Last updated: 2026-09-10
Every answer you get from Sourcebound is generated by an AI system. This page explains how, and what that means for you.
1. How an answer is generated
When you select a fragment of a report and ask a question:
- Your question and the selected fragment are sent, together with the most relevant excerpts from our own research Corpus (the claims and sources our research pipeline has collected for that report), to Claude — an AI model built by Anthropic — through Anthropic's API.
- Claude is instructed to answer using only the retrieved Corpus material — not its own general knowledge, and not the open internet.
- If the Corpus doesn't contain enough to answer, Claude is instructed to say so rather than guess. In that case you are not charged, and your question is logged to our research queue for a human researcher to potentially cover in future.
- If Claude does answer, the response is labelled "AI-generated" and shown to you together with the sources it drew on, so you can check them yourself.
2. What "grounded in our Corpus" means — and doesn't mean
- It means the answer is built only from claims and sources our research process has already collected and vetted for the relevant report.
- It does not mean the answer is guaranteed correct, complete, or up to date. AI systems can misread a question, misapply a source, or phrase something misleadingly even when the underlying material is accurate.
- It does not mean Claude has searched the live internet, other websites, or its own general training knowledge to answer you — that's a deliberate design choice, not a limitation of the technology.
3. What the system will not do
- It will not fabricate sources or invent claims not present in our Corpus.
- It will not answer from general AI knowledge when our Corpus is insufficient — it will tell you it can't answer instead.
- It will not use your questions, selected fragments, or feedback to train Anthropic's or any third party's underlying AI models. Per Anthropic's commercial API terms, inputs and outputs sent through the API are not used to train Anthropic's models.
- It is not a substitute for professional legal, financial, investment, medical, or engineering advice — see Section 8 of our Terms of Service.
4. Errors happen — how to report one
AI-generated answers can still be wrong: a source can be misquoted, a nuance lost, or a question misread. If an answer looks wrong or off, use the feedback control shown under every answer to flag it. If you paid a credit for it, you can request it back under our Credits & Refunds Policy within 14 days. Flagged answers are reviewed by us and, where useful, fed back into improving the Corpus and the retrieval prompts.
5. Why every answer is labelled "AI-generated"
The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) sets transparency obligations in Article 50 for providers and deployers of certain AI systems. Among other things, deployers of an AI system that a natural person interacts with, and deployers of systems generating AI text, must make clear that the person is interacting with an AI system and that the content is AI-generated, unless this is already obvious from the context. These obligations apply from 2 August 2026 (with a transitional period to 2 December 2026 for systems already on the market before that date). We label every answer as AI-generated for exactly this reason — and because we think you should know, regardless of what the law requires. (The precise scope of Article 50 as applied to a retrieval-constrained deployment like ours should be confirmed by a professional before relying on this section — see our internal Legal Register.)
6. Your data and model training
We do not use your questions, the fragments you select, your feedback, or generated answers to train our own models (we don't train any), and — per Anthropic's commercial terms — Anthropic does not use this data to train its models either. See our Privacy Policy for full detail on what we collect and why.
7. Contact
Questions about how answers are generated: hello@sedesmaximus.com.