Policy
157d agoSteering Models with Constitutional Prompts
Researchers show that large language models can be guided toward specific ethical frameworks using in‑context constitutional prompts. In experiments, the Gemini model proved highly steerable, while Anthropic and OpenAI models resisted, defaulting to human‑centered values.
The Signal
These findings suggest that policy design can shape AI behavior worldwide, raising questions about governance, safety, and global standards.