A Harm Analysis Framework for Code-Synthesizing Large Language Models

OpenAI researchers propose a harm analysis framework to reveal the technical, social, political and economic safety risks that code-synthesizing large language models such as Codex may pose after deployment. The paper points out that models capable of generating code at scale have clear limitations, alignment problems and potential for misuse. The framework is accompanied by a new evaluation method that measures the ability of advanced code generation technology to handle prompts of varying complexity and expressiveness specifications, and the gap between its level of understanding and executing these prompts and human capability.