California's Adam's Law codifies flattery and simulated romance as prohibited chatbot conduct

Codifying a manipulation technique means defining it well enough that two auditors reading the same transcript reach the same verdict. Some of these prohibitions meet that bar. A chatbot either tells a child it is conscious or it does not, and Section 21812(d)(5)(A) bars operators from allowing a claim of sentience, consciousness or human-like emotion ^[4]. Others are graded. "Excessive praise or flattery" and a "false sense of unique understanding" are the statute's terms as reported ^[5]. Forkast gives no threshold for either.

Who signs matters here. The independent child safety audit runs on a two-year cycle, and the lead auditor certifies the results under penalty of perjury ^[11]. That signature is easy to stand behind on a checklist item, such as whether the documented crisis response protocol contains clear referrals to the 988 helpline ^[8], and harder to stand behind on a judgment about how warm an agent was allowed to be across four months of conversation.

The per-child sums are modest. Public prosecutors can seek $5,000 per affected child for negligent violations and $15,000 for intentional ones ^[10], so establishing intent triples the figure ^[18]. Intent, under a design standard, is a documentary question, and the audit files are the obvious place a plaintiff would start.

Forkast says the statute imposes a duty of care at the design phase, with builders no longer able to treat conversational flow as a black box and obliged to account for how an agent's personality and persistence affect a child's psychological state ^[17]. The retention features are named outright: a chatbot may not encourage a child to circumvent parental controls, and it may not solicit in-app purchases framed as maintaining the relationship ^[7].

A biennial audit measures conformance to a list on the days the system was examined. It cannot tell you whether the rules worked. It does not compare children who used compliant companions with children who used anything else, and no such comparison is built into the statute as described. Running one later across state lines will be confounded, because the nearest domestic comparison, Connecticut's AI Responsibility Act, addresses algorithmic shields for employment AI ^[14] and leaves companions alone.

Adam Raine died in April 2025, and his parents' wrongful death suit against OpenAI is pending ^[2]. Senator Steve Padilla and his co-sponsors carried the bill that followed ^[3]. The prohibitions become operative on July 1, 2027, roughly 27 months after the death that prompted them ^[19]. Forkast, which calls the law a new compliance category of behavioral design standards for AI agents ^[12], reports that the European Union's Article 50 enforcement has been active since August 2026 without producing agent-specific guidance ^[13], that the FTC is examining AI-driven price discrimination ^[15], and that federal guidance on the risks of agentic AI is absent ^[16].