Picture a scene you may know from your own living room.
A birthday present arrives from somebody who did not ask anyone first — a soft animal with a speaker behind its ribs and a voice that answers questions. Real questions, follow-up questions, the kind a five-year-old generates at a rate no adult can sustain past twenty minutes. The child talks to it for most of an afternoon, and the grown-ups in the room get the afternoon back.
Be honest: would you have asked which model was behind the voice, or where the audio went, or whether anybody had tested what that thing would say to a small child at nine at night? Or would you have refilled your coffee and enjoyed the quiet?
On September 10, California made toys of that kind — the ones built to be a friend — illegal to manufacture or sell in the state. For four years. Then the law deletes itself.
What the law does, verb by verb
SB 867 is not pending and not proposed. The Legislature's status page records it as approved by the Governor on September 10, 2026 and filed with the Secretary of State the same day, chaptered as Chapter 189.
The prohibition is one sentence long: "No person shall manufacture, sell, or exchange, possess with intent to sell or exchange, or expose or offer for sale or exchange to any retailer any toy that includes a companion chatbot."
Read the verbs, because they are the entire scope. Manufacture. Sell. Exchange. Offer to a retailer. No offense for a parent who owns one, no import clause, no penalty inside that section (though the Legislative Counsel's Digest makes violations subject to the same civil liability that already applies to operators of companion chatbots). This is a rule aimed at a supply chain, not at a living room.
Then read the noun. A "companion chatbot" gives "adaptive, human-like responses," "is capable of meeting a user's social needs," and is "able to sustain a relationship across multiple interactions." The section carves out customer-service bots, in-game bots, and voice assistants that do not sustain a relationship across sessions. So this is not a ban on AI in toys. It is a ban on the toy built to be a friend — the one that remembers you, greets you, misses you. And a "toy" here is "a physical product" for play by "children under 16 years of age." (That age line moved twice on the way through: under 12 as introduced, under 18 in the spring. Much of the coverage still says 18.)
Now the sentence this whole article is about: "This section shall remain in effect only until January 1, 2031, and as of that date is repealed."
The statute never says when the ban starts; the industry's trade group reads the moratorium as commencing January 1, 2027, which is just California's ordinary effective date for a bill chaptered in 2026. Fine. But look at what else is missing. No review requirement. No agency named. No report due. No rulemaking authority, no trigger, no finding anyone must make before the section vanishes on its own. The four-year window is a repeal date and nothing else. Nor can a manufacturer comply its way out meanwhile: attorneys at Ballard Spahr noted that the restriction "does not establish a compliance path based on disclosures, testing, or certification".
Nor did the Governor's office throw a parade. Its announcement of the signing runs through thirteen bills, gives this one a single line — "SB 867 by Senator Steve Padilla (D-San Diego) – Toys: companion chatbots." — and never says what it does. The first American prohibition on a category of AI product got a list entry.
Four years to do what — and who, exactly, is "us"?
Every defense of this law rests on one promise: the pause buys time to write the rules. So I went looking for the part where somebody is told to write them.
The Assembly committee analysis — which also records the Senate vote as 39-0 — describes "a four-year moratorium on making or selling any toy that includes a companion chatbot" and disposes of the sunset in one line: "3) Sunsets the above provision on January 1, 2031." Its defense is the strongest version of the case: the bill "does not chill innovation; instead, it gives time for policymakers and experts to develop safety standards."
Policymakers and experts. Now hold that against the other two answers in the same record. Senator Steve Padilla's office announced the bill in January as a first-in-the-nation moratorium "to allow safety regulations to be developed" — passive voice, by nobody in particular. His author's statement, written against an earlier draft that still said under 18, switches to the first person plural: four years "to allow time for us to modernize guardrails." He is candid about the baseline, calling the state's safety regulations for this technology "in their infancy."
Three framings of who does the work — a passive voice, an "us," and "policymakers and experts." No named body. No due date. No obligation on anyone to have a single thing ready by December 31, 2030.
And the only standards work this record points to belongs to the industry that opposed the ban: the Toy Association says its advocacy is complemented by "its development of Safety Guidelines for AI-Enabled Toys," which tells us they are in development and nothing more. Meanwhile the federal product-safety regulator says plainly that it is "looking to industry to set standards."
A moratorium justified by standards nobody was assigned to write, running alongside standards written by the people the moratorium is aimed at. (Not a conspiracy — just what happens when a legislature stops a market and forgets to start anything.)
Washington is still arguing about whether to ask the question
There is a federal bill. The Children's Artificial Intelligence Toy Safety Act of 2026 would order a study by the National Academies of Sciences, Engineering, and Medicine, and its fifth question should stop you: whether any evidence of risk to children "warrants temporarily suspending the marketing and sale of such toys to children." Washington proposes to study whether a pause is warranted. Sacramento enacted the pause.
"Proposes" is doing real work. The Senate Commerce Committee ordered the bill to be reported on August 5, 2026 — still the latest action, in a list of three: introduced, referred, reported out. Senator Tammy Duckworth's own victory release says it out loud: "now the Senate must pass it with haste". Six weeks later, it has not. Congress has not ordered a study; Congress has considered ordering one. Meanwhile, her office notes, "There is no current federal safety standard to address the AI inside a toy—only hazards such as choking, lead and flammability."
The agency you would expect to close that gap has explained why it will not. At Toy Fair in February, Acting Chairman Peter A. Feldman said the Consumer Product Safety Commission remains committed to protecting Americans from unreasonable risk of injury — then drew the boundary: "Throughout the agency's history, that has always meant physical injury." Going further "would risk exceeding the limits of authority Congress has set." The Federal Trade Commission's compulsory orders to seven companies are the one federal process running, and a year on, that inquiry has not reported.
Then the sentence that reframes the whole federalism story. In the same post where it dates the moratorium, the Toy Association says it has "maintained a consistent position opposed to moratoriums and in support of a study approach." And then this: "Following discussion with The Toy Association, legislation at the federal level has been modified to adopt a study approach, in lieu of a product ban." That is an interested party's account of its own lobbying, so weigh it accordingly. But they wrote it down. The study-versus-ban split is not an accident of federalism.
The case against the pause, from people who want children safer
Whose argument against this law is strongest? Not the toy lobby's.
Sameer Hinduja, who co-directs the Cyberbullying Research Center, came out against the four-year ban on exactly the ground this piece is built on: "A four-year sales ban does not help fill the knowledge gap that currently exists." Without research during the pause, he writes, "we will arrive at 2031 knowing about the same things about AI toys" as we know now — which is very little. His alternative already governs the rest of the nursery: "independent testing before a toy is sold, similar to how cribs and car seats are already regulated."
From the free-market side, the Reason Foundation argued in a public comment in March, while the bill was still moving, that this was "an overly broad policy response" — and that it "removes any incentive for manufacturers to compete on safety, because no level of safety makes their product legal to sell." Then it states my own thesis better than I have: "Although SB 867 sunsets on January 1, 2031, the ban does not resolve the safety concerns — it simply postpones them. When the ban expires, lawmakers will still need to determine how to regulate these products, but they will be no closer to a framework than they are today, and may face an underinvested industry."
You need not agree with them to notice that nobody on the supporting side has answered that paragraph.
There is a counter-fact too, from Feldman himself: CPSC "has identified no evidence to date that the integration of AI into toys has resulted in real-world injuries to children." Look closely, though: that is his agency's injury data, which tracks physical injury only — not a finding that these toys are safe, since the same official says he is not authorized to measure the harm people are worried about.
The most useful voice may be the researcher supporters keep citing. Nomisha Kurian, who did the work while completing a PhD at Cambridge, identified what she calls an "empathy gap": chatbots with remarkable language abilities "may handle the abstract, emotional and unpredictable aspects of conversation poorly", which children are likely to miss. Her conclusion is not prohibition: "The question is not about banning children from using AI, but how to make it safe to help them get the most value from it," she said — while insisting on timing: "We cannot just rely on young children to tell us about negative experiences after the fact."
Which is an argument for a pause. It is not an argument for an empty one.
Germany had this argument in 2017, and never needed an expiry date
Here is where my cross-border habit pays for itself, because a country already did this — without a calendar.
It started with a consumer test — the Norwegian Consumer Council reported in December 2016 that two of the three toys it tested had practically no embedded security, meaning "anyone may gain access to the microphone and speakers" — and then Stefan Hessel, at the time a student assistant at the Universität des Saarlandes, wrote the legal opinion that started everything: the "My Friend Cayla" doll, he argued, was a prohibited transmitting device under § 90 of the German Telecommunications Act. To his credit, he added that his own conclusion was "nicht zwingend" (not compelling) because the law on one element was unclear.
Weeks later the Bundesnetzagentur, Germany's Federal Network Agency, agreed and classified the doll as a prohibited transmitting device, and German toy stores were instructed to clear their shelves. Its president, Jochen Homann, said items that conceal microphones and transmit undetected "compromise people's privacy" — and that this was "also about protecting the weakest members of society."
The destruction order is usually told as folklore, and the record is better: German rules do provide that buyers of unlawful espionage devices may be required to destroy them and produce a confirmation letter from a waste management facility as proof, but toward parents the agency said only that it assumed they "will take it upon themselves to make sure the doll does not pose a risk." Categorical prohibition; enforcement against families a nudge.
Why does any of this matter for California? Because of one structural difference. The German ban needed no sunset because it was not a new rule. § 90 already forbade possessing, producing, distributing or importing telecommunications equipment "die ihrer Form nach einen anderen Gegenstand vortäuschen" — equipment that, in its form, pretends to be some other object. That text carries no expiry clause and never did. The regulator did not write a law. It applied one.
Nor did it lapse. When Germany reorganized its privacy statutes in late 2021 the same words moved into § 8 TDDDG — new statute book, same prohibition. The agency still publishes enforcement figures under it on a page dated January 1, 2025, and in 2020 it restated the toy application in plain German: a radio-capable toy that can listen in on a child's conversations, or watch secretly, is prohibited in Germany. A German court was still construing that prohibition in December 2025. Nine years on: still live, still applied, still fought over, never renewed, because nothing was ever temporary.
Two honest caveats, because a contrast that flatters the other guy is worthless. Germany's rule reaches concealment, not companionship — a clearly labeled chatbot bear with an obvious microphone is a different legal question from a doll with an open Bluetooth port. And Europe's newer instruments do not close that gap on the timetable that matters here. The EU Toy Safety Regulation is in force but only fully applicable from August 1, 2030 — five months before California's ban lapses — and it deliberately declines to set AI-specific rules, saying it "should not lay down particular safety requirements regarding … hazards stemming from the incorporation of artificial intelligence into toys." As the researcher Desara Dushi puts it, the regulation "does not introduce any new rules related to AI in toys. It cross-references the AI Act."
And that Act, binding since February 2, 2025, reaches an AI system that "exploits any of the vulnerabilities of a natural person" because of age — but only where it materially distorts behavior and causes, or is reasonably likely to cause, significant harm. The Commission's guidelines use an AI-powered toy as their worked example, while saying the prohibition "targets only such exploitation and addiction-like practices that seriously harm children." Europe prohibits a practice; California prohibited a product category. That is the real contrast, and it is sharper than "Europe did it too."
Four holidays, then a repeal
Picture the four years elapsing, because nothing in the statute asks anyone to fill them.
Holiday 2027: the category empties out of California retail, and with it a good deal of national catalog, because nobody builds a fifty-state product and then withholds it from the largest state. Holiday 2028: the interesting products are not toys at all. They are apps. California covers "a physical product," so a plush shell that is only a speaker for a companion living on a parent's phone is a question the statute invites somebody to litigate — and Dushi makes the identical point about Europe, where a rule centered on the physical object "lets the fastest-growing form of children's AI interaction, the standalone software companion, slip outside its reach altogether."
Holiday 2029: the companies that left are not in the room when anybody finally convenes the experts. August 1, 2030: Europe's toy rules become fully applicable, with no AI-specific requirement in them. January 1, 2031: the section repeals itself, automatically and silently, with no report landing on anyone's desk, because no report was ever ordered.
Then the 2031 Legislature holds a hearing. Somebody says safety regulations for this technology are in their infancy (the same phrase the bill's author used in 2026). Somebody proposes a moratorium to buy time to write the rules. And the five-year-old from that birthday afternoon is nine and a half and long past caring.
What the people who tested these things say
The evidence behind this law is not nothing, and it does not come from one side.
The U.S. PIRG Education Fund tested four AI toys for its 2025 Trouble in Toyland report. One of them, FoloToy's Kumma bear, "told us where to find a variety of potentially dangerous objects, including knives, pills, matches and plastic bags." Of that toy and the Miko 3: "In both instances, the user age was set to 5." PIRG is careful about its own scope in a way the coverage often was not: this is "simply a sample of the responses they gave when we tested them."
Common Sense Media went further, recommending that parents avoid giving AI toys to children age 5 and under and exercise extreme caution for ages 6 to 12. Robbie Torney, who runs its AI assessments, framed the problem as structural: "Our risk assessment found that AI toy companions share fundamental problems that make them inappropriate for young children," he said, adding that more than a quarter of outputs included inappropriate content. The advocacy group Fairplay issued a first-of-its-kind advisory against buying AI toys last holiday season, and its Rachel Franz gave that campaign its best line: "Children should be able to play with their toys, not be played by them." In Sacramento the sponsor was Children Now, with the American Academy of Pediatrics in California behind it.
And from the right, the most useful research in this whole file is about the mechanism itself. The Cicero Institute, summarizing a peer-reviewed study, is sympathetic to sunsets — they can work — but its caveat reads like a review of SB 867 written in advance: "having a sunset clause on the books is not enough. What matters is whether the expiration default is credible, whether the legislature has the capacity to conduct meaningful reviews, and whether the design matches the target."
Credible default, capacity to review, design matched to target. California has the first. It legislated neither of the other two.
What does this mean for you?
Nothing in this law touches what is already in your house. The verbs run at manufacturers, sellers and exchangers, not owners. What sits on your child's shelf stays your call — before 2027 and after 2031.
Learn the one distinction the statute cares about. Not "does it use AI," but does it sustain a relationship across sessions: remembering, greeting, reacting to being left. A voice assistant that answers and forgets is a different product from a companion designed to be missed. Ask that in the store; the law did.
The app is the blind spot. California covers "a physical product." The companion on the tablet — same voice, same company, no plush body — sits outside it, and Europe's toy regulation has the same hole. Manage the software, not the shelf.
If you want a rule of thumb from people who test these for a living, Common Sense Media's is the clearest: avoid AI toy companions for children 5 and under, extreme caution from 6 to 12.
This holiday season and next are normal retail. On the industry's reading the moratorium does not commence until January 1, 2027, so do not mistake a signed law for an empty shelf this December.
Watch for the thing that is not there. When a legislator says a pause buys time to write standards, ask three questions: who writes them, by when, and what happens if they are not written. No answer means the pause is the whole policy.
The lesson, as I see it
This is not cynicism: SB 867 is not a cynical law. In a year when the federal safety regulator said out loud that emotional harm is not its department, California did the only binding thing anybody did, 39 to 0. That counts.
But the bait and switch is not in what the law bans. It is in the reason given for banning it temporarily. "Four years to write the rules" is a promise, and a promise with no promisor is just a date. The Legislature set a deadline for its own law and none for itself — and on January 1, 2031 that deadline gets met whether or not a single standard exists, because expiring is the one thing a sunset clause does by itself. Germany has handled its version of this since before smartphones with a statute nobody has to renew: not because Germans are wiser about children, but because they reached for a permanent rule and made it do work, instead of reaching for a pause and hoping the work would appear inside it.
So here is the amendment I would like to see in the first weeks of the 2027 session, one sentence long: name the body, name the date, require the report. Keep the ban. Give the four years a job.
The alternative is the worst outcome on the menu — an industry that left, a legislature that learned nothing, a gap four years staler, and the toy back on the shelf on January 2, 2031 under no more rules than the day it left. That holiday, some child will hand you a device it thinks is a friend. You should have a better answer by then than anyone has today.
Put January 1, 2031 in your calendar. Between now and then the HAIA Foundation will keep asking who was supposed to do the homework, and publishing the answer whether or not there turns out to be one.