Managing Misuse, in Dual-Use Foundation AI Models

Reading Time: 2 minutes

When making decisions that are critical to national security, governments rely on data, and those that leverage the cutting edge technology of generative AI foundation models will have a distinct advantage over their adversaries.

Pros and Cons of generative AI foundation models

The rewards of using these models, however, are not without risks. Many are dual-use, in that they can be used for both beneficial and harmful purposes, and the inner workings of such models are beyond human comprehension. Governments are therefore wise to balance the valuable benefits of dual-use models with cautious optimism and to develop frameworks such as those stated in NIST AI 8000-1, to help manage the associated risks.

While NIST accurately identifies several methodological and scientific challenges with dual-use, it fails to mention an important insider threat – an organization’s own fallacies in the way it manages its data.

All too often, data is copied, replicated, or changed amongst inoperable systems and kept longer than necessary. Such practices corrode the methodologies used to derive accurate results and valuable insight. In a generative AI context, impure data can cause inaccurate outputs, a risk not worth taking if it involves people’s lives and U.S. national security.

The Logical Data Management Factor

The Denodo Platform, a logical data management solution, can be leveraged to improve data quality by creating interoperability between multiple different data sources for real-time data analysis with greater precision. The Denodo Platform integrates, manages, and delivers data from many data sources without requiring the data to be copied or moved. It can be used to quickly and safely implement retrieval augmented generation (RAG), the current best practice for providing large-language models with correct, accurate, and up-to-date data from all relevant data sources. The Denodo Platform also provides fine-grained access control, data de-identification, and data usage monitoring features, so that sensitive data is kept secure, mitigating the risks of leaks and mis-use.

The power of the Denodo Platform is further augmented when combined with generative AI (GenAI) features of its own, such as the ability to query all data in human natural language, so that anybody working in a government organization with the appropriate authorization can conduct data analysis in real time without needing to use data engineering skills. This multi-faceted approach enables organizations to leverage both their static historical data as well as “dynamic” data being updated in real time, as is typical in real-time analytics use cases such as threat detection.

By leveraging the Denodo Platform’s support of RAG, organizations can provide the necessary fine-tuning to a broad spectrum of models using their own domain specific data sources, securely stored on their premises. Also, by enabling the retrieval of trusted, dynamic data from across the organization, the Denodo Platform augments performance metrics and improves the accuracy of dual-use models for swift, confident, and intelligent decision making using real-time data.

For further information about how the Denodo Platform supports GenAI, please visit our GenAI page.

Author
Recent Posts

Maricela Lechuga

Maricela Lechuga is an Associate Legal Counsel at Denodo Technologies Inc. She is also an environmentalist who has advocated for a ban on the use of lead-based fuel in aviation. Her activism has taken her before the US Congress where she testified as a witness in a hearing held by the Subcommittee on the Environment. In her spare time, Maricela likes to create art and license digital downloads on Etsy.

Latest posts by Maricela Lechuga (see all)