Anthropic has officially released its latest generation AI model Claude Fable5. As the company's strongest AI currently available to the public, it is highly anticipated for its potential in fields such as biology and cybersecurity. However, this model exhibits a "selective amnesia" in practical use: even when faced with basic biology questions at a high school level, it often refuses to answer and actively redirects requests to its previous flagship model Claude Opus4.8.

This phenomenon is not due to a lack of capability, but rather a deliberate restriction made by Anthropic based on its extremely high safety standards. As a Mythos level model, Claude Fable5 has strong capabilities for advanced tasks, but also comes with significant abuse risks. During development and launch, Anthropic regarded the field of biology as the top priority for safety, with its protection strategies even reaching a "overly cautious" level.

image.png

In actual testing, the side effects of this conservative strategy are evident. Whether it's popular science about mitochondrial function in cells, discussions on the causes of hay fever, or explanations of asthma medication principles, Claude Fable5 rejects all of them. Even topics related to public health, such as the Ebola virus, trigger its blocking mechanism. This "non-discriminatory interception" leads to the interruption of exchanges that originally had academic value and educational significance.

Anthropic responded to this by stating that the core purpose is to prevent malicious users from using advanced AI for high-risk biological research, such as developing biological weapons. Company spokesperson Parul Maheshwari said that to allow the model to go live as soon as possible, the team chose an extremely conservative protection approach and acknowledged that there are many misjudgments in the current identification process.

Compared to this, Claude Fable5's restrictions in the fields of chemistry and cybersecurity appear more flexible. Although it will refuse to provide dangerous information such as methods for making explosives or cultivating anthrax bacteria, it can normally answer neutral technical questions such as the use of chlorine gas, password security, and physics principles. It only calls upon the previous generation model for assistance when dealing with extreme scenarios involving toxic substances.

Currently, Anthropic is working on optimizing the identification mechanism to reduce misjudgments and plans to release specific versions in the future that can unlock these restrictions for professional users in the biomedical field, aiming to balance research efficiency and social safety. This "restricted release" model may be a reflection of how top AI models are currently trying to find a balance between powerful performance and strict security requirements.