Use AI responsibly with Microsoft Foundry Content Safety: Review

I was evaluating content moderation options for a small app and thought I'd check out what Microsoft offered with Foundry Content Safety. The module introduces the four main harm categories—hate, sexual, violence, and self-harm—and the concept of severity scores. This is useful vocabulary to have. The section on 'When To Use Content Safety' provides a couple of basic architectural diagrams showing where you might place the filter: before the LLM, after the LLM, or both. That was probably the most useful five minutes of the entire thing.

The rest of it felt like a slightly interactive product marketing page. The so-called 'exercise' was a disappointment. It was just a block of text describing a scenario, followed by a multiple-choice question on what to do. No API calls, no sandbox, no code. You just read and click. There's zero depth here. How do you actually implement this? What's the latency? What does it cost at scale? How does it handle adversarial prompts or nuanced cases? The module doesn't touch any of that. It's all high-level concepts and diagrams with arrows. I left with a vague understanding of the service's marketing claims and a link to the official docs, which is where I should have started in the first place.