On February 24, 2026 Silicon Valley time, the most explosive controversy in the AI sphere was undoubtedly Anthropic's public accusation that multiple Chinese AI labs (DeepSeek, Moonshot AI, MiniMax) had launched an "industrial-scale distillation attack" on its Claude model. This accusation quickly ignited discussion, with Anthropic's official post garnering over 50,000 likes, thousands of reposts, and nearly 7,000 replies, becoming a hot topic among hot topics.
Anthropic detailed in its February 23 official blog and X post: these labs allegedly created over 24,000 fraudulent accounts, systematically "distilling" core capabilities from Claude through over 16 million interactions (queries), including the most differentiating features like agentic reasoning, tool use, and coding. The extracted data was allegedly used to improve their own models, violating terms of service, circumventing regional restrictions, and potentially undermining U.S. AI export controls. Anthropic emphasized that this behavior is not only intellectual property theft but could also enable foreign labs to bypass safety guardrails and input capabilities into military, intelligence, or surveillance systems.
The post immediately polarized opinion. Supporters of Anthropic viewed this as legitimate rights protection: U.S. frontier models require massive compute and R&D investment—how can they be "freeloaded"? Distillation attacks bypass chip export restrictions, threatening national security and technological leadership. Some pointed out that this exposed the "shortcut" nature of China's rapid AI catch-up, calling for industry, cloud providers, and policymakers to unite in response.
However, the backlash was larger and more intense. Numerous users, developers, and independent creators directly mirrored the accusation, forming a strong "who's the real thief" narrative—namely that Anthropic (along with OpenAI, Google, and other U.S. giants) themselves massively scraped public internet content (personal blogs, GitHub code, LibGen pirated books, etc.) to train their models, yet cry "theft" when being "reverse distilled," representing the ultimate double standard.
The most representative high-engagement reply came from renowned developer Jeff Geerling (@geerlingguy), whose post directly rewrote Anthropic's original:
"I've discovered industrial-scale copyright infringement of my content by Anthropic, OpenAI, Google, X, and other companies. These companies created thousands of crawlers, incorporating all my blog posts, open-source code, and book text into paid AI models for massive profit."
This mirror post received thousands of likes and tens of thousands of views, sparking widespread resonance. The comment section was filled with mockery like "throwing stones in glass houses" and "officials can set fires but commoners can't light lamps." Some users even dug up old news about Anthropic paying $1.5 billion in settlement for training data infringement and facing $3 billion in music copyright lawsuits, combined with Elon Musk's repost and Community Notes, directly pushing Anthropic to the forefront as a "hypocrite."
The controversy quickly extended to deeper levels:
- Ethics and intellectual property of AI training data: Does scraping public data count as "stealing"? Does API output distillation count as "fair use"? Where's the boundary?
- Legality of model distillation: Legal distillation for compressing one's own models is accepted, but do cross-company, ToS-circumventing "industrial-scale" operations constitute theft?
- Geopolitical dimensions of U.S.-China AI competition: The accusation's inclusion of terms like "CCP control," "military applications," and "circumventing export controls" ignited nationalist sentiments. On one side: "America First, technology blockade is justified"; on the other: "Chinese innovation is being stigmatized, U.S. giants stole the world's data first."
- Business model vulnerability: If frontier models can be "distilled" to near-parity with 16 million API calls, are massive compute moats and valuation bubbles illusory? Some bluntly stated: "Your moat was never model weights, but data flywheel and user relationships."
This storm is not just a technical debate, but a comprehensive collision of power, morality, and national narratives in the AI industry. Anthropic intended to reinforce its "safety + America First" image, but unexpectedly ignited the community's long-standing resentment toward all big tech's "data plundering." Within 48 hours, the topic evolved from a simple accusation to a global AI community's "turning the tables" carnival, with momentum still building.
In one sentence: When the "teacher" accuses the "student" of copying homework, they forget they first "borrowed" countless books from the world's libraries—this drama is too ironic, yet too real.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接