Anthropic Under Pressure: Criticism of AI Limitation Strategy

Anthropic Facing Criticism

Anthropic, a company specializing in the development of artificial intelligence models, has recently come under fire for its lack of transparency and the limitations imposed on its models. While the company justifies these restrictions for security reasons, it also aims to protect its business from AI model distillation, a technique that would allow competitors to appropriate its technological advancements.

This week, Anthropic decided to modify its strategy in response to negative feedback from developers. The company announced that it would no longer secretly degrade the responses of its Fable 5 model for requests related to advanced model development. From now on, these requests will be redirected to a less capable model, Opus 4.8, and users will be informed of this change. This decision aims to address concerns regarding the provision of intentionally poorer responses. Anthropic apologized for having intentionally provided lesser-quality responses while concealing this fact.

Security or Business Strategy?

Despite these adjustments, Anthropic continues to restrict the use of its most advanced public model for certain AI development projects. The company claims that these limitations are necessary to prevent "foreign adversaries" from exploiting its technologies, which could jeopardize the United States' lead in AI. However, this explanation is being questioned, notably by Nicholas Vincent, a professor at Simon Fraser University, who believes that these measures seem more driven by commercial interests.

Anthropic's restrictions also help protect its business against distillation, or intelligence extraction. This technique allows competitors to query a powerful model, collect its results, and use that data to improve their own systems. Anthropic has warned against Chinese labs using these distillation techniques, but the same threat also comes from developers of open-source models in the United States and Europe.

Anthropic's terms of service also prohibit anyone from using its products to develop competing offerings. This policy applies not only to Chinese labs but also to open-source model developers in the West, raising questions about the notion that the restrictions are solely aimed at countering foreign adversaries.

The Threat of Open-Source Models

Open-source AI models represent serious competition for Anthropic. A study from MIT Sloan showed that these models achieve, on average, 90% of the performance of closed models, closing the gap in just 13 weeks, whereas a year earlier, it took them 27 weeks to bridge that gap. The open-weight model MiMo v 2.5 Pro from Xiaomi, for example, is approaching Anthropic's performance while being much more affordable. While Anthropic's Fable 5 costs $10 per million input tokens and $50 per million output tokens, Xiaomi's model costs only 43 cents per million input tokens and 87 cents per million output tokens.

Artificial Analysis tracks the performance of AI models, and rankings like those from Arena show that Anthropic's models remain at the top for complex tasks such as mathematics, programming, and creative writing. However, the pressure from open-source models, which are "almost as good" and "much cheaper," is undeniable.

A Strategy Under Pressure

Anthropic is not required to share its technology with competitors, but the company is criticized for its lack of transparency. The restrictions imposed by Anthropic aim to curb competition, but they raise questions about the balance between security and business strategy in the AI sector. "This seems to be a business move," said Nicholas Vincent, emphasizing that without explicit targeting of "bad organizations," it is difficult to justify these measures as being security-focused.