AI & transparency

This is a courtesy translation. RevueNova is operated by a French company and the French version of this page is the authoritative one; in case of discrepancy, the French text prevails.

1. Why this page exists

The European artificial-intelligence regulation — Regulation (EU) 2024/1689, the « AI Act » — has required since 2 August 2026 that anyone interacting with an AI system be clearly informed, and that AI-generated content be labelled (Article 50).

RevueNova uses AI centrally: most of its analysis tools produce generated text. This page says which ones, with what, from what, and within which limits. It supplements the terms of use and the privacy policy.

2. What the AI does here

The following tools send all or part of a paper to a language model and return generated output:

• Reading and understanding — Understand, Critique, Methodology, Relevance, Errors, Wins, Outlook, Deep reading, Polish the page.
• Structured extraction — Tables, Equations, Quantitative data, Variables, along with methodology detection and risk-of-bias pre-filling.
• Corpus-level work — cross-paper Q&A, Compare, Meta-analysis, Report.
• Occasional assistance — abstract translation, keyword and search-equation suggestions, monitoring triage, screening validation, Python code suggestions in the analysis workbench.

Everything else is deterministic and goes through no model: bibliographic search and source aggregation, bibliometric indicators, the citation map (snowballing), related papers, the relevance score, the PRISMA diagram, citation formats and exports.

3. What the AI does not do

• No automated decision is applied to you. AI output is a reading aid: it produces no legal effect and does not significantly affect you within the meaning of Article 22 GDPR.
• No automatic sorting of your corpus. At the screening stage the model never replaces a human decision: it screens blind, with no access to your decisions, and its only role is to measure the gap with your own judgements (sensitivity, specificity, Cohen's κ).
• No automatic execution. In the analysis workbench, suggested Python code lands in the editor, never in the kernel: you read it, then you run it yourself.
• No publishing on your behalf. Nothing is added to your notebook without an action from you.

4. Which model, which provider

RevueNova neither develops nor trains models. It is a deployer of models supplied by Anthropic, from the Claude family. Which model is used varies with the weight of the task: a fast model for short work, an intermediate model by default, a deeper model for long analyses.

Anthropic acts as a technical processor, under the conditions described in § 4 of the privacy policy. The content you submit is not used to train third-party models.

5. What is sent to the model

Depending on the tool: the title, the abstract, excerpts or the full text of the paper being worked on, together with your question or your criteria. Your credentials and account data are never sent.

The analysis workbench, specifically

The Python workbench runs entirely in your browser: your data is not uploaded. When you ask the model for code, only your column names and their nature (number, text, empty) are sent — never a cell value.

6. How generated content is labelled

• When a tool that calls a model opens, its panel carries a notice saying its results will be AI-generated — before they even are.
• Every generated result is followed by a provenance line recalling its origin and inviting you to check it.
• Any result inserted into your notebook keeps that notice there, alongside its timestamp and the source paper.
• Exported documents (PDF, Word, HTML, Markdown) carry a footer notice when they contain generated passages.

This labelling applies to content produced from 3 August 2026 onwards: blocks already saved in a notebook before that date do not carry it retroactively.

7. Limits — read this before relying on a result

A language model produces a plausible approximation, not a verified truth. It can be wrong, omit things, or state something inaccurate with confidence. Two safeguards are in place, and they do not remove that risk:

• Verified quotations. For the extraction tools, every value must come with a quotation found word for word in the source text; a quotation that cannot be found is discarded rather than displayed. The page number is recomputed by the application, never asked of the model.
• Provenance kept. Cells filled by the AI stay distinguishable from your manual entries; correcting a cell by hand erases its AI provenance.

Some tools are explicitly marked beta and stay that way as long as their output is projection rather than reading. In every case, scientific responsibility for what you publish remains yours: if you use AI assistance in published work, declare it — the screening validation tool even produces a methods statement ready to be reused.

8. Your controls

• No AI tool starts on its own. Every call follows an explicit action on your part and consumes credits, visibly counted.
• RevueNova remains usable without ever opening an AI tool: search, import, bibliometrics, manual screening, notebook and exports all work with no model.
• You can choose the language of generated analyses from your Settings, and delete your account and all associated data from the same section.

9. How RevueNova was built

For honesty, and because it shapes how the rest of this page should be read: RevueNova was developed by « vibe coding » — that is, having a language model write most of the code, under the direction and review of its authors.

We did so for lack of means, facing a real need for help in this field. We will bring in actual prefrontal cortices as soon as we can.

What this does not change: the rules described above apply to the product as it runs, not to how it was written. The same limits (section 7) hold, and a generated result still has to be checked before you rely on it.

10. Contact us

A question about the use of AI, a result that looks wrong, or a request for information under Article 50 of Regulation (EU) 2024/1689: contact.mquantum@gmail.com.