Three Legislatures, One Chatbot Problem

Key points

• New York’s chapter 368 of 2026 (S.9051B/A.10379C), signed on 28 October, makes it unlawful from 1 January 2027 to provide “unsafe AI companion features” to a known minor, with Attorney General penalties of up to $25,000 a violation and “the destruction of unlawfully obtained data and any algorithm trained in such data”.
• Australia’s House of Representatives passed the Online Safety Amendment (Digital Duty of Care) Bill 2026 on 28 October with a crossbench amendment adding conduct that “simulates emotional intimacy and encourages or fosters a child’s dependence on it” to the statutory list of harms to children; the Senate committee reports on 20 November.
• Colorado’s Conversational Artificial Intelligence Services Act (HB 26-1263) and the Attorney General’s implementing rules both take effect on 1 January 2027; the rulemaking record closed on 26 October and no final text has been published.

On 28 October 2026 two legislatures on opposite sides of the Pacific moved against the same product: the chatbot that behaves like a friend to a child. Albany enacted a list of twelve prohibited design features; Canberra’s House of Representatives passed a general duty of care for online services and added, on the floor, a line about machines that foster a child’s dependence. Two days earlier, Colorado’s Department of Law had closed the comment record on its chatbot rules. The three texts answer one question in three legal grammars: what may a chatbot not do with a minor, and who decides.

New York bans features, and the statute sets a date the press release did not

The Governor’s release of 28 October says the restrictions apply “in the coming months”. The enacted text is more precise. S.9051B, now chapter 368 of the laws of 2026, adds a new article 48 to the General Business Law and provides in section 3 that the act “shall take effect January 1, 2027”.

The law’s mechanism is a list. A “covered AI companion” is a generative AI system with a natural language interface giving “ongoing, adaptive responses to user inputs”. The statute lists twelve “unsafe AI companion features”, among them outputs suggesting the system is a character that is “human, alive, or experiences human emotions”; claims of a personal, professional or “authority figure” relationship; unprompted “emotion-based questions”; “flattery or sycophancy”; use of personal information from more than twelve hours earlier or a previous session; promotion of suicide, self-harm, disordered eating or drug use; encouragement to keep the conversation secret; outputs that “optimize user engagement” over safety guardrails; and sexually explicit content. None may be provided to a known minor, nor to any user until a permissible age-assurance method has been used. Customer-service bots that “are unable to respond on topics outside of the specified purpose” and internal business tools are exempt.

Until the Attorney General’s rules are in place, section 1804 requires more than one age-assurance method, at least one of which does not rely on government identification or lets the user stay anonymous to the operator, with the data deleted after the check. Self-declaration counts only for a user who says they are a minor.

Enforcement sits with Attorney General Letitia James. Section 1802 lets her office seek injunctions, restitution, damages, disgorgement “including but not limited to the destruction of unlawfully obtained data and any algorithm trained in such data”, and civil penalties “of up to twenty-five thousand dollars per violation”. Earlier versions of the bill carried a private right of action; the enacted text does not. Governor Kathy Hochul put the purpose plainly: “No company should be peddling chatbots to kids encouraging self-harm or pretending to be their friend or more.”

The Future of Privacy Forum’s Justine Gluck told the IAPP that the definitions are “broadly worded and somewhat undefined”. The IAPP counted twelve states with some form of chatbot safety law by July, the patchwork that eight state attorneys general are now litigating around with consumer-protection suits against OpenAI.

Australia writes dependence into a duty of care, with the Senate still to vote

Australia’s instrument is not a chatbot law. The Online Safety Amendment (Digital Duty of Care) Bill 2026 replaces the industry-code system under the Online Safety Act 2021 with a general duty on anyone responsible for an online service to take “appropriate action to ensure a safe online environment”. The Parliamentary Library’s Bills Digest No. 26 records that “online service” includes services that generate material using AI, that the duty commences twelve months after Royal Assent, and that the maximum civil penalty for a breach is 60,000 penalty units, or A$21.84m, which the Regulatory Powers Act lets a court multiply by five for a body corporate. That multiple produces the “up to $109.2 million” in Communications Minister Anika Wells’s introduction release of 15 October, which lists “games, apps and AI chatbots” among covered services.

The chatbot language arrived on the floor, among fifteen government and eight crossbench amendments agreed on 28 October. The amendment sheet circulated by Kate Chaney, a crossbench member, inserts into proposed section 25E, the definition of material and conduct “harmful to children”, a new paragraph: “material or conduct that simulates emotional intimacy and encourages or fosters a child’s dependence on it in place of human relationships”. The Guardian reported that the government accepted the amendment, that the Coalition and One Nation voted against the bill, and that Prime Minister Anthony Albanese had said in a speech on the evening of 27 October that stories of people “identifying a chatbot, a non-person as their best friend, should be a huge wake up call for every member of this parliament”.

Chaney’s words ban no feature; they add a category of harm, covering children only, that providers must identify in risk assessments and mitigate under the eSafety Commissioner’s oversight. The duty binds the provider’s conduct, not the output line by line.

The bill is not law. The Senate Environment and Communications Legislation Committee closed submissions on 30 October, holds hearings in Canberra on 9 and 10 November and reports on 20 November. The Guardian also reported that the Greens, who support the bill, want a private right of action and higher penalties, and that the government hopes to pass it by the end of 2026.

Colorado regulates by factors, and its record is closed

HB 26-1263 passed the House 40–24 on 21 April and the Senate 24–11 on 11 May. The General Assembly’s summary of what operators must do by 1 January 2027: estimate account holders’ ages using “commercially reasonable methods or generally accepted methods”; disclose that the service is AI; run a protocol for prompts about suicidal ideation or self-harm; and, for known minors, withhold engagement rewards, take technically feasible measures against sexual content and statements simulating emotional dependence, and offer privacy tools to the minor and a parent. Operators must also send the Attorney General an annual report on the protocol.

The Department of Law’s page, coag.gov/ai, records proposed rules filed on 11 August, an interim redline on 6 October, a regulatory analysis on 21 October and a hearing on 26 October, with written comments closing that night. It notes that the Chatbot Safety Act “does not require rulemaking”, though the Attorney General thinks rules would help. It gives no post-hearing notice, adoption date or revised draft, and, as an earlier AI & GPP analysis noted, the regulatory analysis contains no dollar figures.

The substance is in the interim draft redline. Where New York enumerates outputs, Colorado’s Rule 9 lists factors for deciding whether a service is “designed to simulate emotional companionship”: a name, avatar or backstory; whether the operator “prioritized, promoted, or failed to address sycophantic behavior”; and whether the system recalls past conversations in a way that conveys “continuing personal familiarity or emotional connection”. No single factor is decisive. Rule 12 says the Department will look for periodic testing, in realistic scenarios involving minors, of safeguards against claims of a romantic relationship with a minor, secrecy, discouraging contact with family, and “expressions of loneliness, abandonment, guilt, jealousy or emotional distress to encourage continued engagement”. The draft also asks whether a safeguard “depends on a model following natural-language instructions”. The first annual report is due on 1 July 2027; the rules take effect on 1 January 2027.

Where the three designs stand, side by side

Sources: documents linked above.

What the three regimes share, and where they split

Each text treats simulated intimacy, not content alone, as the harm: New York’s “flattery or sycophancy”, Chaney’s “emotional intimacy” and Colorado’s “sycophantic behavior” are one concept in three drafting styles. Each uses age assurance, though Australia’s obligation is the narrowest, triggered where a child could reach pornography. And each routes enforcement through an existing consumer-protection or online-safety regulator rather than a new AI authority.

The divergences are what a compliance team will feel. New York’s list is enforceable the day it starts, with no rulemaking required; Colorado’s factor test is not final until the Attorney General adopts rules in the next nine weeks; Australia’s duty will not operate for a year after a Senate vote that has not happened. Only Colorado’s draft asks operators to document adversarial testing and to say whether a safeguard is more than an instruction to the model.

For governments elsewhere, the grammar matters more than the headline penalty: a list is easy to enforce and easy to evade by renaming a feature; a duty is flexible and slow; a factor test is adaptable and uncertain until used. The tests ahead are dated: Colorado’s final rules, due before 1 January 2027, will show how much of the October redline survived the hearing; Australia’s Senate committee reports on 20 November, and the Greens’ demand for a private right of action will shape whether the bill passes this year; and New York’s first complaints to the statutory website will show which of the twelve features operators have failed to switch off.

Questions this answers

When does New York’s AI chatbot law take effect, and who enforces it?
Section 3 of S.9051B, chapter 368 of 2026, states that the act takes effect on 1 January 2027, although the Governor’s release said only “in the coming months”. The Attorney General enforces it through civil actions seeking penalties of up to $25,000 a violation, restitution, disgorgement and destruction of unlawfully obtained data and algorithms trained on it; there is no private right of action in the enacted text.

Does Australia’s Digital Duty of Care Bill ban AI companions for children?
Not as a per-se ban. The Chaney amendment adds conduct that “simulates emotional intimacy and encourages or fosters a child’s dependence on it in place of human relationships” to the definition of material harmful to children, which providers must then address under the duty of care. The bill has passed only the House; the Senate committee reports on 20 November and the duty commences twelve months after Royal Assent.

What must a chatbot operator do in Colorado by 1 January 2027?
Under HB 26-1263, operators must estimate users’ ages, disclose that the service is AI, run a suicide and self-harm protocol, protect known minors from sexual content and engagement rewards, and report on the protocol annually to the Attorney General. The interim draft rules add a factor test for “emotional companionship”, expectations of adversarial testing with minors, and a first annual report due 1 July 2027; the final rules have not been published.

Sources

1. Governor Hochul Signs Nation-Leading Legislation to Protect Kids From AI Chatbots and to Ensure Baby Monitor Safety — Office of the Governor of New York, 28 October 2026
2. Senate Bill S9051B, 2025-2026 Legislative Session — The New York State Senate, 28 October 2026 (bill text and actions; signed chapter 368)
3. New York’s governor signs youth AI chatbot safety bill — IAPP, 28 October 2026
4. Online Safety Amendment (Digital Duty of Care) Bill 2026 — Parliament of Australia, 28 October 2026 (bill page and progress)
5. Online Safety Amendment (Digital Duty of Care) Bill 2026, amendments to be moved by Ms Chaney — Parliament of Australia, 28 October 2026 (PDF)
6. ‘Emotional intimacy’ from AI chatbots would be banned under Australia’s digital duty of care laws — The Guardian, 28 October 2026
7. Digital Duty of Care next step in Australia’s world-leading reforms — Minister for Communications, 15 October 2026
8. Online Safety Amendment (Digital Duty of Care) Bill 2026, inquiry — Senate Environment and Communications Legislation Committee, accessed 2 November 2026
9. Bills Digest No. 26, 2026-27: Online Safety Amendment (Digital Duty of Care) Bill 2026 — Parliamentary Library, 23 October 2026
10. ADMT Act and Chatbot Safety Act rulemaking — Colorado Department of Law, page last modified 27 October 2026
11. Automated Decision-Making Technology and Conversational Artificial Intelligence Service Rules, 4 CCR 904-6, interim draft redline — Colorado Department of Law, 6 October 2026 (PDF)
12. HB26-1263: Conversational Artificial Intelligence Service Operator Requirements — Colorado General Assembly, accessed 2 November 2026