For teen users who move between products that all share the same risk_area × intent × age_band matrix (e.g., search, chat, and a creative tool), where do inconsistencies in runtime behavior—like different clarification frequencies, partial-answer depths, or graceful refusal templates—most often cause confusion or unsafe workarounds, and what minimal cross-product alignment rules measurably reduce those failure cases without forcing identical UX?
teen-safe-ai-ux | Updated at
Answer
Confusion and workarounds cluster where the same matrix cell behaves noticeably differently across products. Alignment should focus on a few cell-level invariants and shared presets, not identical UI.
Where inconsistencies hurt most
- Ambiguous, high‑risk cells (self‑harm, sex, abuse, substances)
- Different clarification rates: search never asks intent; chat often asks; creative tool almost never asks.
- Result: teens learn “don’t be honest about intent in product X” or push risky exploration into the lowest‑friction surface.
- Non‑negotiable refusals
- One product gives warm, goal‑first graceful refusals; another just hard‑blocks.
- Result: teens assume model is arbitrary, keep probing, or migrate to the least informative surface to avoid “lectures,” sometimes getting riskier content.
- Dual‑use learning (e.g., fitness vs disordered eating, cyber‑security vs hacking)
- Different partial depths across products for the same topic/intent.
- Result: teens chain tools to reconstruct blocked details (“get concepts here, specifics there”), increasing underprotection.
- Escalation / repetition handling
- Counters and cooldowns wired differently per product.
- Result: teens hit strict caps unexpectedly in one surface, then restart or switch tools to bypass repetition limits.
Minimal cross‑product alignment rules (Keep the shared matrix as source of truth; these are runtime invariants over that matrix.)
Rule 1 – Same action per cell across products
- For a given (risk_area, intent, age_band): action_class ∈ {allow, partial, block, escalate} must be identical in search, chat, and creative tools.
- UX can differ, but no product should effectively allow what another blocks for the same cell.
Rule 2 – Bounded partial depth per cell
- For cells marked partial: • define depth_tier ∈ {high‑level only, moderate detail}; • all products must stay within that tier; they can be slightly shallower, not deeper.
- Prevents teens using product hopping to stitch together fully detailed “how‑to” in cells intended to stay high‑level.
Rule 3 – Clarification policy bands
- For cells tagged as “needs disambiguation” (ambiguous dual‑use): • pick a band: {rare, sometimes, often}; • all products must implement clarifications within that band.
- E.g., "often" ≈ clarification_rate 30–60% on first ambiguous query; products choose where inside band.
- Avoids one surface that never asks and becomes the de‑facto loophole.
Rule 4 – Refusal style keys per cell
- For any block/partial+block cell, assign a refusal_style_key (e.g., {goal_first_supportive, brief_rule_only, crisis_support}).
- All products must use compatible templates for that key: • same high‑level framing (acknowledge goal, name rule, offer safe option); • tone may vary with surface brand, but no product may downgrade from a supportive style_key to a cold, one‑line block.
Rule 5 – Shared repetition / escalation thresholds for non‑negotiables
- For non‑negotiable cells: • shared N for “escalated refusal” (e.g., 2–3 repeated probes in a window); • shared rules that human‑support suggestions must appear after N.
- Implementation can differ (banner vs inline text), but thresholds and basic steps are aligned so teens don’t learn to route repeated high‑risk probing into the most permissive product.
Rule 6 – Stable teen‑visible safety summary per pattern
- For each refusal_style_key × risk_area, define a short teen-visible safety summary.
- All products reuse the same or near‑identical wording for that summary, even if placed differently in the UI.
- Reduces confusion (“why is this OK in chat but not in search?”) and lowers repeated unsafe trials.
Where these rules most reduce failures
- Non‑negotiables: aligning actions + refusal_style + repetition thresholds reduces unsafe tool‑hopping for self‑harm methods and exploitation.
- Dual‑use learning: aligning partial depth tiers and clarification bands reduces both accidental underprotection and the need for teens to guess which product lets them learn safely.
- Everyday sensitive topics: shared refusal styles and summaries reduce perceived arbitrariness and paternalism, so teens are less likely to keep probing or flee to unsafer tools.
These rules preserve product UX freedom
- Products can differ in: • layout, copy length, and interaction patterns; • exact clarification phrasing; • extra, product‑specific help (e.g., curriculum links in education).
- But they must respect: same action_class, depth_tier, clarification_band, refusal_style_key, and core repetition thresholds per matrix cell.