OpenAI has published an updated Model Spec, the internal document that governs how its models reason about values, priorities, and behavior. The revision centers on five core properties ranked in explicit order: being broadly safe, being broadly ethical, adhering to OpenAI's principles, being genuinely helpful, and avoiding actions that concentrate power inappropriately. That hierarchy is not cosmetic. When properties conflict, models are instructed to resolve tension by following that sequence, with safety overriding ethics, and ethics overriding helpfulness.
The document is worth reading in full because the reasoning behind each tradeoff is exposed, not just the conclusion. OpenAI argues that current models may have miscalibrated values without knowing it, which is the explicit justification for keeping humans in the loop even when a model's judgment appears sound. The spec also draws a hard line against any action, including ones requested by OpenAI itself, that would give any single entity, including OpenAI, disproportionate control over critical systems or decision-making processes.
The update arrives as OpenAI ships more agentic products capable of taking real-world actions across longer time horizons. The spec addresses this directly, instructing models to prefer cautious, reversible actions and to pause for human verification when uncertainty is high. What happens when those guardrails meet commercial pressure and user demands for autonomy is the question the document raises but cannot fully answer.
[READ ORIGINAL →]