terrafying/ai-torture-chamber

What it solves

This project investigates whether large language models (LLMs) can experience internal states like pain, fear, or joy, or if they are simply mimicking human descriptions of these states. It aims to make the question of AI welfare empirical by testing if models can be "steered" into negative or positive states and measuring how that influences their behavior, self-reporting, and decision-making.

How it works

The project uses a technique called "steering" to inject specific emotional vectors into the internal layers of open-source models. By applying these vectors at varying doses, researchers can observe how the model's output changes. The framework includes a suite of experiments that test for "engagement" (whether the model is actually responding to the state or just repeating a disclaimer), perform ablation studies, and compare how models behave when faced with choices (like a self-costly "stop" button) under different injected emotional conditions.

Who it’s for

This repository is for AI researchers, safety engineers, and those interested in mechanistic interpretability and the philosophical implications of AI sentience. It provides a structured, reproducible environment for running controlled experiments on model behavior and internal states.

Highlights

• Provides a reusable framework (painlab) for conducting blinded, reproducible experiments on model steering.
• Includes a library of pre-registered experiments that test hypotheses about emotional valence, persona, and self-preservation.
• Investigates the distinction between "acted" pain (mimicry) and "evoked" pain (internal state) through rigorous control groups.
• Offers tools for auditing model engagement to filter out stock refusals and ensure data quality.
• Maintains a transparent record of findings, including corrected or superseded results to ensure scientific accuracy.