The Bouncer Problem: Teens & AI Companions

A unanimous vote with a catch

Nazzaro also reported that "some senators raised concerns about the age-verification measure." She wrote that Sen. Alex Padilla, a California Democrat, "who voted yes, said there are concerns about 'potential privacy and security risks' with the age-verification component, suggesting it may need to be 'fine-tuned.'" Sen. Ted Cruz, a Texas Republican, also voted yes. Two days earlier, he and three other senators from both parties had introduced the rival bill, the CHATBOT Act. According to the Senate Commerce Committee's announcement, it would require "family accounts" and parental consent.

Joe Mullin is a senior policy analyst at the Electronic Frontier Foundation, a digital civil-liberties advocacy group. On May 8, he wrote that even the narrowed bill would shut people out: "Millions of Americans do not have current government ID, accounts at major banks, or stable access to the kinds of digital identity systems the bill contemplates." The same day, a free-speech advocacy group, the Foundation for Individual Rights and Expression, or FIRE, published an essay in RealClearPolitics, which it reposted on its own site. It argued that age checks would force people to give up their anonymity.

California had already chosen a different road. In October 2025, Gov. Gavin Newsom vetoed a stricter bill. "AB 1064 imposes such broad restrictions on the use of conversational AI tools that it may unintentionally lead to a total ban on the use of these products by minors," he wrote in his veto letter, as quoted by Sam Mondros of the San Francisco Standard. This year the Legislature passed a narrower law, SB 1119. It is named for a California teenager whose family has sued OpenAI. On September 10, the day Newsom signed it, Colin Lecher of CalMatters reported that OpenAI supported the law. According to Lecher, the company said the measure "pairs strong protections with continued access to useful AI tools."

The courts are a third front. On June 1, Florida became the first state to sue OpenAI. "People are getting hurt, parents are getting deceived, and they need to pay for it," Florida Attorney General James Uthmeier, a Republican, told reporters. Diana Novak Jones of Reuters reported the remark, in a story the Hawaii Tribune-Herald also carried.

By summer the argument had at least five positions: ban minors and verify everyone's age; keep access but regulate how the products are designed; put parents in charge; reject age checks as a threat to speech and privacy; and sue. The companies themselves are split. The safety side has the lawsuits and a unanimous committee. Its opponents argue that age-check mandates are unconstitutional. In 1997 the Supreme Court struck down an online child-protection law, in part over the burden of age checks on adults. In 2025, though, it upheld a Texas age-check law.

The bouncer, then and now

Supporters of age limits compare the internet to places where the law already keeps children out. "They would lock up that liquor store, they would put that strip club operator in jail if they had kids there. But in the virtual space, it's like the Wild West 24/7, 365," Sen. Marsha Blackburn, a Tennessee Republican, said at a Senate hearing. Veronica Irwin quoted her in an explainer for Transformer, a newsletter on AI policy. Irwin answered with the cost the comparison hides: "A bouncer who checks someone's ID doesn't retain that data in his brain, or at least not with the precision that is permitted online. He does not muzzle someone he bars from entry, either."

A Supreme Court justice used the bouncer image in 1997. The case was Reno v. ACLU, a challenge to a 1996 law meant to protect minors from "indecent" material online. In a separate opinion, Justice Sandra Day O'Connor described websites that asked visitors for an adult identification number or a credit card number, "much like a bouncer checks a person's driver's license before admitting him to a nightclub." The Court's opinion accepted "the legitimacy and importance of the congressional goal of protecting children from harmful materials." It struck down the law's indecency provisions anyway, mainly because they were too vague and too broad. The burden of age checks on adults was a large part of that reasoning. The Court stated the test in one sentence: "That burden on adult speech is unacceptable if less restrictive alternatives would be at least as effective in achieving the legitimate purpose that the statute was enacted to serve."

The law has moved since. In 2025, the Supreme Court upheld a Texas law requiring age checks on pornography sites. The case was Free Speech Coalition v. Paxton. The official summary of the decision gives the reason the Court applied a middle level of review: the law "only incidentally burdens the protected speech of adults." Paxton concerned sexual material, not conversations with a machine, and nothing in it says how a court would treat age checks for an AI companion.

A harm filed under "accusations"

The May 8 essay from FIRE, the free-speech advocacy group, builds part of its case against age checks on an imagined user: "Would someone trapped in an abusive relationship be more, or less, willing to seek advice from a chatbot if they had to surrender their privacy?" The question points to a real cost.

When the essay turns to why the bills exist, it writes: "Yet artificial intelligence, particularly chatbots, has become Washington's latest political punching bag. Accusations of manipulation and harm are driving a slew of legislative proposals to censor this emerging technology." Elsewhere it refers to "concerns about AI." It does not mention teenagers' deaths, the lawsuits or the settlements. One harm, to a hypothetical person, is weighed in full. The other is called "accusations" and left there.

The most charitable reading is that "accusations" is accurate. The causes of the deaths are allegations in court. But a cause unproven in a particular death is different from a risk left off the scale entirely. A risk can be uncertain and still belong on the scale. Another group on FIRE's side makes the same legal case without that dismissal. In his post for the Electronic Frontier Foundation, Joe Mullin argues against the same bill and still writes: "The concerns driving this bill are real. Some AI systems have engaged in troubling interactions with vulnerable users, including minors." FIRE's legal argument survives. Its treatment of the other value does not.

A cause settled in a clause

The press release that Sen. Josh Hawley's office put out on the day of the vote says: "Parents of children who lost their lives or committed self-harm at the direction of AI chatbots watched as the Committee advanced the bill." In a subordinate clause, the phrase "at the direction of AI chatbots" settles the question the lawsuits exist to decide. At least one company denies it. In November 2025, Angela Yang of NBC News reported on OpenAI's court filing in a California wrongful-death suit. Of the teenager at the center of the suit, the filing says his "death, while devastating, was not caused by ChatGPT." That denial is also a party's claim, no more settled than the lawsuit it answers.

Nobody expects a press release to argue the other side. But a risk can be serious enough to legislate against without proof of cause in every case. The bill does not need the causal claim, and the clause adds nothing but a disputed claim stated as background fact.

The release lists what the GUARD Act "would" do, and age verification is not on the list. Yet it is the provision in the bill's own name, and the one Sen. Alex Padilla questioned that day. The committee had just narrowed that requirement from every chatbot to AI companions only, and the release could have claimed credit for the change. It did not. The bill's main cost goes unnamed in its sponsor's own account of it.

FIRE's essay and the release fail in opposite directions. The essay treats the harm to teenagers as an accusation. The release treats a disputed cause as settled and leaves out the cost of verification. Hardly Thinking calls this pattern G2: Liberty versus Safety. It is a real trade-off between freedom and protection from harm, argued as though only one of the two values existed.

The careful case that skips a step

In a July 2 op-ed in STAT, Ryan K. McBain made one of the more careful cases for regulation. McBain is an assistant professor at Harvard Medical School and a senior policy researcher at RAND. Starting from Meta's acknowledgment "that teens may lie about their age," he writes:

"A better default is: Treat all users as minors unless they have been verified as adults. Companion features that simulate romance, intimacy, or devotion should be disabled for minors and limited by default when age is uncertain. The burden should be on companies to prove that a user is an adult before unlocking higher-risk features, not on parents to find every account or on teenagers to choose the safest settings."

The passage names the alternatives it rejects and who should carry the burden instead, but not the cost of its own default. "Verified as adults" means that every adult who wants the unlocked features must prove their age to a company. FIRE opposes that mechanism, and so does the Electronic Frontier Foundation. The mechanism's cost is the one Joe Mullin named in his post for that group: millions of Americans without current government ID or accounts at major banks. The op-ed does not mention adults' privacy.

The charitable reading is strong. McBain's default covers only "higher-risk features," a narrower reach than the GUARD Act's. He may judge the cost to adults worth paying, and he may be right. What does not survive is the silence. The op-ed decides, without saying so, how many adults must hand over identifying data, though it is otherwise careful to name who carries each burden. Readers who want teenagers protected are likely to agree with every word of it. That is why the missing step matters.

When 10 percent becomes 31

J. B. Branch directs federal AI policy work at the consumer advocacy group Public Citizen. He made one of the fuller cases for a ban in "AI Companions Are Not Your Teen's Friend," an essay in the Fall 2025 issue of Issues in Science and Technology. He cites a survey by Common Sense Media, a child-safety advocacy group that also publishes research: "At least 31% said they find speaking with an AI companion more enjoyable than speaking with their friends."

The survey report says something narrower: "Thirty-one percent of teens find conversations with AI companions as satisfying or more satisfying than those with real-life friends (21% say about the same quality, and 10% say more satisfying). However, 67% still find AI conversations less satisfying than human conversations." The share who found the AI more satisfying was 10 percent. The 31 percent includes everyone who rated the two about the same. "At least" pushes up a number that was already the ceiling. The 67 percent who found AI conversations less satisfying are missing from the essay.

The essay opens with a lawsuit and reports it carefully, as allegations. That makes the number harder to excuse. The essay does not say how it got from the survey's wording to its own. The charitable reading is that the essay's concern does not depend on the larger figure. That 31 percent of teenagers find an AI companion at least as good as their friends is striking on its own. The concern survives. The number, as printed, does not.

"Small numbers," with no number

On April 29, the day before the committee vote, a free-market think tank, the R Street Institute, published a commentary against the GUARD Act. It acknowledges "a number of sobering cases" involving teenagers and chatbots. It then argues that the bill's supporters "focus on small numbers of worst-case outcomes from youth interactions with generative AI to the complete exclusion of all the ways that young people can benefit and learn from chatbots." It continues: "A recent Pew Research poll shows that a majority of U.S. teens use chatbots, and more of them say that AI bots have a positive impact on their lives than the opposite."

"Small numbers" is a measurement, and the commentary gives none. Nobody has one. In her Transformer explainer, Veronica Irwin wrote that "There's no official count of how many children have been harmed." She then made an estimate anyway. She combined an OpenAI figure for all ChatGPT users in a given week with a Pew finding that about two-thirds of teenagers have used a chatbot: "Taken together, that suggests a significant number of teens are discussing suicide with AI chatbots." A weekly rate for all users and a share of teenagers who have ever used any chatbot do not combine into a number. "Significant" stands in for one.

The Pew report that R Street cites says something different: "While 36% say it'll have a positive impact on them personally over the next 20 years, far fewer (15%) say this effect will be negative." That is a forecast about AI in general over 20 years, not a report that chatbots are helping teenagers now.

R Street worries that policy built on worst cases can do harm, and that worry holds up. But the honest version of its sentence would say that nobody knows how many teenagers have been harmed. That uncertainty cuts against confident claims on both sides.

The Pattern Underneath: liberty versus safety

A rule that protects people from harm almost always limits someone's freedom. The honest argument is about how much of one to trade for the other. The common failure is to argue as if only one value were real. Hardly Thinking calls the pattern G2: Liberty versus Safety. The argument can fail in two directions.

At the liberty pole, people ignore real harms to others in the name of freedom. FIRE's essay is the plainest case in this fight: the harm to teenagers enters only as "accusations."

At the safety pole, any conceivable harm justifies any rule. The Hawley press release states a contested cause as fact and leaves out the cost. According to Sam Mondros's report in the San Francisco Standard, Jim Steyer, the founder and chief executive of Common Sense Media, said of the bill Gov. Gavin Newsom vetoed in 2025: "This is a very straightforward, simple choice." In his veto letter, the governor warned of "a total ban on the use of these products by minors." A choice with that cost on one side was not simple.

The two poles meet at the age question. Every position needs an answer to it: a government ID, a bank account, a signal from the phone's operating system, a guess from how someone uses the product, or the user's own word. Only some of the people in this fight say that their answer has a price.

Two findings fall mostly on one side. The first, stating a disputed cause of death as fact, falls mostly on the safety side and in neutral reporting. Four sources do it: the Hawley release; a release from state Sen. Steve Padilla, the author of SB 1119 and a different person from Sen. Alex Padilla; a statement by Assemblymember Rebecca Bauer-Kahan, the author of the vetoed bill, quoted in the San Francisco Standard; and the CalMatters report, which does it through the word "after." The R Street commentary comes close from the other side, when it describes "sobering cases in which troubled youth interacting with AI platforms like ChatGPT have found these chatbots all too willing to reinforce negative inputs, even encouraging suicide." OpenAI's denial is the matching claim from the other direction. The liberty side mostly avoids the deaths altogether. As FIRE's essay shows, that is its own way of leaving a value unweighed.

The second, misdescribing the bill, falls on the liberty side. R Street, FIRE and others wrote around the time of the vote and described the original text. That is a matter of timing. The Reason Foundation, a libertarian think tank, has no such excuse. Its backgrounder is dated May 19, eight days after the amended text limited verification to AI companions. It says the bill "would impose mandatory identity verification on every American chatbot user."

The tools, used and abused

The tools on Hardly Thinking are questions a reader can run on any argument, including their own. Each has a fair use and a weaponized one.

B3: The Trade-off Surfacing asks: compared to what, and at the cost of what? Veronica Irwin used it fairly when she wrote that an online age check, unlike a bouncer, keeps a record. Newsom used it in his 2025 veto letter. It goes missing in the Hawley press release and in Ryan McBain's STAT op-ed. The release leaves verification out, and the op-ed leaves adults out. Its weaponized form, economism, treats every firm commitment as an error simply because it has a cost. No clear case of it turned up in this fight.

A2: The Denominator Demand asks of any number: out of how many, and answering which question? OpenAI's statement supporting California's law, quoted by Rachael Myrow of KQED, said that "nearly nine in ten teens who use ChatGPT" turn to it for learning and similar uses in a given week. The figure counts only teenagers who already use the product and leaves out how many use it for companionship. A2's weaponized form, denominator-flooding, drowns a real harm in a huge total until it seems to vanish. R Street's "small numbers" is a quieter version: the harm is called small, and no number is given at all.

A4: The Mirror Check asks: did I scrutinize this claim, or did I just like its conclusion? J. B. Branch's figure in his Issues in Science and Technology essay and the R Street Institute's reading of the Pew poll form a mirror pair. One made a risk look larger than the survey found. The other made a benefit look larger and nearer than the poll found. If you caught one of the two more quickly than the other, you have just run the Mirror Check on yourself.

For many people, the weighting of freedom and safety flips with the topic: liberty for what they want to do, safety from what they don't. So take the rule about AI companions you favor and run B3 on it first. To make this rule work, who has to prove their age, how, and what does it cost the people it is not meant to protect? Then ask whether you would accept that answer on an issue where you usually take the other side.

Sen. Alex Padilla, a California Democrat on the Judiciary Committee, did something like this on April 30. He voted to advance the bill banning minors from AI companions and, the same day, raised concerns about the age checks' "potential privacy and security risks." Few people in this fight have weighed both values out loud, and few of the rest of us do either.

Reader review

Every HT article ends by asking readers to check it against its own rules. The question is not what you think of AI companions but whether this article met the standards it applies to others.

1. Find the passage where this article was hardest on the side you agree with, and the one where it was softest on the side you oppose. Were they judged by the same standard?
2. The article runs A2, the Denominator Demand, on OpenAI's "nine in ten" and R Street's "small numbers." Check one number in this article against its linked source. Out of how many, and does the article say anything stronger than the source?
3. Run B3, the Trade-off Surfacing, on the article itself. It counts what age checks would cost adults in some detail. Does it count what no rule would cost teenagers as carefully?
4. Of the five passages examined at length, three come from people who want more restriction on companies and two from people who want less. Is that how the fight fell, or a lean?
5. The lawsuits and the companies got less room than the legislators and advocates. Pick one. What might the tools have found there?