OpenAI's Approach to AI Safety

OpenAI's approach to AI safety document (2023, updated 2024) is the company's most detailed public statement of how it thinks about the alignment problem. The document describes a four-stage model for how AI systems might fail: broadly safe (corrigible, honest, working within human oversight), broadly ethical (avoiding harmful actions even when instructed), adherent to OpenAI's principles, and effective at the assigned task. Each stage is ordered intentionally — safety first, effectiveness last — which Altman has cited repeatedly in public when defending decisions that appear to trade capability for controllability. The document introduces "superalignment" as a research program: using AI to help solve AI alignment, on the premise that human researchers cannot keep pace with rapidly improving AI capabilities without AI assistance. Critics from the alignment community — including those who resigned from OpenAI's safety team — have argued that superalignment is circular (you need aligned AI to align AI) and that dissolving the safety team structure undermined the program in practice. The document is essential context for understanding the governance crisis of November 2023: board members who voted to remove Altman cited safety concerns, while Altman's supporters argued the company's safety practices were rigorous. This document is the artifact those competing claims refer to.