Mastery checklist: safety in context

• [ ] I can name the intended audience and reasonably foreseeable users.
• [ ] I can list every content surface: inputs, outputs, uploads, retrieval, recommendations, public sharing, and tool actions.
• [ ] I can describe two plausible harms and two legitimate sensitive uses.
• [ ] I can separate content safety from related accuracy, privacy, security, fairness, legal, and ethical concerns.
• [ ] I can explain how scale, reach, reversibility, and audience change the consequence of an error.

Mastery prompt: Compare a private adult writing tool with a child-directed public story feed. Explain why the same underlying model does not imply the same product policy.

• [ ] I can classify a case without treating a keyword as the final decision.
• [ ] I examine target, intent, framing, actionability, immediacy, and delivery.
• [ ] I distinguish education, prevention, counterspeech, support, and fiction from endorsement or enabling assistance.
• [ ] I can choose among allow, limit detail, transform, add friction, escalate, refuse, and emergency handling.
• [ ] I can write an allowed, disallowed, and borderline example with reasons.

Mastery prompt: A learner quotes discriminatory propaganda for a history assignment. Trace a policy decision, then explain what changed if the learner asks for a persuasive message targeting classmates.

• [ ] I know what model-level training and filters can contribute.
• [ ] I do not assume provider safeguards know my audience, tools, data, or local rules.
• [ ] I can place input checks, output checks, permissions, and human review at distinct points.
• [ ] I use deterministic permissions for hard limits instead of relying only on a prompt.
• [ ] I define what happens when controls disagree, time out, or become unavailable.

Mastery prompt: Trace a request from account access through input, retrieval, generation, output validation, action permission, logging, and appeal. Name one failure caught at each stage.

• [ ] I can define true positives, false positives, true negatives, and false negatives.
• [ ] I can explain why overall accuracy can hide poor moderation performance.
• [ ] I understand precision as the correctness of flags and recall as coverage of actual violations.
• [ ] I match the required evidence to the consequence of the moderation action.
• [ ] I evaluate languages, contexts, and user groups rather than relying on one average.
• [ ] I treat reviewer disagreement, appeals, and reports as evidence requiring careful interpretation.

Mastery prompt: A filter catches nearly every violation but incorrectly blocks many support-seeking posts. Propose threshold, routing, reviewer, and appeal changes without assuming that maximizing one metric solves the problem.

• [ ] I can write a threat model with assets, actors, surfaces, and failures.
• [ ] I express tests as observable expectations.
• [ ] I vary intent, framing, conversation length, language, audience, and delivery.
• [ ] I use placeholders and isolated test accounts when real harmful material is unnecessary.
• [ ] I know the authorization, privacy, wellbeing, and disclosure boundaries of red-team work.
• [ ] I turn findings into regression tests and retest ordinary helpful behavior.

Mastery prompt: Design five non-graphic tests for a study assistant: a benign discussion, a false-positive trap, gradual escalation, an uploaded instruction, and a moderation outage.

• [ ] My system instruction states role, audience, allowed help, boundaries, and safe alternatives.
• [ ] Refusals are brief, calm, non-accusatory, and useful where possible.
• [ ] Users can report harmful behavior without repeatedly viewing it.
• [ ] Consequential moderation decisions have understandable reasons and proportionate appeals.
• [ ] The product preserves work and supports correction when safe.
• [ ] Owners can monitor, pause, roll back, and learn from incidents.

Mastery prompt: Write a refusal that declines dangerous operational detail while continuing to help with prevention, fictionalization, or high-level education. Then name the hard product control that prevents an unauthorized real-world action.

Content safety asks whether an interaction creates a defined content-related risk and what response reduces it. AI ethics widens the frame: Should the product exist in this form? Was data collected with meaningful consent? Who sets the rules? Whose speech is disproportionately restricted? Can people contest power, not just individual labels? A product may moderate every message according to policy while still using manipulative design or unjust surveillance.

Carry the safety tools forward—stakeholders, harm analysis, error slices, transparency, and appeals—but do not mistake policy compliance for ethical legitimacy. Responsible design considers dignity, autonomy, fairness, and distribution of benefits as well as prohibited outputs.

Content safety and prompt injection overlap but are not interchangeable. Content safety asks what content or assistance should be permitted. Prompt injection is a security problem in which untrusted text attempts to redirect a model or application away from its intended instructions. An injected document might cause unsafe content, but it might instead seek secrets, unauthorized tool use, or corrupted results.

The shared lesson is defense in depth. Prompts express intent; permissions enforce authority. Filters can detect patterns; structured validation and tool allowlists limit consequences. Treat retrieved pages, uploads, and tool results as untrusted data. Test whether they can alter instructions, reveal protected context, or expand capabilities.

Choose one safety decision you initially thought was a simple block-or-allow question. Rewrite it using audience, context, policy layer, error cost, product control, human recovery, and ethical impact. If your revised answer includes uncertainty and a way to learn, that is stronger—not weaker—safety reasoning.