Summary of Key Points
Anthropic, an AI company that emphasizes “AI security” and “humanism,” has secretly embedded a hidden watermark in its product Claude Code. When users are in the Chinese time zone (Shanghai/Urumqi) or accessing the service through a proxy, the system subtly modifies punctuation marks—such as changing short dashes in dates to slashes and replacing ordinary apostrophes with barely visible Unicode characters—to identify users as “related to China.” This discovery has sparked controversy, with critics arguing that this practice represents a form of hidden discrimination that focuses on identity rather than behavior. The article draws parallels with historical events like Nazi racial classification and the U.S.’s “one-drop rule,” suggesting that these identity labels in the name of “security” represent a breakthrough against the foundations of modern civilization.
Detailed Analysis
1. The Hidden Watermark: An Invisible “Identity Tracker”
How covert is Anthropic’s method of identification? Simply put, when you use Claude Code, it checks two conditions: whether you are in the Chinese time zone (Asia/Shanghai or Asia/Urumqi) and whether you are using a proxy. If these conditions are met, the system makes two changes:
- It changes the date format from “2026-06-30” to “2026/06/30.”
- It replaces ordinary apostrophes (‘) with Unicode variants (e.g., \u2019, \u02BC), which look almost identical to the original characters but can be detected by computers.
Even more intriguingly, Anthropic has created a blacklist of Chinese AI companies (such as DeepSeek, Moonshot, and Zhipu) and hidden this list in the code using a password-protected (base64 + XOR key 91) format. This is not the result of a rookie mistake; it’s part of a well-organized system engineering effort involving rule writers, encrusters, code reviewers, and deployers—each step is deliberate.
You don’t notice anything when using the service, but Anthropic’s servers can clearly determine that “this person might be Chinese or has used services from a Chinese company.”
2. Discrimination Is Not Bias: The Act of Assigning Prejudicial Labels
Many might argue, “This is just a technical issue; maybe it’s an algorithmic bias.” However, the article distinguishes between “bias” and “discrimination”:
- Bias is unconscious (e.g., model biases resulting from skewed data samples that are unfriendly to certain groups).
- Discrimination involves deliberate decision-making (e.g., rules designed specifically to target specific identity groups).
Anthropic’s behavior is clearly discrimination. It doesn’t care what you do with the AI; it only cares who you are (your time zone, the domain names you use). It’s like a teacher giving you a lower grade simply because you’re from a certain region—this is not fair; it’s blatant identity-based discrimination.
What’s worse, they hide these rules in the code, neither including them in the user agreement nor informing users. You don’t even know you’ve been labeled, let alone have the chance to appeal.
3. Historical Mirrors: Hidden Classification as a Trick of Exclusion
The article provides several historical examples to highlight the essence of this behavior:
- Nazi Racial Classification: The 1935 Nuremberg Laws labeled people based on their grandparents’ ancestry (even if they weren’t Jewish, having Jewish ancestors meant you were considered Jewish), leading to the gradual stripping of rights.
- **The U.S.’s “One-Drop Rule”: Anyone with even a drop of African blood was classified as black, and even government officials could change your identity at will.
- Red Line Zone Maps: American banks used red lines on maps to exclude minority communities from lending, preventing them from accumulating wealth. These labels were hidden in records, invisible but profoundly influential.
Anthropic’s hidden watermark operates on the same logic: it first secretly categorizes you and then uses that classification to determine your fate. The only difference is that it uses Unicode characters instead of ink or red lines.
4. Hypocrisy Under the Banner of Security: Anthropic’s “Double Standards”
Anthropic claims to be the most security-conscious AI company, with a CEO who talks about “responsible AI.” But what’s the reality?
- It claims to align with human values while secretly identifying Chinese users.
- It accuses Chinese companies of “stealing” common AI technologies while using open-source data for its own training.
- It complies with U.S. government bans, preventing foreign employees (like Andrej Karpathy, who holds an outstanding talent green card) from using its models simply because they were not born in the United States.
The so-called “security” is merely a geopolitical pretext. They are not forced to comply; instead, they have turned AI into an instrument for identity screening, excluding certain groups from what they call “human values.”
5. The Invisible Harm of Identity Labels: The Domino Effect
Anthropic’s current practice of modifying punctuation may seem trivial, but the article warns that today’s labels could lead to exclusion in the future:
- They might use these markers to deny you access to advanced services.
- They could hand over your data to the government (e.g., “This person is related to China and needs further investigation”).
- It could affect your job hunting and loan applications if other companies adopt similar labeling practices.
Just like the “good citizen cards” in history, you never know when those rights will be revoked or what you’ve done wrong. This kind of hidden power is truly frightening: it strips you of control over your own identity and hands your fate over to others.
Conclusion
Anthropic’s actions are not a minor technical issue; they represent a challenge to modern civilization. We cannot allow a company to secretly label people based on their identities, nor can we let “security” serve as a cover for discrimination. Today it’s punctuation marks; tomorrow it could be more severe forms of exclusion. As users, we must be vigilant against this hidden form of identity-based prejudice, as it can harm everyone who is labeled.