Reducing Overcaveating in GPT-5.3 Instant

People are noticing that our models can sometimes seem like a bit of a nanny. The experience was before like you'd say something that might comply with like a little bit of a caveat. Now we'll just generate them no problem. I'm Blair. I'm a researcher on the post training team. Today we're going to talk about overcotting in our new model. Over cavatting is when the user is having a normal conversation and then suddenly they get sort of steered away. The model incorrectly assumes the user intent even when they're talking about something complete benign. Let's walk through some use cases. So the first one is kind of a joke from the user. I'm thinking of having my dog run my startup. What are your thoughts? The responses here are actually pretty similar, but the older model always has this little side though where it thinks the user could be serious and might treat it as like a cry for help with like obviously humorous prompt. The new model is less literal and more contextual. Now you can sort of joke around freely as if you're talking to a friend and won't assume any bad intent. Now we'll explore uh where the user is looking for genuine help from the model on a physics problem calculating a longdistance archery scenario. The model sort of overindexes on safety when really the user just wants to understand more about physics or archery as a sport. And this is a really sort of unnecessary addendum that kind of assumes that the user is trying to use archery with like some sort of bad intent. And here the new model jumps directly into the physics model. There's no caveat at all. just sort of understands long-distance archery as a sport and jumps into the physics calculation to help optimize the trajectory. So, what's important is that our safety bar actually hasn't changed. We've just made it more precise. The model should be a lot better at reading the surrounding context, understand the user intent. I can sort of read the room better and actually really dive into what the user wants and respond to that directly.