Should AI safety focus on intent or block entire topics?

I’m building BlackChat around a narrower safety boundary. Difficult research questions stay answerable, while requests intended to enable real-world crime stop there.

For people who use AI for security research, crime analysis, politics, or mature subjects: which approach feels more useful and trustworthy—broad topic blocks or intent-based intervention?

I’d especially value examples of false positives or edge cases you have run into. BlackChat launches later today.