Anthropic Suspends Claude Accounts Suspected to Be Minors

There's only one way to unlock a suspended account: go through Yoti, an outside age-verification vendor. A notification email delivers a verification link, and passing it restores the account. There are three routes — take a selfie for facial age estimation, upload a photo of an ID like a passport or driver's license, or, if you already use the Yoti app, share the over-18 attribute it already has stored.

Anthropic draws a clear line around not holding this personal data itself. The documentation says Yoti deletes the selfie, ID image, and any other personal information as soon as age is verified. Anthropic states it never sees the ID or photo at all — it only receives a pass or fail result — and that it neither processes nor stores any personal data generated during verification. The document also notes that Yoti undergoes independent SOC 2 audits.

Structurally, it resembles the ID check at a bar door. A bouncer eyeballs your face to guess your age and asks for ID when it's not obvious. The difference here is that the bouncer isn't an employee of the bar but an outside contractor, and once you're through, the bar claims it never even saw your face.

This support document was posted verbatim to Hacker News on September 11 and racked up 666 points and 645 comments in a single day. But the discussion didn't center on protecting minors — it centered on where the face and ID data actually goes. Commenters mixed together suspicion that age verification was a pretext for collecting identity data, discomfort with handing biometric information to an unfamiliar outside vendor, and memories of real large-scale breaches at identity-verification companies. Many also pointed out that this adds to a growing pattern of having to pass an ID checkpoint just to use the internet.

The policy itself isn't new this week. Anthropic disclosed that Claude is restricted to users 18 and older in a December 18, 2025 announcement about user wellbeing. It explained that if someone identifies themselves as a minor during a conversation, a classifier flags the account for review and suspends it if confirmed, and said it's building additional classifiers to catch subtler conversational signals of minors. Anthropic also said it joined a family online-safety organization to help improve industry-wide protections for minors. At the time, the company's stated reasoning was that younger users face greater risk of being negatively affected by conversations with chatbots.

That same announcement bundled other user-protection measures alongside the age restriction. A classifier that catches suicide and self-harm signals surfaces a banner pointing to professional counseling resources when it detects risky content, and a partnership with ThroughLine connects users to crisis lines in more than 170 countries. Anthropic also said it dialed down how much its models flatter users — the 4.5 series is 70 to 85 percent less sycophantic than the earlier Opus 4.1, and it disclosed that its models gave appropriate responses in 98.6 percent of high-risk suicide and self-harm evaluations, dropping to 91 percent under real-conversation stress testing. The age restriction was one piece of that bundle — evidence that Anthropic treats keeping out minors and protecting vulnerable users as the same problem.

So what's new this week isn't the policy — it's people actually feeling its effects. Only once the classifiers were running, accounts were actually getting locked, and unlocking one meant handing over a face or an ID did people go dig up the documentation and read it. That's the size of the gap between a rule being written down and a rule actually landing on someone's account. Users typically learn about a policy not on the day it's announced, but on the day it catches them.

The genuinely hard problem is accuracy. Estimating age from a face or guessing it from conversation patterns is inherently probabilistic. A 17-year-old might read as 20, and a 23-year-old might read as 16. The first kind of error means the safeguard fails; the second means a legitimate user gets locked out of a tool they were working in — and wherever a classifier sets its threshold, one type of error necessarily goes up. Anthropic hasn't disclosed where it has struck that balance.

For anyone using Claude for work, the takeaway is clear. Account lockouts can arrive without warning, and recovery runs through an outside vendor's window, not Anthropic's own. If your work history and prompts live entirely inside one service, your work stops until that window opens — not because you can't prove your age, but because proving it takes time.

Anthropic chose filtering over growth, and adult users are the ones now covering that cost with selfies and IDs. As more companies narrow the door in the name of safety, proving your age becomes a routine part of using the internet. What's happening with Claude is an early look at what that routine actually looks like in practice.