Microsoft and NVIDIA Publish Guide to Managing Risks of Open-Weight AI Models in Enterprises

At a glance

Microsoft and NVIDIA have released a joint guide on how companies can safely use open-weight AI models (models whose weights are public, so anyone can examine and modify them). This guide is the first part of a series.

Key points

1. 1

   How risky a model is depends more on the purpose and environment in which it is used than on whether it is open or closed.

2. 2

   To describe risks in consistent terms, the guide sets out five categories: cybersecurity, data protection, content, human-AI interaction, and bias.

3. 3

   It recommends using model evaluation results to identify risks and decide which safeguards are needed.

4. 4

   Evaluating the model alone does not guarantee the safety of an agent or app, so permissions, data access, and similar factors should also be reviewed.

Why it matters

It gives companies a starting point for deciding who is responsible for which risks when they run AI models themselves.