Q: 9
What should you implement to prevent hateful responses from being returned by a generative Al
solution?
Options
Discussion
Option B. since filtering actively stops the bad output from even reaching the user. If it just asked to lower the chance (not prevent fully), C might be valid but "prevent" flips it for me. Open to other takes though if I missed something specific in Azure AI docs.
I don’t think B is right this time, I’d pick A. Abuse monitoring can flag hateful content as it comes up and block it from being returned. Content filtering is useful too, but abuse monitoring seems more direct here. Anyone see reports where A covers this better?
B here. Content filtering actually blocks the negative stuff from getting through, so it directly prevents hateful responses rather than just lowering the odds. I think that's what they're looking for with "prevent".
I’d say A. Abuse monitoring feels like the right thing because you're watching for bad stuff in real-time and can act fast if something hateful comes up. Not 100 percent sure though, maybe someone else sees it differently.
Probably B here, since content filtering in Azure AI is what actually blocks hateful responses from ever making it out. Fine-tuning and prompts help lower the risk but only B consistently prevents. Let me know if you see it differently.
Content filtering is the core control to block hateful outputs from generative AI, so B makes sense here. The other options might help reduce risk but don’t actually prevent those responses. Pretty sure B is correct for this one.
D imo. Prompt engineering actually lets you shape the kind of outputs you get, so if you write your prompts carefully to avoid hateful context, the system won't generate those responses in the first place. I get that filtering is more direct, but prompt tweaks can sometimes totally prevent bad content if you control input tightly. Not 100% sure though since Azure has some built-in stuff.
B is the way to go since content filtering actually stops hateful responses from ever being returned. Fine-tuning (C) lowers the chance, but doesn't guarantee prevention. Pretty sure that's what Microsoft wants here-disagree if you see it differently.
B makes more sense because content filtering actually blocks hateful stuff before it reaches users. Fine-tuning (C) just lowers the risk, but doesn't guarantee prevention. Pretty sure B fits best given "prevent" is in the question, but open if anyone sees it different.
C fine-tuning. I think adjusting the model helps reduce those bad outputs directly, so that's what I'd pick here.
Be respectful. No spam.
