Vellum policy evolution
Before/after stance changes across captured policy versions, with exact citations. If no before/after delta is available yet, AIRIN shows the latest citation-backed stance events instead.
No before/after stance delta is available for this filter yet. Latest citation-backed stance events are shown below.
Aug 3, 2026traininglow
model training
Latest stance: no training claim
“These technologies help us improve our Website and communications, see which areas and features are popular, count visits, and track clicks. You may be able to adjust your browser settings to remove or reject browser cookies. Please note that removing or rejecting cookies could affect the availability and functionality of our Services. Conversational Content: consistent with your preferences, we collect the content of your conversations, including the messages and files you send to the model provider and the responses generated for you, and use it to improve the quality and accuracy of our Services, including evaluating and improving the systems that power Vellum, but do not use it to train AI models. You can opt in or out of this collection at any time during onboarding or in Settings > Permissions & Privacy. This is separate from how third-party model providers handle your data, described under the "Third-Party AI Providers" section below.”
Open citationJul 10, 2026traininglow
model training
Latest stance: no training claim
“These technologies help us improve our Website and communications, see which areas and features are popular, count visits, and track clicks. You may be able to adjust your browser settings to remove or reject browser cookies. Please note that removing or rejecting cookies could affect the availability and functionality of our Services. Conversational Content: consistent with your preferences, we collect the content of your conversations, including the messages and files you send to the model provider and the responses generated for you, and use it to improve the quality and accuracy of our Services, including evaluating and improving the systems that power Vellum, but do not use it to train AI models. You can opt in or out of this collection at any time during onboarding or in Settings > Permissions & Privacy. This is separate from how third-party model providers handle your data, described under the "Third-Party AI Providers" section below.”
Open citationJun 17, 2026traininglow
model training
Latest stance: no training claim
“The Vellum Assistant uses third-party AI providers to generate responses, transcribe and synthesize speech, search the web, and perform other AI-powered tasks. When you use these features, the relevant inputs are transmitted to the third-party provider that performs the task. Each provider operates under its own privacy notice; we have summarized below the categories of data sent to each, and link to each provider's notice at the end of this section. Inference (chat) providers. When you send a message in the Vellum Assistant, the content of your message, any files you attach, and the surrounding conversation context are sent to the AI provider that hosts the model you have selected: Vellum's managed inference service routes your data to Anthropic (Claude family) using Vellum's credentials. If you connect your own API key in your account settings, your data is sent to Anthropic , OpenAI , or Google (via Google Vertex AI) using your credentials and is governed by the terms of your relationship with that provider. Under the published terms of Anthropic, OpenAI, and Google's commercial APIs, the inputs you submit and the responses generated for you are not used by these providers to train or improve their underlying models, and are not used by these providers for advertising. Image generation. When you generate images in the Vellum Assistant, the prompt you submit is sent to Google (via Google Vertex AI Imagen).”
Open citationGenerated from live stance events. Informational only, not legal advice.