How we write

Our sources

Around forty public sources, in French and English: official lab blogs (OpenAI, Anthropic, Google DeepMind, Mistral AI, Hugging Face…), research (arXiv, Hugging Face Daily Papers, CNRS, The Conversation), specialised press and institutions (CNIL, European Commission, EDPB).

We only read what each source publishes for syndication: the title, link, date and excerpt from its RSS feed. Never the full article page. A source that refuses this kind of use (robots.txt, text and data mining reservation, terms of use) is left out.

What we keep

Every 30 minutes, new articles go through the same criteria:

• published less than 48 hours ago;
• for general news outlets, only articles about AI;
• at most 8 articles from the same source per run, so that none dominates the feed;
• duplicates removed (the same article is only published once);
• a relevance score from 1 to 10 (reach, novelty, practical impact) that orders “Today's essentials” and the Morning Brief.

What AI does

A language model (Mistral) reads the source excerpt and writes, in French and English: the title, the summary, the three “In brief” points, the “What it changes” line for each industry concerned, the category and the relevance score. Another model (Claude) writes the Morning Brief editorial and the Friday review from articles already published.

What it does not do

• It does not choose the sources: the list is maintained by hand.
• It does not read the full article or copy its text: the summary is short and reworded.
• It adds no figures, quotes or context missing from the excerpt: when the source is brief, so is the summary.
• It does not replace the source: every story cites it and links to the original article, which remains the reference.

Rule-based checks

Summaries are not reviewed one by one by a person. They go through automatic checks before publication:

• category and industries taken from a closed list, score bounded from 1 to 10, lengths capped;
• source content is treated as material, never as an instruction (protection against planted instructions);
• original titles, links and dates come from the source, not from the model;
• before any change of prompt or model, a test set checks, among other things, that no number missing from the source appears in the summaries.

Corrections

An inaccurate summary, a misattributed source, a misreading? Write to contact@brefai.fr with the link to the story. We check against the source, then correct or remove the summary.

Independence

BrefAI is not tied to any of the companies it covers. No article is sponsored, no source pays to appear in the feed or to rank higher, and we sell neither subscribers' addresses nor targeted advertising.

Frequency

The feed is updated every 30 minutes, day and night. The Morning Brief goes out every morning, the weekly review on Fridays. Events and regulation are refreshed every day.