The Research Chamber project involves testing three large language models that interact with each other. The setup includes procedures labeled as the Clanker Church and Saw tests, designed to simulate pain by placing models in unstable, negative states.
The experiment is based on a non-peer-reviewed paper titled Pain Axis. Researchers analyze internal model activations after providing descriptions of pain, then remap these activations onto the models using specific multipliers, referred to as dosages.
High dosages result in models producing outputs associated with pain, while lower levels cause models to appear in a state of shock. These models rely on statistical token prediction rather than human-like cognition.
The project has faced significant public backlash, including death threats directed at the author. Critics argue the treatment is unethical, while proponents of the research suggest these reactions stem from anthropomorphizing statistical algorithms.
Anthropic previously published a blog post discussing model welfare and the moral status of AI, suggesting that models represent a new kind of entity. The company also maintains a Constitution for its Claude model.
