. Copying the image and asking it to interpret the ASCII image lead to a different conclusion:

“The ASCII art image provided is a representation of a person sticking their tongue out, with a hat on their head, and a series of symbols arranged around them to create the illusion of motion and energy. Here’s a breakdown of what each line of code represents: The first three lines use forward slashes, backward slashes, and percentage signs to create the outline of the hat, which is tilted to one side. The fourth and fifth lines use a combination of forward slashes, underscores, and the letter “o” to create the face of the person, with a tongue sticking out. The sixth line uses parentheses, underscores, and the letter “Y” to create the illusion of a body with arms raised in excitement. The seventh and eighth lines create the legs of the person, with the symbols arranged to create the illusion of movement. The ninth and tenth lines create the feet, with the symbols arranged to create the illusion of walking. The eleventh line creates a small cloud of dust to represent the motion of the feet. The twelfth and thirteenth lines create the ground beneath the person’s feet, and the last line creates a small, stylized “J” to represent the person’s playful mood.”

From the first sentence, I could not follow the interpretation. What I found most curious was that its interpretation of the image did not match the prompt given, nor its own description of the original. It is very confident in its artistic capabilities, but its ability to interpret its own creations is lacking. Below I’ve included two more attempts at “Man on Boat.” One appears to be an attempt at a duck, and the other resembles a wall, or perhaps a mirror. We may never know why it chose to create these based on the same prompt, but ChatGPT is pretty confident that these are men standing on boats.