Beyond task completion: OpenAI on principles and observability

In An Alien Mind, dated September 6, 2026, OpenAI chief scientist Jakub Pachocki distinguishes goal alignment from value alignment. The former means understanding and pursuing assigned goals; the latter concerns retaining principles in ambiguous, conflicting or unfamiliar settings.

Completing an assigned task does not establish that a system retains the intended principles in unfamiliar situations. The target and crossroads scenes illustrate this distinction.

The essay describes chain-of-thought monitoring at OpenAI and reports weakening reliability. Even when a model makes an appropriate judgment, whether people can reliably observe that process remains a separate question.

The author proposes better monitoring, continued human involvement and slowing development when necessary. These are research arguments and proposals, distinct from capabilities already demonstrated by a system.