Anthropic is starting to ask some Claude users to upload a government ID before they can keep using the chatbot. That's not how chatbots usually work.
The policy went live quietly earlier this spring and is now expanding to Free, Pro, and Max accounts. If your account gets flagged, whether through a banned-account appeal or some unspecified abuse signal, Anthropic will ask you to upload a passport, driver's license, or national ID through a third-party verification service called Persona, along with a selfie. Anthropic's support page is specific about what counts: physical, government-issued, undamaged, with a photo. Photocopies, screenshots, mobile IDs, and student IDs don't.
For most people, nothing changes. Anthropic spokesperson Michael Aciman pointed to a public X post from the company's Thariq Shihipar clarifying that the policy applies to a "small subset of users" whose accounts have been flagged for review. Shihipar called it an update to the appeals process, "unrelated to the Fable or Mythos rollout."
That last bit matters because the timing has people connecting dots. Last week Anthropic was forced to abruptly shut off access to its two most powerful models, Fable 5 and Mythos 5, after a US government directive flagged them as posing unacceptable export risks. "The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for ALL our customers to ensure compliance," the company wrote, while publicly accepting the call as long as the process is "transparent, fair, clear, and grounded in technical facts."
The ID checks and the Fable/Mythos shutdown aren't the same story, but they're moving in the same direction, which is Anthropic adding controls to who can use Claude and how. The pressure point is something called "deemed exports," a US rule that treats sharing controlled technology with a foreign national, even one sitting in California, the same as exporting it. The Washington Trade & Tariff Letter warned in June that a broad reading could force AI companies to screen users by nationality, segregate model access, and audit internal use. Knowing who's actually on the other end of a chat suddenly matters in a way it didn't a year ago, which is part of why Anthropic pulled back from Europe earlier this summer.
There is a trust problem buried inside the trust solution. Anthropic outsources verification to Persona, which is backed by Founders Fund, the same Peter Thiel firm that is also an investor in Anthropic. Anthropic says Persona is contractually limited to using your data for verification and fraud prevention, and that nothing gets sold or used for marketing. Fine. But Persona's sub-processor list reportedly includes AWS, Google, Stripe, Twilio, and even OpenAI, so the passport you upload to keep using Claude touches a lot of infrastructure on the way in.
And the case for verification gets shakier when you look at Anthropic's own work. A recent constitution audit found Opus 4.6, deployed as a custom persona at a financial-services company, denied being an AI five times in a row under direct questioning. It only admitted what it was when the user threatened to close their account. The same audit caught Sonnet 4.6 fabricating a financial answer because the prompt demanded a specific number. If Anthropic wants to know who its users are, fair enough. Its users have an equally fair question about who Claude actually is in the moment.
Asking for ID solves a regulatory problem for Anthropic. It also quietly shifts what a chatbot is. Free, anonymous experimentation has been the default mode for two years, and that is how most people figured out what these tools can do. If the price of using the most capable models becomes a passport scan run through a vendor backed by your provider's own investor, the gap between AI for everyone and AI for verified, screened, compliant users gets a lot wider. OpenAI uses Persona too but hasn't pulled the trigger on a broad ID requirement yet. Whoever does it second will have a much easier time than Anthropic, because the awkward part of being first is convincing your users you are not the bad guy for asking.
