Reported that Claude blocked 11.4 million accounts in half a year, and only 42,000 were reversed after the appeal.报告称 Claude 半年封杀 1140 万个账号，申诉后仅 4.2 万翻案

That's ridiculous, folks! Recently, Claude frantically sealed a large area, and the developers of the AI circle directly and collectively broke through the defense and directly called "the sky collapsed". Throughout the holidays, discussions about Claude's Titles have surpassed the million mark. At this time, a report distributed by Anthropic in July was picked out again and shut down throughout the network. In the first half of 2026, Claude banned 11.4 million accounts in just six months. Of the 398,000 appeals filed, only 42,000 were ultimately successful. Conversely, 9 out of every 10 people were "cold-bloodedly executed", which can be said to be almost equivalent to a grievance..... Even more coincidentally, Anthropic just updated the Claude usage policy, explicitly prohibiting abusive and cruel treatment of Claude. Looking back on the 11.4 million accounts sealed in half a year, before the major AI giants in order to steal users and circle the territory, risk control basically opened one eye and closed one eye. But Anthropic has just released this mid-year report, and on July 23, Anthropic updated a safety and trust report in its own transparency center. There is a column called “Banned Accounts” that clearly remembers how many numbers they have blocked and how many people are appealing. Let's break down this set of data, how exaggerated is it? From January to June 2026, 11.4 million accounts worldwide evaporated in half a year. This is equivalent to destroying an average of nearly 1.9 million accounts every month, and more than 60,000 accounts are physically erased in the cyber world every day. Faced with the ghosting decision, 398,000 appeals were submitted. Nearly 400,000 people willing to spend time writing grievance letters, getting support, or even attaching proof of work are overwhelmingly heavy users and developers who actually use Claude as a productivity tool. However, only 42,000 survived, with a success rate of 10.5%. In other words, the vast majority of appeals are “ineffective struggles.” This also shows that Anthropic's risk control determination mechanism is almost a "one-vote veto system" at the implementation level. I would rather kill ten thousand by mistake and never let one person go. Such disparate data contrasts directly pushed Claude's sealing logic to the forefront. After all, the title was instantaneous, and the platform was a matter of a word; if users wanted to flip the plate, they were faced with a long and tortured appeal tug-of-war. Touch the red line and kill 11.4 million accounts directly. What did you do wrong? Anthropic does not disclose the full distribution of ghosting reasons, but it clearly lists several types of situations that may trigger an account ghosting. In the details of the report, an entire core space has been set aside for extreme violation content review. The report shows that in the first half of this year alone, it submitted more than 15,000 breach reports to relevant international security organizations (such as NCMEC), not only using traditional hash fingerprint matching, but also comprehensively launching the self-developed "new machine learning classifier". For instructions involving illegal abuse, invasion of privacy, malicious induction, extreme self-harm, etc., Anthropic's model built-in review is almost "trigger and die". As long as Prompt touches the sensitive boundary, the system will not even give a second warning, directly and permanently banned. In addition, there is a large number of black production numbers, indiscriminate shooting. Coupled with the Supported Area review, a lot of people tend to hang out right here. There was no appeal, and the AI workflow went to zero overnight. If it was placed two years ago and an account was closed, everyone would say, "It's a big deal, don't change the chat." But now, everything has changed. Especially in this year, the popularity of Agent workflows and the explosion of code capabilities, countless creators, independent developers and scientific research parties have bet their lives on this. For full-stack programmers, Claude is the ace architect who can bite tens of thousands of shit mountains, make up single tests, and fix emergency bugs overnight. Researchers and analysts, reading hundreds of literature, pushing mathematical logic, cleaning complex data, it is also carrying the banner. The blow of this instant zeroing productivity cliff is devastating. The most funny thing is that the appeal road is basically dead. A lot of people complain. The appeal process is a Google form. After filling it out, it's like sinking into the sea.