How our AI monitoring is produced

1. Purpose

To keep the association’s members, public authorities and partners informed about personal care news in Cameroon and worldwide. The system makes no decision about anyone, scores or profiles no one, and processes no health data. It gives neither medical nor legal advice.

2. Status under Regulation (EU) 2024/1689

The association is Cameroonian; it nevertheless applies the EU Artificial Intelligence Act, because its content is read in the European Union and as a matter of rigour. In our analysis, which is not legal advice:

• No prohibited practice (Article 5): no manipulation, no social scoring, no biometrics.
• Not a high-risk system (Article 6 and Annex III): news monitoring falls under none of the listed areas.
• Article 50 transparency obligations, applicable since 2 August 2026 (Article 113):
  • Text published to inform the public (paragraph 4): every brief, the ticker and every email state that they are AI-generated. We do so even when a board member has reviewed the text, although the Regulation then waives this disclosure.
  • Machine-readable marking (paragraph 2): there is no single standard for text yet. We combine the data-ai-generated attribute in pages, JSON-LD structured data carrying the IPTC digital source type trainedAlgorithmicMedia, the genere_par_ia field of our API and, in every email, x-ai-generated and digital-source-type metadata embedded in the message.
  • Clear information at first exposure (paragraph 5): the “AI” label is always visible in the ticker, at the top of each brief and at the top of each email.
• AI literacy (Article 4): board members who run the monitor must read the operating notice before they can open mailings or approve a newsletter; this acknowledgement is recorded.

3. Production chain

1. Collection every 20 minutes from the sources listed below. The robot identifies itself (Veille3A/1.0) and honours robots.txt and text-and-data-mining reservations (TDMRep protocol): a source that opts out is not read. Only the title, link, date and an excerpt of at most 500 characters are kept; no article is reproduced.
2. Sorting in two stages: a weighted lexicon discards off-topic items, then a language model rates relevance, section and geographic scope, and writes a one-sentence summary.
3. Writing of the brief at the start of each three-hour slot (midnight, 3 am, 6 am… Yaoundé time).
4. Automatic checks on the generated text: every statement must cite a collected item; any number missing from the cited sources gets the statement rejected; geographic scope comes from the sources, not the model; if nothing passes, the brief only uses the already-verified summaries.
5. Human oversight: the board can withdraw any brief, require approval of every brief before publication, and by default approves the monthly newsletter before it is sent. It can suspend collection, publication or mailings at any time.
6. Email distribution only to people who have consented, checked again at the moment of each sending, with nothing sent between 10 pm and 6 am.

4. Models used

• Loading…

Open-weight models by Meta and Mistral AI, run by Cloudflare Workers AI. No subscriber data is sent to the models.

5. Human oversight in force

• Loading…

6. Activity over the last 30 days

7. Known limits

• A summary may distort the meaning of an article or leave out an important nuance.
• Few Cameroonian outlets publish a machine-readable feed: national coverage is thinner than international coverage.
• Institutional and English-language sources are over-represented.
• French data are comparison points; they do not describe the situation in Cameroon.
• An item may appear several hours after publication.

8. Personal data and rights

To produce the monitor, the system processes no personal data other than names appearing in public articles. Subscribers are enrolled on the basis of their consent, collected without pre-ticked boxes and evidenced by a time-stamped log; withdrawal takes one click from any email. Rights of access, rectification, erasure and objection under Cameroonian Law no. 2024/017 and, for EU residents, the GDPR: [email protected].

9. Reporting an error

Each brief has a “Report an error” link. The board reviews the report and withdraws the brief if needed; the number of withdrawn briefs appears in the activity figures above.

10. Sources read