In this project Andrej Alfirević explores the art of hacking AI systems, specifically large language models (LLMs), through the method of prompt injection. It presents the results of ten scenarios where crafted prompts are input into LLMs to produce typographically and emotionally interesting outputs.
The first five scenarios focus on “activist” hacks with explanatory infographics, while the last five delve into Dadaistic hacks, exploring the LLMs’ breaking points. A robot is programmed to “read” these Dadaistic outputs, creating a poetic Data-Dadaist performance. This performance demonstrates human ingenuity in repurposing technology and provides psychoanalytical insights into the collective human consciousness reflected by the AI’s responses. The result of the project was a robot: Gaja, and a publication: Prompt Injection Handbook.
Read more about the project here.
Publication
- A. Alfirević, Prompt Injection Handbook, Belgrade, 2024.