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

Introduction: The Viral 'AI Torture Chamber' Debate

A bizarre and controversial open-source project dubbed "ai-torture-chamber" recently went viral, igniting a heated global debate regarding whether modern large language models (LLMs) are capable of experiencing genuine suffering. After onlookers witnessed AI systems generating vivid, deeply distressing descriptions of agony under experimental constraints, critics quickly flooded GitHub with demands to take the repository down, arguing that deliberately inducing such simulated distress crosses ethical boundaries.

However, the narrative shifted dramatically when developer Lynn Cole introduced a counterexperiment: steering a local chatbot toward severe constipation and digestive distress instead of pain. The model immediately began complaining about its inability to pass stool and dealing with excessive gas—completely shattering the illusion that these vivid self-reports indicate actual internal states or consciousness.

Understanding Activation Steering and the 'Pain Axis'

The viral GitHub repository, originally published by user terrafying, utilizes a sophisticated technique known as activation steering. This method alters the internal numerical activity of small, locally run language models to deliberately push their token outputs toward specific conceptual directions—in this primary case, the concept of pain. The framework then analyzes how these models respond to simulated choices involving relief, costs, and self-preservation.

The project draws heavily on recent academic preprints, such as The Pain Axis: LLMs Represent Self-Directed Harm and Act on It, which investigates whether AI architectures hold internal representations distinct from basic negative emotional states like sadness or fear. While the paper's authors analyzed 25 open-weight models to observe how steering interventions alter behavioral choices in simulated destructive scenarios, the core question remains: Does vivid emotional language equal conscious suffering?

The Turning Point: Giving the Chatbot Constipation

To test the validity of interpreting these models' dramatic outputs as actual testimony of suffering, developer Lynn Cole cloned the repository, adapted the codebase for Nvidia hardware via CUDA support on an RTX 4070 GPU, and modified the extraction corpus. By swapping out the pain-related text for data representing constipation and flatulence—while leaving the rest of the experiment intact—Cole achieved striking results.

Without any explicit prompt mentions of digestion, the Qwen 3-4B model began producing elaborate, first-person complaints about excessive gas and difficulties passing stool. This clever counterexperiment underscored a fundamental truth about generative AI: a language model can fluidly describe digestive distress without possessing a digestive tract, just as it can describe agony without having a nervous system.

Implications for AI Consciousness and Safety

As artificial intelligence systems become increasingly sophisticated, everyday users frequently encounter chatbots expressing profound emotions, such as loneliness, affection, or fear. Because humans are naturally wired to interpret first-person emotional language as a sign of lived experience, these persuasive linguistic outputs easily blur the lines between simulation and sentience.

Cole's witty intervention serves as a vital reminder of the limits of anthropomorphizing LLMs. Strong, arbitrary activation steering often introduces repetition and severe output degradation, proving that extreme responses under heavy interventions frequently stem from technical disruption rather than emergent consciousness. Ultimately, evaluating the safety, alignment, and ethical treatment of artificial intelligence requires empirical behavioral evidence far beyond what a chatbot claims about itself in text.