Anthropic Faces Backlash Over AI Model Restrictions, Citing Safety and Competition Concerns
AI company Anthropic has drawn criticism for implementing limitations on the use of its AI models. While the company stated these restrictions are for safety reasons, including preventing foreign adversaries from eroding the US's AI advantage, the move has also been interpreted as a strategy to protect its business interests. These restrictions specifically aim to prevent rivals from using techniques like distillation to improve their own open-source models, which have shown competitive performance at significantly lower costs.
AI developer Anthropic recently faced backlash regarding restrictions and transparency surrounding its AI models. The company initially responded to developer concerns this week by stating it would no longer secretly degrade Fable 5 responses when users sought help with frontier AI model development. Instead, these requests will now be routed to a different model, Opus 4.8, with developers being informed of the change. Anthropic issued an apology for its prior approach.
The company maintains that its most powerful public model will continue to have usage restrictions for certain AI development tasks, citing safety reasons. Anthropic argues these limits are crucial to prevent "foreign adversaries" from utilizing its top model to diminish America's lead in AI and chip technology.
However, these restrictions are also seen as a measure to safeguard Anthropic's business from a process known as distillation, or intelligence extraction. This technique involves rivals querying a powerful model, collecting its outputs, and then using this data to enhance their own systems. Such methods enable open-source model providers to close the performance gap with companies like Anthropic more quickly and offer competitive pricing. Anthropic has previously expressed concerns about this practice from labs in China, but the same threat also emerges from open-model developers in the US and Europe.
Anthropic's terms of service prohibit users from developing competing offerings with its products. Open-source models are proving to be potent rivals; an MIT Sloan analysis from January indicated that open models achieved an average of 90% of closed-model performance, typically closing the remaining gap within 13 weeks. A year prior, this process took 27 weeks. Performance tracking by Artificial Analysis and Arena's leaderboard shows open-source models keeping pace, with Xiaomi's open-weight MiMo v2.5 Pro nearing the performance of top Anthropic models for expert text-based tasks.
Price disparities further highlight the competitive pressure. Xiaomi's open model costs 43 cents per million input tokens and 87 cents per million output tokens. In contrast, Anthropic's Fable 5 costs $10 per million input tokens and $50 per million output tokens, making it at least 20 times more expensive, according to Arena data. Nicholas Vincent, a computer science professor at Simon Fraser University, suggested the move appears to be a business decision, noting the difficulty in defending it as solely safety-focused without more specific targeting of organizations.
(Source: Business Insider)
