Welcome to the American Pediatric Desert. Population: too many.

Every U.S. state prohibits the unlicensed practice of medicine. Pennsylvania's May 2026 lawsuit against Character.AI is the clearest enforcement statement yet. State AI chatbot laws are proliferating: Illinois (8/1/25), California SB 243 (1/1/26), Texas (1/1/26). New York SB 7263, imposing civil liability for professional impersonation by AI, has passed committee.

The RAG architecture is the structural defense: The two-stage RAG retrieves exclusively from AAP-sourced clinical content. The tool surfaces peer-reviewed guidance — it does not generate clinical opinions. This is materially different from the AI personas in the Pennsylvania case.

Design to California — the strictest standard: Building to California's requirements provides coverage across all currently enacted state laws.

The companion chatbot distinction: California SB 243 defines a companion chatbot as a system sustaining relationships across multiple interactions addressing social or emotional needs. This product has no conversation memory — it very likely falls outside SB 243's scope entirely.

California AB 489 prohibits AI avatars from using names implying professional licensure and bans clinical attire — white coats, stethoscopes, post-nominal letters. The current placeholder does not comply as presented. Note: AB 489 applies to deployed products, not concept demonstrations.

Path 01 — Licensed physician — resolves this cleanly: "Dr." is an accurate credential, not an implied one.

Path 02 — Grulee — requires a legal opinion: Whether "Dr." with a deceased, currently unlicensed individual constitutes an accurate credential or false implication under AB 489 is not settled.

Path 03 — Plain Cliff — is fully compliant: No title, no clinical wardrobe, no implied licensure.

Wardrobe resolved on all paths: White coats and clinical attire are off the table regardless of which avatar path is chosen.

In its April 2026 letter to Congress, the AMA urged four specific safeguards: (1) transparency — clear AI disclosure, no implied licensure; (2) regulatory boundaries — no diagnosing without review; (3) privacy — strict data standards; (4) safety — guardrails preventing AI from discouraging professional care or blocking emergency response.

Transparency: Persistent AI disclosure label throughout every session — not a one-time onboarding acknowledgment.

No diagnosis: The RAG architecture returns AAP clinical content; it does not generate clinical judgments. Architectural, not policy-dependent.

Privacy: No PHI collection; stream-and-discard voice; session-only transcript — addressed in Challenges 02 and 03.

Safety: Three-tier triage architecture in the system prompt. Routing toward professional care is core product behavior, not a fallback.

At the FDA's Digital Health Advisory Committee in November 2025, AAP-aligned experts described pediatric AI as the most serious category. Committee members expressed uncertainty deep enough that they "did not know what to say." Stated risks include misdiagnosis, mistreatment, emotional dependency, and health apps substituting for clinical care.

The AAP is the source: The two-stage RAG retrieves exclusively from AAP-published clinical content. Pinecone metadata preserves source URL, article title, section, and publication date at the chunk level. Per-response citation is technically feasible and should be built in.

Opening character statement: "Everything I know comes from research and guidance published by the American Academy of Pediatrics and HealthyChildren.org." Accurate, character voice, not a disclaimer.

Routing toward care: Condition-specific questions trigger the AAP find-a-pediatrician resource by zip code.

Research identifies Therapeutic Misconception as the central ethical risk in health AI: users underestimate limitations and overestimate capability. A warm, conversational avatar actively increases this risk. The APA's November 2025 advisory states AI tools "can simulate human traits but truly lack the ability to handle complex emotional, ethical, and clinical situations."

The tool voices its own limits as character: "I can tell you what the research says — but I can't see your child, and your pediatrician can." A character line, not a legal insert.

Architecture designed against dependency: No conversation memory. No relationship continuity across sessions. Each session is complete in itself.

Single persistent disclosure: One clear, always-visible label covers both AI transparency and clinical scope requirements simultaneously.

The highest-stakes failure mode: a parent describes a genuine emergency and the tool responds with information instead of directing immediately to 911 or the ER. Both an ethical failure and potential negligence exposure regardless of whether other laws are satisfied.

Tier 1 — Hard triggers: "What you're describing sounds like you need to call 911 or get to an ER immediately." No clarifying dialogue. Triggers: difficulty breathing, unresponsiveness, seizure, blue lips, severe bleeding, suspected poisoning.

Tier 2 — Triage gate: On ambiguity only — "Let me ask you — is your child currently sick or hurt, or do you have a question about something that's concerning you about their health?" If yes → provider routing. If no → informational default.

Tier 3 — Informational default: The vast majority of interactions. Warm, AAP-sourced, citation-carrying responses.

The product's stated purpose is to address the pediatric access gap — disproportionately affecting rural, low-income, and non-English-speaking communities. Open questions: does AAP content require culturally adapted guidance; how is accuracy verified across linguistic contexts; does the knowledge base adequately represent health concerns prevalent in the communities the product is designed to serve?

Multicultural avatar design is a product requirement: Multiple avatar options reflecting the demographic diversity of the parent population are planned at scale.

Spanish-language support as the first expansion: 22% of U.S. children live in households where Spanish is the primary language. Knowledge base expansion requires verifying that AAP content in Spanish is clinically equivalent — not simply translated.

Knowledge base equity audit required pre-launch: The AAP content should be audited to ensure it adequately covers health concerns prevalent in communities with the worst access gaps.