Browse the docs
Privacy and security
No training on your data
ASDK doesn't train AI models on your content or your visitors' chats, and neither do its AI providers. What is sent to them, why, and what it means for you.
Last updated
On this page
A common question before adding an AI assistant to a website: will my content, or my customers' questions, end up training someone's AI model? With ASDK, the answer is no. This page explains what that commitment covers and how it works.
The commitment
- ASDK doesn't use your content or your visitors' conversations to train foundation models. That covers your pages, files, text snippets, Q&A pairs and every conversation in the widget.
- ASDK never uses one customer's content to answer another customer's visitors. Each site's knowledge is kept in its own namespace, and an agent can only search the knowledge of the site it runs on.
- The AI model providers ASDK uses process data under terms that don't allow them to train their models on it.
This is part of our privacy policy (section 4) and our Data processing addendum.
What is sent to AI model providers, and why
ASDK uses AI models for two jobs:
| Job | What is sent | When |
|---|---|---|
| Making your content searchable | Passages of your content | When ASDK reads a page, file or snippet, and again when it changes |
| Writing answers | The visitor's question, recent messages of the conversation, the most relevant passages of your content and the agent's instructions | Each time a visitor asks something |
Only what the task needs is sent. Your whole knowledge base is never sent to write one answer: only the few passages that match the question.
What the providers say
The subprocessors page lists three AI model providers, used as configured, and summarizes their terms:
- OpenAI states that data sent to its API is not used to train its models unless the customer opts in, and ASDK doesn't opt in.
- Anthropic's commercial terms state that it may not train models on customer content.
- Google's Gemini API terms state that prompts and responses on paid services aren't used to improve its products. ASDK uses only the paid services.
Providers may keep requests for a limited time to detect misuse. OpenAI, for example, states that it keeps abuse-monitoring logs for up to 30 days by default. The privacy policy and the subprocessors page link to each provider's own terms.
What "no training" doesn't mean
To be precise about the limits of the promise:
- Your content is processed. Agents can't answer from your content without reading it: ASDK stores it, makes it searchable and sends relevant passages to a model for each answer. Processing to run your agents is the service, not training.
- ASDK improves its own service. We look at aggregated, technical signals, such as how often answers are grounded in your content or how long they take, to make the product better. That's not training a foundation model on your content.
- Public pages are public. ASDK only reads what your website already shows to everyone. If your pages are public, other companies' crawlers, including AI companies' crawlers, may read them independently of ASDK. Your robots.txt controls those; our article on robots.txt for AI bots explains how.
What it means for you
- You can add internal documents such as price lists or product sheets without them becoming part of a public model. Remember that agents can quote anything in your knowledge to your visitors, so add only what visitors may see.
- Your visitors' questions stay yours. They're stored for your Inbox and insights, under your control as controller, and kept for 12 months by default. See Data retention.
- You can tell your customers so. A sentence such as "Questions you ask our website assistant aren't used to train AI models" is accurate for ASDK.
Questions from your security or legal team
The security page describes how ASDK protects data, and the subprocessors page lists every provider with its location. For a security questionnaire, write to hello@asdk.com.
Related: Data we store, Subprocessors, and the glossary entry on large language modelsA large language model (LLM) is an AI model trained on vast amounts of text to predict the next token, which lets it write, summarize, translate and answer..