AI transparency

1. You are talking to an AI

When you ask a question in Gelex, the answer is written by an AI model. Every answer says so and names the model that wrote it, also when you reopen it later. Text shown as a source quotation, with a citation and grade, is the publisher's text, not AI output.

2. AI text that leaves the app is marked

A draft made from an answer records that it was AI-generated and by which model, for as long as it exists. When you download it as a Word document, the file carries that in its properties (AIGenerated and the model), so other software can tell. If you publish AI-written text on a matter of public interest, the law may require you to say so (AI Act Article 50(4)); that duty is yours as the publisher.

3. Where no AI is used

Collecting the law, reading its structure, deciding which version was in force when, grading each document A to D, and building the search index are done by deterministic code from the publisher's stored files. No AI model is involved, so the same file always gives the same result, and a grade can be explained from measurements. The citation check that marks an answer's citations verified or not is also deterministic.

4. Whose model writes the answer

Gelex does not run its own language model, and we do not train any model on your content. Answers come from:

• In our app: the model your organisation connects with its own API key: any OpenAI-compatible or Anthropic-compatible API, for example Anthropic, OpenAI, Azure OpenAI, Mistral, DeepSeek, Cerebras, xAI, OpenRouter, Google Gemini through its OpenAI-compatible endpoint, or a model you host yourself.
• In your own assistant: through our MCP connector, the assistant you already use (such as Claude, ChatGPT, Cursor or GitHub Copilot) searches and reads Gelex while it answers in its own app. There the answer is written and shown by that assistant, not by us.

That provider is responsible for its model, and you for choosing it (see terms). We tell you which model answered.

5. How answers are checked

• Every citation in an answer is checked against what the model actually read while answering. It is marked verified if the model read that provision, not verified if the provision exists but the model did not read it, and not found if we do not hold it.
• A quotation of more than 40 characters next to a verified citation is compared with the provision's text and marked exact or inexact.
• Each cited document shows its grade, the dates its version was in force, and a fingerprint of the publisher's original file.
• We do not check the model's reasoning or conclusions. A verified citation means the model read the source, and an exact quotation means the words are there, not that the answer is right.

6. Limits

AI models can misread sources, leave out exceptions, mix up jurisdictions or dates, or state things with more confidence than the sources support. Read the cited provisions before relying on an answer, and get professional advice where the stakes call for it. See Not legal advice.

7. Risk classification

We have assessed that Gelex is not a high-risk AI system under Annex III of the AI Act. It is a research tool: it does not take or recommend decisions about people's access to justice, benefits, employment or credit, and it is not offered to judicial authorities to research and interpret facts and the law for them (Annex III point 8(a)). If we offer it to courts for that purpose, we will assess it again before we do.

8. Questions and reports

Report an answer that misrepresents a source, or a wrong citation check, with Wrong answer under the answer or to support@gelex.eu. Corrections to the corpus itself: takedown and corrections.