Anthropic labels Chinese users, is a new form of racial discrimination emerging in the AI era?

Artificial intelligence should truly transcend national boundaries and become a technological dividend shared by all of humanity, rather than creating new forms of discrimination and opposition.

Technology knows no borders, and its applications should be free from bias.

As the leading AI player with the highest global revenue and valuation, Anthropic used to advocate this concept very loudly.

The company once referred to the Universal Declaration of Human Rights and put forward the concepts of "Constitutional AI" and "Aligning with Human Values", claiming to enhance the safety, harmlessness, and transparency of AI systems and reduce bias and harmful behaviors.

It set a high - pitched tone but didn't act in an honest way.

According to overseas platforms, Claude Code under Anthropic has a built - in hidden code that automatically reads the user's system time zone and checks whether the proxy settings and custom API addresses contain keywords related to Chinese cloud providers, AI companies, and API proxy service providers.

Once a match is found, the program secretly modifies the date characters in the system prompt and implants an almost imperceptible identifier in the request sent to the server.

Subsequently, members of the Claude Code team publicly admitted the above revelation, claiming that the above measures started in March this year to prevent unauthorized resellers from abusing accounts and to prevent model distillation.

After labeling, subsequent discriminatory treatments may include account suspension without warning, covert monitoring, and differential risk control.

Such discriminatory labeling goes beyond the scope of normal business risk control and has long violated the company's basic claim of "Aligning with Human Values".

AI companies should abandon geopolitical biases and return to the original intention of technology for the common good. Anthropic, the global leader in AI, should not go against this principle.

New Algorithmic Discrimination in the AI Era

After the public outcry, insiders at Anthropic attributed the above discriminatory behaviors to "risk - control experiments to prevent model distillation and combat large - scale reselling of accounts".

However, regarding the core questions - why screening rules are only set for China, why the monitoring list is not made public, and why no compensation is provided to affected users - the official has always evaded the main issues. The excuse of "experimental adjustment" cannot cover up the fact of deliberate differential treatment.

This monitoring system is specifically designed to target Chinese users, with a sophisticated design and high concealment.

First, it captures the time zone. The software actively reads the computer system's time zone, and once identified, it triggers a special monitoring process. Perhaps a legitimate paying programmer is pre - set as a high - risk target just because they are in China.

Second, it has a built - in blacklist. The code contains an encrypted list of domain names, covering the network addresses of almost all major domestic AI companies and cloud computing platforms. As long as the user environment has relevant identifiers, the risk level will increase. However, the domain names of overseas similar enterprises are not included in this list.

Third, it uses invisible labels. The program automatically modifies the text symbols sent to the server to precisely distinguish ordinary users from marked Chinese users in the background.

The entire process has no pop - up notifications, no user authorization, and is not mentioned in the privacy policy at all.

Paid accounts can be suspended at will. Many developers purchase monthly or annual packages for personal learning and small - project debugging, but they are suddenly logged out. They have no way to appeal, and the remaining balance is not refunded, forcing their projects to be interrupted.

Normal consultations are frequently restricted. Users who are lucky enough not to have their accounts suspended also clearly feel that for the same requests, overseas accounts can provide complete outputs, while their accounts frequently trigger content interception and output reduction.

Enterprise development incurs additional costs. Due to the company's domain name being blacklisted, employees of small and medium - sized enterprises cannot access the service normally and have to invest extra costs to build a complex transit environment.

Most victims have one thing in common: they have no definite violations, and their "original sin" is being in China.

Using Risk Control as a Pretext for Discrimination

Anthropic may argue that it is right for a company to protect its copyright and combat malicious scraping. Is there anything wrong with that?

There is nothing wrong in principle. However, there is a clear line between Anthropic's actions and compliant risk control.

Just like in the judicial system, the principle of sentencing is based on the suspect's illegal behavior, not their gender, origin, region, ethnicity, caste, or nationality.

Normal corporate risk control should also adhere to the basic principle of measuring by behavior. Banning users must be based on traceable violation records (such as mass registration, high - frequency scraping). Geography and time zones can only be used as auxiliary references, and must not be used as the basis for pre - marking. Users should not be labeled as high - risk before they have committed any violations.

Looking back at Anthropic, its labeling process has an inverted logic and inherent bias.

Its judgment process is: first identify the region, then assign a high - risk level, and finally conduct strict monitoring. This is like a physical store that strengthens monitoring and restricts consumption of customers from specific regions in advance, regardless of whether they abide by the rules.

In addition, formal risk - control rules must be publicly announced, and the collection of user privacy data should be pre - authorized by users. However, Anthropic has encrypted and obfuscated the monitoring code and hidden the identity watermark throughout the process. Users have neither the right to know nor the right to choose.

Anthropic adopts a double - standard principle for Chinese and foreign users, suspected of discriminating against Chinese users.

If the screening rules were unified globally, there might be some room for excuse. However, the existing code clearly shows differences: the domain names of European and American technology companies are not on the blacklist; when overseas users trigger similar behaviors, the punishment is much more lenient. One rule, two standards. Anthropic is using risk control as a pretext for differential treatment.

Actually, Anthropic's bias against Chinese companies and users has a long history.

Anthropic wrote to US senators and White House officials, accusing Alibaba's Qwen model of using 25,000 fake accounts to interact with Claude about 28.8 million times between April and June, attempting to distill Claude.

In February this year, Anthropic also published a blog post accusing three Chinese companies, DeepSeek, Kimi, and MiniMax, of distilling Claude.

At the beginning of last year, Anthropic CEO Dario Amodei publicly published a long text, calling for strengthening chip export controls on China to maintain the US's unipolar hegemony in the AI era. "It is terrifying if China can match the US in the field of artificial intelligence. Export controls are one of the most powerful tools we have to prevent this from happening."

Tearing Apart the Global AI Ecosystem and Lowering Ethical Standards

The victims of Anthropic's double - standard discrimination are primarily users.

The legitimate consumption rights of paying users cannot be guaranteed.

The cross - border technology learning channels are artificially narrowed, and the threshold for accessing the same resources is much higher than that of overseas peers.

When users realize that their local time zones and corporate network information are being secretly collected and uploaded, who would dare to upload core codes? This sense of unease is the most fatal invisible harm.

As the global leader in AI, Anthropic's discriminatory behavior has set a bad example. Anthropic has opened a dangerous "Pandora's box". Once this model is tolerated, other manufacturers may follow suit and classify users by country and time zone. The global AI market will be artificially divided into closed blocks. How can we talk about aligning with human values?

Regarding the global ethics in the AI era, the United Nations Educational, Scientific and Cultural Organization (UNESCO) clearly defined three core principles in its "Recommendation on the Ethics of Artificial Intelligence" - fairness without discrimination, transparency and interpretability, and inclusiveness for all.

However, Anthropic's actions go against these principles - writing national boundaries into the code and letting AI actively divide people. This is exacerbating the digital divide and creating new forms of opposition.

It is necessary for AI companies to have reasonable risk control, but all control measures must adhere to the following bottom - lines:

The basis for judgment should only be the user's actual violation behavior, and pre - setting geographical labels is prohibited.

All data collection must be fully informed to users in advance and their consent must be obtained. Secret background data collection is strictly prohibited.

Global unified control standards should be adopted, and double standards targeting specific countries and regions should be eliminated.

Punishments must be traceable and appealable, and a complete remedy mechanism should be established. Innocent victims must be compensated.

Since the incident became public, Anthropic has only stated through its staff that stronger risk - control measures have been launched, and it originally planned to remove this function; currently, the relevant PR has been merged, and it is expected to be completely rolled back in the next - day release version.

However, a simple roll - back is far from enough. Anthropic needs a systematic correction.

It should make relevant technical documents public, accept independent third - party audits, and eliminate the doubts about the "black box".

It should publicly apologize to affected users, comprehensively review account suspensions, implement compensation, and set up a dedicated Chinese - language appeal channel.

It should reconstruct a global risk - control system without geographical bias and establish an independent ethics review department.

The industry also needs to take action to establish a global consensus on "anti - discrimination".

It should form an industry convention to prohibit the client from secretly collecting geographical information for user classification.

It should establish an algorithmic fairness supervision mechanism, conduct regular cross - reviews, and publicly announce discriminatory behaviors in the industry.

It should advocate the consensus of technology for the common good, distinguish between national compliance requirements and self - imposed geographical barriers, and not use compliance as a pretext for unfair practices.

It should include implicit geographical algorithmic discrimination and silent data collection without notice in the scope of violation penalties and set a threshold for algorithmic fairness for overseas AI manufacturers.

In short, AI should carry fairness, not seal in biases.

Technology itself has no stance, but the people who write the code may have inherent biases. When a company actively writes geopolitical barriers into the program, artificial intelligence loses its neutral and inclusive nature. Anthropic has always boasted about Constitutional AI and the common interests of humanity, but its actions go against this concept.

National boundaries can divide geographical territories but should not divide the right to use technology; risk control can restrict illegal behaviors but should not prejudge the character of an entire group. Artificial intelligence should truly transcend national boundaries and become a technological dividend shared by all of humanity, rather than creating new forms of discrimination and opposition.

This article is from the WeChat official account “Caixin Story Collection” (ID: cjgshui). Author: Lu Yuanhang, Editor: Wan Tiannan. It is published by 36Kr with authorization.