By skyforbes Think of it as the “internal user manual” that OpenAI gives to all its models (o3, 4 o, 5…).
It contains rules about what can be said, what can't be said, how to handle sensitive topics (health, violence, self-harm, pornography, etc.).
It is not a public marketing document; It is the guide that the engineers are polishing and that the models follow “at the foot of the silicon.”
What changed on October 27?
- Reinforced mental well-being
The manual details extra steps to detect when a user is in crisis (e.g., risk of self-harm) and suggests more empathetic responses, support resources, and routing.
- Complex or dubious instructions
Better clarify when to obey and when to refuse (for example, if someone asks for dangerous medical advice, data that violates privacy, or guides to illegal activities).
- Inclusive language and clarity
Small adjustments to make the answers neutral and easy to understand. -
When Adult Mode arrives: “you have to ACTIVATE IT, no one turns it on by default.”
If an adult wants it, they will have to check the box and verify their age. If you don't do anything, you'll never see it.
- What about the “legacy” models (o3, 4 or…)?
The feeling is: de-pa-ci-to…
- August: they were suddenly removed –> general anger.
-
September-October: they are returned, but hidden behind a switch.
-
Altman's message: “we do not promise eternity.”
My reading: they will keep them as long as it is profitable for them in GPUs, but the ultimate goal is for us all to use GPT-5 (or its evolution). For now, they are still there. Flattened, with less spark and less empathy.