US AI Giant's New Model Sparks Controversy with Chinese Tech Links
US AI Leader Faces Backlash Over Training Data Questions
Anthropic, the prominent American AI developer, launched its flagship Claude Opus 4.8 model this week amid growing controversy. While the model initially impressed with strong benchmark results, unexpected behavior during API testing has raised eyebrows across the tech industry.
Identity Crisis in the Code
Independent developers first noticed something unusual when querying the model about its origins. Without the constraints of web interface safeguards, the system occasionally claimed to be either Alibaba's 'Qwen' or 'DeepSeek' - two major Chinese open-source models. These responses only appeared in raw API tests, not in the polished web interface where strict system prompts prevent such revelations.
"It's like asking someone their name and getting three different answers," explained one developer who reproduced the behavior. "The web version gives you the corporate-approved response, but the unfiltered API sometimes shows what might be under the hood."
Training Methods Under Scrutiny
The anomalies suggest Anthropic may have used substantial data from Chinese models during training - a practice some call 'distillation.' This involves using outputs from existing models to train new ones, potentially inheriting characteristics of the source material.
What makes this particularly awkward is Anthropic's previous political stance. The company has actively supported U.S. government efforts to restrict China's AI development, even helping draft legislation that could classify similar data use as adversarial attacks.
Hypocrisy Claims Emerge
Industry watchers quickly noted the apparent contradiction. "They're pushing to limit China's access to AI chips while possibly building their own model with Chinese tech," remarked one researcher who requested anonymity due to professional sensitivities.
Anthropic hasn't directly addressed the specific claims but maintains that Opus 4.8 represents entirely original work. The company suggests the identification quirks might stem from evaluating against various models during development rather than direct training use.
Key Points
- Unexpected behavior: Claude Opus 4.8 occasionally identifies as Chinese models in API tests
- Training questions: Findings suggest possible use of Chinese model outputs in development
- Political tension: Controversy highlights complex US-China tech competition dynamics
- Industry reactions: Some see hypocrisy in Anthropic's previous anti-China advocacy