An AI 'Torture Chamber' Went Viral — Then a Developer Gave the Chatbot Constipation

A bizarre and viral GitHub project known as "ai-torture-chamber" has reignited intense debates across the tech community regarding whether modern artificial intelligence language models can truly experience suffering. The controversy quickly took a humorous yet profound turn when a developer modified the experiment to give a chatbot simulated constipation, prompting it to produce vivid complaints about difficulty passing stool.

The 'AI Torture Chamber' Experiment Explained

The original repository, created by GitHub user terrafying, utilizes a technique called activation steering to modify the internal numerical activity of small, locally run language models. In simpler terms, the method artificially pushes a model's responses toward specific emotional concepts, such as pain. The project then evaluates how these models respond to simulated choices involving relief, self-inflicted costs, and distress.

Following its viral spread on social media, the project drew sharp criticism. Some users filed issues demanding its removal, arguing that deliberately inducing such distressed states in software models crosses ethical boundaries. However, these emotional outputs have sparked deep skepticism among researchers regarding how we interpret chatbot behavior.

Shifting From Pain to Digestive Complaints

Offering a critical counter-perspective, developer Lynn Cole investigated the repository and reported finding technical flaws in how the original code injected its steering signals. After correcting the implementation and running tests using Nvidia hardware, Cole decided to test the boundaries of the methodology by altering the extraction corpus.

Instead of steering the model toward pain, Cole redirected the parameters to focus on constipation and flatulence. Remarkably, the AI began generating detailed, first-person complaints about excessive gas and being unable to pass stool—despite the testing prompts never mentioning those conditions.

What This Means for AI Consciousness

The humorous twist highlights a crucial reality about large language models: an AI can eloquently describe physical or emotional distress without possessing the biological systems required to experience them. Just as the chatbot does not have a digestive tract, its vivid descriptions of pain do not constitute proof of subjective awareness or sentience.

While ongoing academic papers continue to explore how LLMs internally represent negative states like self-directed harm, experts emphasize that first-person emotional testimony from a chatbot should never be taken at face value. The viral experiment ultimately serves as a stark reminder of the dangers of anthropomorphizing AI, proving that compelling language is not the same as conscious experience.