AI transparency
Last updated October 7, 2026 · Our approach to the EU AI Act (Regulation (EU) 2024/1689, as amended in 2026)
Where Groundtruth uses AI
| Feature | Uses an AI model? | How it's labelled |
|---|---|---|
| Ask, in the web app, with your own AI key | Yes: your Claude, OpenAI or Gemini model | “AI-generated · model” on every answer, plus signed machine-readable marking |
| Ask, in the web app, on the free AI credit (no key yet) | Yes: Claude Haiku 5.5 on Groundtruth's account, up to $0.50 per person | “AI-generated · claude-haiku-5-5” on every answer, plus signed machine-readable marking |
| Ask, without a key once the credit is used | No: sentences are extracted verbatim from your sources | “Not AI-generated” |
| Setup from your website | Yes: your AI key, or the free credit, reads up to five public pages of your website and drafts pages, FAQ answers, records and proposed decisions | Written as the “Website import” agent with the “AI” badge, an “AI draft (model)” note in the text, and marked for your review; decisions stay “proposed” until you accept them |
| Ask, called by your agents (MCP/API) | No model on our side: we return sources; your agent's model writes the answer | Disclosure field in every response |
| Contract extraction | Yes: your AI key reads the contract text | Marked as AI-extracted; you review every field before saving; the record keeps the marking |
| Write mode (alternatives, Lab, trims) | Yes, only when you ask: your AI key suggests wordings, marks passages or proposes cuts | Every suggestion is labelled as AI (with the model). It never rewrites your text: you choose, edit or ignore each suggestion, and trims only fade words until you confirm |
| Structure Agent | Yes, only when you press Structure: your AI key (a cheaper model if you choose) or the free credit reorganises the page | Shown as an “AI suggestion” with the model, next to your current page. Nothing changes until you click Replace; the activity log records the model and the old version stays restorable |
| Contract, Churn and Freshness agents | No: deterministic, explainable rules you can configure | Badged as an automated agent; every finding states its reasons |
| Groundtruth Capture (browser extension) | No: it saves the page, document or thread you chose. Optional smart fill suggests a title and tags with your own Anthropic key, and you review them before saving | Saved under your name, marked “via Groundtruth Capture”; not labelled as AI |
| Content written by your own AI agents (MCP) | Your agent's model | “AI” badge on every write, a notice on pages last edited by an agent, and last_edited_by_ai_agent in the page's markdown |
Machine-readable marking (Art. 50(2))
AI-generated text from Groundtruth carries a signed provenance record: the generator, model, time, a SHA-256 of the exact text and an Ed25519 signature. It's in every API/MCP response (provenance) and in the web page (data-ai-provenance). Verify any marking with our public key at /.well-known/ai-provenance or POST /api/provenance/verify. Our Terms forbid removing it.
Human oversight
- Agents can add and edit knowledge but never delete it. Work they finish goes to Needs review for a person.
- Every change is attributed to a person or an agent in the activity log.
- Answers refuse rather than guess when your workspace has no source.
- Pages have owners and a verified date; stale pages are flagged.
What Groundtruth is not for
Groundtruth isn't designed or assessed for the high-risk uses in Annex III of the AI Act (for example hiring or evaluating employees, creditworthiness or insurance pricing of individuals, education admissions, essential public services, law enforcement, migration or justice), nor for any practice prohibited by Article 5. Our Terms don't allow these uses. The Churn Agent scores business accounts from logged interactions; it doesn't use biometric data and doesn't infer anyone's emotions.
Models and data
We don't train AI models. When you add a key, your questions and the matching workspace excerpts are sent to that provider (Anthropic, OpenAI or Google) under your own account. On the free AI credit they go to Anthropic under Groundtruth's account (Anthropic doesn't train on API data). Neither is used by Groundtruth for training.
AI literacy
Our team follows an internal AI-literacy programme, and our docs explain how each AI feature works and its limits so your team can use it knowingly.
Contact
Questions, or think something isn't labelled correctly? ai@groundtruthhq.tech