When Data Becomes a Battleground: Anthropic, Distillation, and DataClaw's Counterattack
*A story about walls, gardens, and a crowbar*
Have you ever wondered where your AI conversations go? They flow like rivers into company servers—stored, analyzed, and possibly used to train the next version of the model. You created this data, but it no longer belongs to you. That has been the default rule of the AI era: users produce data, companies own data. Until someone decided to change that rule.
The Walled Garden
Anthropic, the company known for "AI safety," recently published a hard-hitting article, *Detecting and Preventing Distillation Attacks*, claiming that three Chinese AI companies—DeepSeek, Moonshot, and MiniMax—made over 16 million distillation calls against Claude.
Distillation means using a large model's (teacher) outputs to train a smaller model (student). It is common practice—arguably, without distillation there would be no modern AI.
But netizens quickly dug up Anthropic's own history:
- Claude was trained on massive amounts of internet data
- Anthropic itself has been accused of distilling OpenAI's models
- More embarrassingly, it once downloaded e-books from piracy sites to train its models
- Anthropic. (2026). *Detecting and Preventing Distillation Attacks*. https://www.anthropic.com
- POM. *DataClaw*. https://github.com/peteromallet/dataclaw
- Hugging Face. https://huggingface.co
Musk criticized them directly: "Anthropic previously misused others' training data and paid billions in settlements."
This is classic double standards: when I use your data, it's "learning"; when you use mine, it's an "attack."
Who Owns the Fruits of the Garden?
The deeper question: who owns the data an AI model produces? The traditional answer is the provider—the terms of service say your usage data improves their models. But there is a paradox: if Claude's answers derive from public internet data created by countless people, on what basis can Anthropic claim exclusive rights over its outputs?
> Knowledge flows one way: from the public to the company, but not back.
The Crowbar
Shortly after Anthropic's article, an open-source developer known as POM (Peter O'Malley) published his 155,000 real local Claude Code conversation logs and released an open-source tool: DataClaw.
The name is pointed: Claude's logo is a claw; DataClaw is the claw that claws data back from Anthropic's walled garden.
What DataClaw does: it one-click exports your conversation histories from AI coding tools (Claude Code, Codex, Gemini CLI, etc.) into standard-format datasets that can be uploaded to Hugging Face. Musk reposted it with one word: "Cool."
The Art of Privacy
DataClaw does not blindly upload raw data—it includes layered privacy protection:
| Original | Sanitized |
|---|---|
| /Users/alice/project/secret.py | ./project/secret.py (relative path only) |
| alice@company.com | user_7a3f9e2 (anonymized) |
| sk-abc123xyz789 (API key) | [REDACTED] (auto-detected and removed) |
| DB passwords, tokens | Multi-pass scanning, auto-removal |
Every step requires your confirmation: select source → confirm scope → local preview → privacy scan → confirm and push.
Why It Matters
1. Data sovereignty made concrete. With pip install dataclaw, the decision—share or keep—returns to you.
2. Undervalued data. Anthropic's strong reaction shows these conversations were undervalued. Most training data today is synthetic or static code from public repos. Real-world debugging, requirement decomposition, and iteration are almost never covered. Developer-AI pairing conversations are golden data: real problem-solving, human thinking patterns, learning from mistakes.
3. A legal gray zone. Are AI-generated outputs bound by the original model's terms of service? Can companies reverse-train on user behavior? Do users have the right to export? These questions are nearly blank in law.
Limitations
The author is honest—the README says: "This is NOT foolproof." Automated sanitization may miss things; sensitive info can hide in unexpected ways. Review your data before publishing. Whether individual contributions will be effectively used, or just scattered across Hugging Face, remains unanswered. This is a starting point, not a solution.
The Bigger Picture
The internet began decentralized; Web 2.0 turned data from "public property" into "platform assets"; AI sharpens this further—you now produce *training data* that builds models that may serve or replace you. DataClaw represents quiet but firm resistance: "You built a wall? I'll plant a garden."
Conclusion: Garden or Wall?
This is not a simple right-versus-wrong judgment. Anthropic protects massive investments; POM believes knowledge should be free and data sovereignty belongs to users. The real question is what future we want. DataClaw cannot answer that, but it gives us a choice: rather than waiting for others to decide where your data goes, choose yourself. Even if you never publish, periodically exporting your own conversation records preserves your genuine thinking process—your own digital asset.
References
*"Knowledge is like air. Trying to monopolize it is like trying to monopolize breathing."* — an open-source community developer