$1,000 per violation. Plus attorney's fees.
Private right of action.

CA SB 243 — User Disclosure

“Users must be told they are interacting with an AI chatbot. Disclosure must be clear, conspicuous, and recorded.”

Cortex Workflows + Vault disclosure ledger

Every conversation session captures the disclosure shown to the user, the disclosure version, the user acknowledgment (where required), and the timestamp — all recorded as a Vault event tied to the session ID.

Audit evidence: Per-session disclosure record auditable by the California AG. Reconstruct any conversation's disclosure compliance from one query.

CA SB 243 — Self-Harm and Suicide Safety Protocols

“Operators must implement protocols for users expressing suicidal ideation or self-harm intent.”

Guard content classifier + human_review escalation

DeBERTa-based content classifier runs on user input AND chatbot output. Detection triggers immediate routing to a human_review workflow step with crisis-resource surfacing and Vault-logged escalation.

Audit evidence: Per-detection audit trail: classifier confidence, message excerpt, escalation action, resource provided, human reviewer identity.

FTC Five "Don'ts" for AI Chatbots

“Don't misrepresent what the AI is. Don't deploy without harm mitigation. Don't exploit emotional relationships for commercial gain. Don't make unsubstantiated claims. Don't use automation to mislead.”

Cedar policy library + cross-model verification

Declarative Cedar policies encode the company's content standards; cross-model verification routes any output flagged by policy through a majority-vote consensus check before it ships to the user.

Audit evidence: Per-message policy evaluation record showing which rules matched, which models verified, which decisions were made. The FTC's five-don'ts compliance is provable, not asserted.

Known-Minor Safeguards (SB 243)

“When the operator knows or should reasonably know that the user is a minor, additional protections apply.”

Gate per-user risk classification + Cortex routing

User metadata (declared age, signals from the session) informs which Cedar policy tier applies, which model is selected, and which safety floor governs the conversation. Minors get a stricter policy without operator code changes.

Audit evidence: Per-session policy tier record showing which floor was applied and why. The age-based protection works because it's centralized, not because every conversation got it right on its own.

Federal AI Policy Statement (FTC, due March 11, 2026)

“Per Executive Order 14178 (Dec 11, 2025), the FTC must publish its formal AI policy statement on how authority over unfair and deceptive practices applies to AI.”

Vault per-conversation evidence pack

Full session transcript, policies evaluated, decisions made, disclosures shown, classifier outputs, escalations triggered — all in one chain-linked export. FTC investigates? You hand them the evidence pack.

Audit evidence: Per-session export in FTC-ready format. The whole conversation reconstructed cryptographically, not pieced together from logs.