Microsoft’s Six-Principle RAI Standard

▶ Watch (00:28)

Microsoft’s responsible AI standard has six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Kim framed all six as equally important, with no priority ordering among them. The hardware context matters here. Copilot Plus PCs carry an NPU (neural processing unit) that handles on-device model inference, reducing latency and keeping data off the cloud. The Windows Copilot Runtime APIs, available in the Windows App SDK, sit on top of that NPU and give developers local AI capabilities including text generation via Phi Silica, background segmentation, and text extraction.

Phi Silica Content Moderation Controls

▶ Watch (04:08)

Phi Silica ships with content moderation built into the API. Four harm categories (hate, sexual, violent, and self-harm) each have separate sensitivity controls for input and output. “Medium” is the default; “low” is the stricter setting, which Kim called counterintuitive. In the Phi Silica harm-filter live demo, setting output violent-content to “low” returned “Response blocked by policy” on a “kill mobs in Minecraft” prompt. Setting input to “low” blocked the prompt before any generation: “Prompt blocked by policy.” A system prompt adds a third option, steering outputs toward safe framings without blocking.

“Phi Silica has a built-in content moderation layer” — Diana Kim

Non-Technical RAI: Transparency, Feedback, and User Control

▶ Watch (10:55)

Technical controls are one piece. Kim’s argument was that no filter is infallible, so the product layer matters equally. The Click to Do on-screen summarization demo made three design choices visible. First, AI output appears in a separate popup; the source text stays untouched. Second, a feedback button sits at the bottom of that popup for reporting problems. Third, a “Learn More” link opens documentation covering feature scope, privacy information, and model transparency notes. Each choice is deliberate, not decorative.

“But we are sure not to write over the user’s initial text.” — Diana Kim

Notable Quotes

Phi Silica has a built-in content moderation layer Diana Kim · ▶ 7:37

that no technical intervention is infallible. Diana Kim · ▶ 10:44

But we are sure not to write over the user’s initial text. Diana Kim · ▶ 12:33

Key Takeaways

  • Phi Silica filters four harm categories at two sensitivity levels, configurable separately for input and output.
  • System prompts work as a second mitigation layer when filter-level blocking is too blunt for context-specific apps.
  • Responsible AI requires transparency docs, feedback buttons, and user checkpoints alongside technical content filters.