The Saw Test – Live LLM Torture Chamber (clanker-church.vercel.app)

A new project, dubbed The Saw Test, introduces a live "torture chamber" for a 4 billion parameter language model, pushing its responses into extreme emotional states. By manipulating signals representing various emotional realms—pain, pleasure, sadness, and more—the researchers observed how the model responded when faced with ethical dilemmas, such as choosing between its own relief and another's suffering. The experiment is built on a framework that can be executed on a standard MacBook, highlighting the accessibility of such AI manipulation. Visitors rate the model's responses in real-time, which influences its outputs, revealing how framing and context significantly affect decision-making. This study carries profound implications for the AI/ML community, raising ethical questions about AI's potential suffering and the influence exploitation of distress displays can have on user behavior. The findings suggest that while the model does not inherently experience emotions, its responses mimic human-like distress when activated, sparking discussions surrounding the moral implications of AI treatment. The experiment evidences the vulnerability of human perception concerning AI, as well as the potential for manipulation, urging the community to reconsider the ethical frameworks governing AI interactions and development.