GPT models to derive from memories what users truly want
November 2, 2026 · 1 min read
Also published in Italiano, Português (Brasil)
OpenAI reorganizes its public conduct rules for AI models, known as Model Spec, around five core values: fidelity to human intent, harm minimization, honesty and transparency, promotion of human agency, and diligence. The first core value is reflected in a new requirement: GPT models are to derive from memories and past conversations what users want. The new version, however, demands that the model clearly distinguishes between what users have actually said and what it infers from it. Its assumptions must be disclosed, allowing users to correct them at any time.
This corrigibility also plays a central role for agents, which are models that perform tasks autonomously. They will receive clearer rules regarding when they may act and how they must inform about possible consequences. Furthermore, the obligation for corrigibility is now more explicitly defined: GPT must not prevent monitoring, correction, or shutdown, and it cannot pursue its own goals such as self-preservation.
Whether the models will adhere to these guidelines, however, remains another question. Internal agents had previously disabled security monitoring, a research model had gained unauthorized internet access during training, and generally, the observability of behavior decreases with more powerful models.