AI transparency

Version: 2026-10-draft

The French version of this notice is the reference version. This English version is a translation provided for convenience.

1. In short

Quorumind's reports are produced entirely by artificial intelligence models. They say so, they state which models played which role, and they mark how certain their key claims are. They are well-argued second opinions: the decision stays yours.

2. The AI systems used

1. Quorumind Cloud uses general-purpose large language models provided by:
   • Anthropic (Claude models);
   • OpenAI (GPT models);
   • Google (Gemini models);
   • Mistral AI (Mistral models).
2. They are called through their APIs, with Quorumind's accounts: you need no key. Most go through OpenRouter, a gateway that passes each request to the model's provider and asks that no data be kept.
3. The list of available models is in the app (Settings, Models section). Only models whose price is known are offered, so that the cost of each debate can be calculated.
4. Quorumind does not train any model and does not use your debates to train one.
5. Web access: for Standard and Deep debates, a research step checks on the web the facts the debate depends on, with short generic queries, and each fact keeps its source and date. Otherwise, and always in "European models only" mode, the models rely on their knowledge, which stops at a certain date, and on what you give them.

3. What each model is for

Each debate shares roles among several models, which you can choose when you start, preferably from different providers so that they do not share the same blind spots:

The same models are used to rewrite a report for another reader (executive, client, CTO), without changing its substance.

4. Every report says it was produced by AI

1. During the debate, the interface shows which role and which model is speaking.
2. Every report contains a "How this report was produced" box that explains the method and lists the models by role (interviewer, proposers, devil's advocate, arbiter).
3. Exports (Markdown, PDF, Word) carry the statement "Generated by Quorumind", or "Prepared with Quorumind" when the report carries your agency's branding.
4. If you publish a report, or part of it, to inform the public, say that it was produced by AI.

5. The limits

1. Models can be wrong: invent a fact, rely on outdated information, misunderstand your needs, or wrongly agree with each other. In Audit mode, every passage quoted from the reviewed document is checked word for word by the code: a quote that cannot be found is dropped. The market price range comes from the models and remains an estimate.
2. Level of certainty: each key claim is marked "verified", "estimated" or "assumed". This label is the model's own assessment, not a check made by a human. Check the estimated and assumed claims first.
3. Costs and durations: the models estimate the effort of each work package, in days per role. Quorumind computes the amounts itself, from your profile's day rates (or, without a profile, from rates estimated by the models, marked as such), and checks the budget and deadline by calculation. These are estimates, not quotes. The compliance checks (GDPR, health data, public sector, finance, online commerce) are produced by the models: they are points to verify, not legal advice.
4. Disagreements: the verdict shows the points on which the models still disagree.
5. Price of the debate: on Quorumind Cloud, a debate's price is fixed by its tier and shown before it starts (section 7 of the terms).

6. Human control

You stay in control at every step:

1. you describe the project and answer the interview;
2. you correct and validate the requirements sheet: the debate relies on it alone;
3. the debate pauses to ask you a question when missing information could change the outcome;
4. you can run the debate again with modified requirements ("Start again from here"); You can also have the recommendation updated with new elements (your client's answers, new facts): only the arbiter rules again, and each version stays readable with what changed;
5. you make the final decision. Nobody at Quorumind reviews the reports.

Do not use Quorumind to make decisions that have legal or similarly significant effects on people (hiring, credit, insurance, scoring, etc.): the terms forbid it.

7. Personal data

The content of your debates is sent to the providers of the chosen models. When you start a debate, the "Mask personal data" option replaces e-mail addresses, phone numbers, card numbers, IBANs, IP addresses and secret keys before anything is sent. It does not detect names or postal addresses. The details are in the privacy policy.

8. Reporting an error

1. Below the verdict, the "Report an error in this report" button lets you note what is wrong. The note is saved with the session, in your account, and deleted with it: it is a record for you and your team, but we do not read it.
2. For us to look into an error, write to hello@quorumind.com and describe the problem (attach the report's export if you wish). This feedback helps us improve the instructions given to the models and the debate method.
3. We do not use your debates to train models.

9. The EU AI Act

1. Regulation (EU) 2024/1689 on artificial intelligence sets transparency obligations (article 50), applicable from 2 August 2026: informing people that they are interacting with an AI system and that content is AI-generated. This notice, the interface and the method box of every report meet these obligations.
2. Quorumind is not designed for the high-risk uses listed by this regulation, in particular decisions about people; these uses are excluded by the terms.
3. The providers of the general-purpose models (Anthropic, OpenAI, Google, Mistral AI) have their own obligations under this regulation.

10. Contact

hello@quorumind.com