An off switch for controlling dual-use knowledge in AI models

Among the knowledge AI models possess, there is dual-use knowledge—such as cybersecurity or virology—that can be used for both benevolent and malicious intent. Existing filtering methods had limitations in that they required retraining the entire model or caused performance degradation.

GRAM (Gradient-Routed Auxiliary Modules), jointly researched by Anthropic and AE Studio, is a method that allocates dedicated modules within the model for specific knowledge categories. During training, general text is learned by the entire network, but when hazardous knowledge data is input, the system is designed so that only the Auxiliary Module for that category is updated.

The key benefits of this technology are as follows:

• Selective control: You can remove a model's hazardous capabilities by deleting specific hazardous knowledge modules, or retain functionality for trusted users by leaving those modules intact.
• Efficiency: With just a single model training run, numerous versions of the model can be implemented depending on the combination of modules turned on/off, dramatically reducing the cost of having to retrain the model each time.
• Performance maintenance: Since hazardous knowledge does not spread throughout the entire model but accumulates only in specific modules, the model's general performance is not affected.

Experimental results demonstrated GRAM's effectiveness: when reconfigured to 'forget' a specific topic, it showed almost identical performance to a model trained from scratch excluding that topic.