Why Do ChatGPT, Claude and Codex Accounts Get Blocked? 20 Factors Worth Knowing
Why do Anthropic and OpenAI block accounts even of those who have used them for years, legally and paying, and by what signs do the services determine who you are and where you are?
The box "don't train on my data" does not turn off analysis: companies keep access to chats, and requests with ill intent go to the security service and may reach the police. Years of use and payment do not protect against deactivation — accounts are blocked over a proxy, an Iranian phone number, a flag after a trip. Anthropic's rules that took effect on November 12, 2026 extend the restrictions to personal accounts for corporate needs, to ownership and control, the headquarters and employees abroad; the company explains them by data transfer to the state, military use, distillation and US interests. Alexander Volchek's twenty factors — from IP address and time zone to device fingerprint, payment, account data and request content — show that with today's use of AI your location cannot truly be hidden, and he considers workarounds through access resellers unworkable.
What to watch for
Key takeaways
The box "don't train on my data" only means the model is not trained on it: the corporation keeps access to the data and says so openly. A request with ill intent immediately goes for analysis by separate models, then to a person in the security service, who can pass it further — up to the police; there are known cases of OpenAI handing data to the police.
A user from Russia used ChatGPT for years, connecting through various VPNs — the long-standing account was deactivated and the appeal rejected. Alexander Volchek concludes that neither long use nor payment prevents a later deactivation.
At the beginning of October, a user from Iran had an account that worked on other devices, but the Windows app did not accept an Iranian number with the +98 code, and support replied that this country cannot be used. The restriction was triggered by an entirely different criterion.
On September 4, 2025, Anthropic announced restrictions for foreign companies under the control of unsupported jurisdictions and mentioned ownership of more than 50%. On May 12, 2026, the company added details: the user's location, the headquarters, ownership of the company, employees abroad.
A person went to another country on a business trip or to see a grandmother and came back, but a flag remained in the system — and can later have a negative effect on the whole account history, however he explains the trip.
On October 8, 2026, Anthropic announced an update to its general set of rules, which took effect on November 12; the company called the regional section a clarification of how the existing policy is applied. According to Alexander, a wave of blocks began after that.
Anthropic describes the restrictions in terms of business, but a personal account set up for corporate needs is also subject to blocking. A Belarusian company's developer moving to Serbia does not lift the restriction, foreign registration does not cancel the restriction on ownership, and a headquarters in Russia is an independent ground.
Alexander considers the opinion that a Belarusian, Russian or Iranian passport leads to a restriction in any case to be very relative: for his friend with a Belarusian passport and residency in another country, the check at Anthropic went fine, although he did not send the Belarusian passport itself.
On June 12, the US government demanded that access by foreign nationals to Fable 5 and Mythos 5 be restricted — including those in the US. This would have covered even Anthropic employees with a foreign passport and a green card. Fable 5.1 and 5.2 are available to a wide range of people; Mythos remains restricted.
The main threat named is training competing models on Claude's answers: according to Alexander, millions of accounts are created for distillation, and in China API access to Anthropic is sold five times cheaper through a middle layer that takes the data. Without distillation, he believes, Chinese models would most likely not exist at all.
Restrictions have already begun: a select few got Mythos, and there are closed systems for biology, physics and mathematics and systems only for states. Alexander is sure that Anthropic and OpenAI will at some point verify the identity of all users; what holds them back is that otherwise the companies would lose the world's strongest developers.
The first two of the twenty factors are the country of connection by IP address, which Anthropic directly confirms, and whether the browser's time zone matches the presumed location. Alexander considers them probably the most important: the whole history of changes to this data is stored.
The Russian language by itself does not prove someone is in Russia, but as Alexander perceives it, Anthropic links him in particular very strongly with Russia — a citizen of Belarus who works a lot in Russian. For the system, he believes, Russian automatically means the Russian market.
The card's country of issue, the billing address and the country of connection are checked by payment protection, and failed payments and chargebacks are taken into account. Where Anthropic is blocked, people pay through third-party services without understanding how the payment goes through and what the company will see in the end.
ChatGPT and Claude Code are used so widely — questions about the nearest café, a charger, the school, the children's emails, Codex's access to the computer and home server — that, according to Alexander, it is truly impossible to hide your location.
Alexander does not support attempts to get around the services through middle layers reselling access to the Claude API or ChatGPT API and considers them unworkable: Anthropic obliges resellers to keep track of location. Salesforce, for example, temporarily switches off the Claude API when a user moves to another country and offers something else.
What this episode is about
A ToTheMoon solo episode: Alexander Volchek explains why Anthropic and OpenAI block Claude, Claude Code and ChatGPT accounts — including those of people who have used them for two or three years. He makes a point up front: the episode was not made to help anyone get around these systems, but for people who use AI legally yet do not fully understand what data the system analyzes.
He begins with the box "don't train on my data": the model is not trained on it, but the corporation keeps access to the data. Requests are selectively analyzed, and a request with ill intent goes to separate models, then to a person in the security service, and may reach the police — there are known cases of OpenAI handing data to the police.
Then come real cases: a block in January 2024 over access from an unsupported country and a proxy, a long-standing Russian ChatGPT account deactivated with the appeal rejected, an Iranian number with the +98 code that the Windows client did not accept, a closed Belarusian account. Alexander's conclusion: neither long use nor payment protects against deactivation.
The main part of the first half is Anthropic's regional policy: restrictions for companies under the control of unsupported jurisdictions from September 4, 2025, details on May 12, 2026, and the update to the general set of rules announced on October 8 and in force from November 12, 2026. Alexander goes through examples from the published terms — a business trip to Russia, a developer who moved from Belarus to Serbia, a Kazakh company with a Russian parent, a headquarters in Russia — and asks how the company is going to determine all this at all.
He speaks separately about passports — in his view a passport by itself is not a verdict — about his own case with YouTube, where a restriction was lifted only after a month-by-month description of his residence, and about the US government's June 12 demand to restrict access by foreign nationals to Fable 5 and Mythos 5, including those in the US.
Alexander lists the reasons for Anthropic's policy following the company's explanation: forced transfer of data to the state, military and intelligence use through subsidiaries, distillation — training competing models on Claude's answers, for which, in his words, millions of accounts are created — and protecting the strategic interests of the US and its allies. He is sure that Anthropic and OpenAI will eventually verify the identity of all users, and explains what holds them back.
The second half covers the twenty factors: country of connection and time zone, connection type and address reputation, movement history, language and regional settings, device fingerprint and links between accounts, contradictions in technical data, automation, payment geography and history, account data, identity verification, the company and its owners, geolocation, WebRTC and request content. The examples come from his own experience: the besolid.com website, which itself notices account sharing, connecting through Xfinity and Starlink in Mendocino, the link between the Russian language and Russia.
In closing, Alexander explains why, with today's use of ChatGPT and Claude Code, your location truly cannot be hidden, and says that he does not support attempts to get around the services through API access resellers and considers them unworkable: Anthropic obliges such intermediaries to keep track of location, and Salesforce temporarily switches off the Claude API when a user moves.
The episode comes down to one idea: an account block is not a reaction to a single act but the system's conclusion from many signals, most of which the user never thinks about. Anthropic's published rules concern countries, ownership and control, but compliance with them is decided by the IP address, time zone, device, payment, account data and request content. The views on future identity verification of all users, on the role of distillation and on US policy after January are Alexander Volchek's assessments, not announced company plans; the services' own official terms are what is worth checking.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 67 segments: 67 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…