AI Companion training data availability

I have the same concern as we test AI Companion at my organization. This article (from Zoom) indicates the following:

Zoom’s unique federated approach to AI is designed to deliver high-quality results by dynamically incorporating Zoom’s own large language model (LLM) in addition to Meta Llama 2, OpenAI, and Anthropic. We evaluate and update our models periodically and models used to support Zoom AI Companion features may change from time to time. Features in AI Companion may use third-party AI-model providers, and we will share relevant data with those third-parties when you use the feature.

For example, if you use Chat Compose to help draft responses in Team Chat, the content used to provide the service as described in the section above — the selected chat thread, chat participant names, and your draft response — will be sent to the relevant model to provide the service. Similarly, AI Companion features used in Zoom Meetings that provide summaries or answer questions generate a temporary meeting transcript that will be sent to the relevant model to provide the service.

Even when data is shared, we do not allow third-party AI models to use your data to improve or train their models. Your data may be temporarily retained by those third-parties for Trust and Safety purposes or to comply with legal obligations. Data processed by Zoom AI Companion may be processed within U.S.-based data centers.

It seems like Zoom provides user content to third parties to provide their services, but the AI model itself is not trained on your data. It would be nice to get a more detailed explanation from Zoom, but this might be the best we've got right now.