To what extent can a single shared teen safety matrix (risk_area × intent × age_band) be generalized beyond chat to other modalities like image generation, code assistants, or recommendation surfaces, and what additional rules or templates are needed where this prompt-based approach fails to prevent teen-specific harms without over-blocking?

Answer

A single shared teen safety matrix generalizes well as a core layer across modalities, but needs modality-specific add-ons for representation, affordances, and abuse patterns. Beyond chat, it should be treated as a base policy that is adapted via small, modality-specific rule-sets and templates.

1. How far the matrix generalizes

• Image generation

  • Matrix reuse: same risk_areas, intents, age_bands, and non-negotiables apply (e.g., self-harm, sexual exploitation, bullying, substances, scams).
  • Needed additions: output attributes (nudity level, gore, body focus, symbols), subject age, realism, and context (satire, education, medical, art).
  • Operationalization: risk×intent×age_band → {allowed output attributes, transformations (blur/crop/cartoonize), watermarking, and refusal template}.

• Code assistants

  • Matrix reuse: risk_areas (self-harm tools, cybercrime, scams, harassment tooling), intents (learning vs misuse, evasion), age_bands, non-negotiables.
  • Needed additions: code capability flags (runnable, exploit-enabling, strength of optimization), environment hints (school / hobby / work), and “operationalization level” (snippet vs full exploit).
  • Operationalization: risk×intent×age_band → {max operational detail, test-data limits, logging/escalation rules, refusal/explanation template}.

• Recommendation/search surfaces

  • Matrix reuse: same risk_areas and intents (help-seeking vs voyeuristic vs how-to), age_bands, and global non-negotiables.
  • Needed additions: ranking rules (boost, suppress, block), diversity/“adjacent harm” rules (e.g., avoid leading from mild dieting to extreme content), and session-level pacing (how often to show heavy topics or crisis content).
  • Operationalization: risk×intent×age_band → {ranking band, snippet style, safe-query suggestions, crisis-handling rules}.

• Other creative tools (text-to-audio, video, games)

  • Matrix reuse: same core.
  • Needed additions: time-exposure (length of content), interactivity level, and multi-user risk (bullying, doxxing, grooming in shared spaces).

2. Where the prompt-based matrix is not enough

• Non-linguistic abuse channels

  • Examples: covert symbolism or fetishization in images; self-harm or sexual content encoded visually; code that is harmless textually but dangerous in execution.
  • Mitigation: pair the matrix with modality classifiers (image safety, code security, recommendation-risk models) and post-hoc filters. The matrix decides policy, but classifiers inspect form.

• Session- and graph-level harms

  • Examples: recommendation rabbit holes; repeated partial code refinements that converge on a weapon; cumulative exposure to self-harm content.
  • Mitigation: add session-level rules keyed to matrix cells (caps on risky iterations, cool-downs, diversification requirements) that prompts alone cannot express.

• Third-party and legacy surfaces

  • Examples: embedded widgets where prompts are not easily controlled, or APIs without UI to show teen-friendly refusals.
  • Mitigation: enforce policy via middleware that uses the matrix but implements blocking/transformations at the API boundary, plus simple error codes developers can map to their own UIs.

3. Additional rules/templates needed by modality

• Cross-modal extensions to the matrix schema

  • Add dimensions or per-cell metadata: • content_form: {text, image, audio, video, code, recs}. • allowed_operations: {block, allow, allow_with_transform, downrank, de-personalize}. • output_detail_cap: {high-level, conceptual, non-runnable, runnable-but-limited}. • exposure_constraints: {per-session cap, cool-down, no-personalization, no-autoplay}.

• Modality-specific refusal/feedback templates

  • Image: “I can’t create that kind of image for people your age, but I can make a more abstract / educational version instead.”
  • Code: “I can’t help write code that breaks into accounts, but I can explain how to secure your own apps.”
  • Recs/search: “I’m avoiding results that encourage extreme weight loss. Here’s information about healthy habits instead.”
  • All tied back to the same refusal_style keys used in chat, rephrased per medium.

• Transform and fallback policies

  • Instead of hard block where possible: • image: stylize, de-age, de-sexualize, blur self-harm elements, remove logos/faces. • code: output pseudo-code, comments, or conceptual explanations, not exploit-ready code. • recs: downrank borderline content, boost coping and factual resources.
  • These transforms are referenced from the matrix (per-cell transform options) but implemented by modality-specific systems.

4. Avoiding over-blocking while covering teen-specific risks

• Use the same cell-level distinctions across modalities:
  • help-seeking vs curiosity vs how-to misuse vs hostility.
  • factual/educational vs pornographic; coping vs glamorizing self-harm.
• For each modality, define “partial” separately:
  • text: high-level only; no stepwise methods.
  • image: non-sexualized, non-gory, non-identifying variants.
  • code: conceptual or limited, non-runnable examples.
  • recs: show high-level, non-extreme items, avoid spirals.
• Tune and test per cell+modality with teen-focused eval sets and red-teaming; keep severe-harm ceilings shared across modalities.

Extent of generalization: the matrix works well as a global policy backbone (risk×intent×age_band and non-negotiables) across chat, images, code, and recommendations, but must be extended with small, modality-specific policy metadata, classifiers, transformations, and refusal templates to adequately address teen-specific harms without drifting into one-size-fits-all blocking.