Where AI is in QualCanvas — and where it isn't.

No. When you ask QualCanvas to auto-code, summarize a theme, or sketch a framework, the QualCanvas backend forwards your prompt to the model provider you've chosen (OpenAI, Anthropic, or Google), receives the response, and returns it to your browser. Source transcripts already stored in your project may be included in that request. QualCanvas stores selected AI work products—such as suggestions, summaries and chat messages—so you can review them later; usage analytics store feature, provider, model and token counts.

The model providers we support contractually exclude this API traffic from training. Their policies:

• OpenAI API data usage policy — API traffic excluded from training.
• Anthropic commercial terms — API content not used for training without explicit opt-in.
• Google Vertex AI data governance — customer data not used for training.

Bring your own key. If you'd rather inference be billed and observed by your own provider account, you can configure a personal API key in Account Settings. With BYOK, QualCanvas uses your key when forwarding requests; the provider relationship is between you and them, and your IRB can audit the provider directly.

Three things, all opt-in per action, all reviewable, all reversible:

1. Auto-code suggestions. The model proposes code names for spans you select. Every suggestion is reviewed before it lands; nothing is silently applied.
2. Theme summarization. When you ask, the model drafts a paragraph synthesizing the codes you've grouped under a theme. You edit it before it ships.
3. Framework synthesis. For cross-case work, the model can sketch a framework matrix from your coded cases. You restructure it as you see fit.

No analysis runs without your action. There is no background AI watching your project. The model never applies a code without your accept-click; the model never overwrites a memo.

Usage and research records. For each AI call we record: which feature was used (auto-code / theme / framework), which provider, which model, and the token count returned by the provider. This is what we use to bill, debug, and report on AI behavior.

We do not put full prompt text into general request or analytics logs. The project database can store model suggestions, summaries, chat messages, job results and the source excerpts needed to make those records useful and auditable. Those records are part of your project data and are included in account deletion and export workflows.

When you accept or reject a suggestion, the accept/reject decision is recorded against the coded span so your methods statement can report it (e.g. "model accepted 14/22 suggestions; researcher revised 6"). That record holds the resulting content, reasoning where supplied, and the user's decision.

On Team and Institutions tiers, this usage and decision record is visible to project admins.

AI features are never automatic — you have to click an "AI suggest" button for any inference to happen. So the simplest way to run an AI-free study is to not click those buttons. There is no background activity that needs to be turned off.

If your IRB requires a stronger guarantee, contact [email protected]. A project-enforced AI-disable control is not currently available, so studies requiring a technical prohibition should not rely on QualCanvas until that control is released.