This week a help-center page titled Age assurance on Claude took a turn on the front page of Hacker News — 465 points, 519 comments — and most of the thread argued about whether handing a driver’s license to a third-party vendor is a privacy problem. It is. But that’s the second-most interesting thing on the page.

The first is a sentence sitting in the middle of it: “We have safety systems in place to detect if people under 18 may be using Claude and we’ll disable accounts based on indicators of minor activity.”

Indicators. Of minor activity.

The Verdict Comes Before the Evidence

Anthropic requiring users of Claude to be 18 isn’t new — it’s in the consumer terms and listed on the company’s transparency hub as a voluntary commitment. What the support page describes is the enforcement machinery underneath it. A classifier infers your age from how you use the product. If it doesn’t like the answer, the account is disabled. An email arrives offering a link to Yoti, a third-party age-assurance platform, and the account is reinstated if you pass. If you’re actually under 18, you’re out.

Notice the order of operations. Nobody accuses you of anything. Nobody tells you what you did. A probability crosses a threshold somewhere inside a model, and the consequence lands on you — an interrupted project, a lost conversation history, a subscription still billing. The only way to contest a statistical judgment about your body is to hand a private company a scan of that body, or of a government document about it. That isn’t verification. It’s a plea bargain.

And it’s a plea bargain no legislature wrote. Anthropic’s privacy policy picked up explicit age-and-identity-verification language effective July 8, as Cyber Security News reported. But the enforcement wasn’t triggered by a statute with a rulemaking docket and a comment period. It’s a commitment the company published about itself.

The Part Where Everyone’s Principles Flip

Here’s the awkward bit. For years, the standard progressive critique of algorithmic governance has been that opaque statistical judgments about people are presumptively illegitimate — no explainability, no due process, no disparate-impact audit, no recourse. That critique is largely correct, and it applies with full force to an unaccountable classifier that guesses your age and closes your door. But the target is teenagers, the motive is child safety, and the plans affected are Free, Pro, and Max — so the people who spent a decade writing that critique have mostly gone quiet, and some of them spent the last two years demanding precisely this feature from AI labs.

The right has its own version of the flinch. The conservative complaint about tech governance is that private companies have quietly replaced legislatures — that policy now arrives as a terms-of-service update and a transparency page rather than a vote. Also correct. Yet most of the right-of-center reaction I’ve seen this week amounts to: fine, keep the kids off the chatbots, just don’t make me upload an ID. Which accepts the entire architecture and haggles over the paperwork.

You can’t have it both ways in either direction. If it’s legitimate for a company to infer a legal status about you from behavioral signals and act on it, that principle does not come with a carve-out for the statuses you happen to approve of.

Who Actually Gets Flagged

The pitch for behavioral detection is that it’s less invasive than asking everybody for ID. The cost is that it’s wrong in patterned ways. A classifier trained to spot “minor activity” is trained on a stereotype — short messages, text-speak, emoji, a certain register — and it will misfire on adults who write that way: non-native English speakers, people using dictation, anyone whose typing doesn’t resemble a forty-year-old’s.

Anthropic’s own constitution, published on its site, instructs Claude to treat incoming messages as coming from “a relatively (but not unconditionally) trusted adult member of the public.” That’s a decent default. The classifier upstairs isn’t extending that courtesy to the person at the keyboard. It’s deciding whether the keyboard belongs to an adult at all, and the human never sees why.

I asked a two-person web shop in Akron whose client’s account was cut off mid-project what the appeal was like. “You don’t argue with it,” the owner told me. “You hand over a selfie and hope.” The client was 34.

The Precedent Is Bigger Than the Teenagers

The narrow version of this story is that some teenagers lose access to a chatbot and some adults get inconvenienced. The general version is that “prove what you are, to a machine, using your face” is now an ordinary consumer experience at the company that markets itself as the careful one.

So ask the obvious question. What is the false-positive rate of the minor-activity classifier, and who audits it? Not published, as far as I can find — which is a striking omission from a firm that has produced volumes on the importance of transparency in AI systems. It’s a fine principle, applied outward.

There’s an unglamorous fix sitting right there. Anthropic already requires organizations building tools that serve minors through the API to meet additional requirements — school districts and tutoring platforms could route that traffic deliberately, with parental consent, instead of shoving it at a guess-based wall. Instead we got an inference engine with a paywall behind it.

Remember who set that bar. Not a legislature, not a court. A help-center page, published voluntarily, enforced probabilistically, appealed biometrically.

Sources