AI Policy
What a human writes, what a machine touches.
Last updated: August 2026
1. Who writes what
- Human Written or decided by a person, with no machine input on substance.
- Human, assisted Written by a person. A model may have helped upstream (finding sources, suggesting variants); the final choice is human.
- AI, reviewed Produced by a model from material chosen by a human, then read and approved before publication.
True of everything below, without exception
- Human The choice of topics and sources. Entirely human. Nothing enters an edition through automatic selection or algorithmic ranking of the news. Press articles, wire copy, releases and information sent in by newsrooms are picked one by one.
- AI, reviewed Proofreading. Every text is proofread before publication, with help from a model that flags inconsistencies, grammar mistakes and typos. Corrections address form, never substance.
| What you are reading | Who | How it works |
|---|---|---|
| Brief | AI, reviewed | The topic and the source are picked one by one by a person. A model then condenses the source into a brief. The resulting text is proofread before publication. |
| Note | AI, reviewed | Written by a model, to keep the format consistent. Proofread before publication. |
| Explainer, definition | AI, reviewed | When the content is purely factual and a human rewrite adds nothing, it is written with a model, then checked. |
| Headline | Human, assisted | Written by a person. A model may suggest variants; the final choice is always human. |
| Analysis, editorial, investigation | Human, assisted | Written by a person. A model may help upstream to find or cross-check sources, never to write the text in their place. |
| Illustration | AI, reviewed | Generated by a model. It depicts no real scene and is never presented as a photograph. Some are animated into video from the still image. |
| Podcast | AI, reviewed | When an edition offers one: the text is adapted for audio by a model, then read by a synthetic voice cloned from that of a member of the newsroom. It is not an invented voice. |
Editions in several languages
When a publication appears in several languages, not everything follows the same path:
- Briefs are translated by a model from the original edition, then proofread before sending.
- Long-form content is, depending on the publication: written directly in the language by a person; translated by a model; or adapted by a model trained on our own human translations, so that the tone stays ours.
- Which path each publication takes is set out in the next section.
2. Publication by publication
The framework above applies to all our publications. Here is what sets each one apart.
No machine-generated text. Briefs and long-form content are written by a person. A model is used only upstream, to find and cross-check sources, and downstream, for proofreading. The illustrations, however, are generated.
Briefs condensed by a model from manually selected sources. Analyses, editorials and headlines written by a person.
A single edition is written, then carried into the other languages: briefs are translated by a model, long-form content adapted by a model trained on our own human translations, to keep the newsroom's tone. Everything is proofread before sending.
Briefs condensed by a model from manually selected sources. Analyses and long-form content written by a person.
Briefs condensed by a model from manually selected sources. Analyses and long-form content written by a person. The versions in the other languages are translated by a model, briefs and long-form alike, without adaptation. Proofread before sending.
Briefs condensed by a model from manually selected sources. Analyses and long-form content written by a person.
Satirical publication. The texts do not report real facts and must not be read as news. They are written by a person, with the same proofreading as our other titles.
3. What we never do
- Publish a text no human has read.
- Let a machine invent a quote, a testimony or a figure: every quote comes from an identified source.
- Present a generated image as a real photograph.
- Reply automatically to a reader or a newsroom. An email sent to a newsroom is read and handled by a person.
- Pass our subscribers' personal data to an AI model provider, or use it to train one.
4. The models we use
We build no models of our own. We use these, as customers:
| Use | Provider | Models |
|---|---|---|
| Condensing, assisted writing, translation, proofreading | OpenAI (United States) | GPT-4.1, GPT-4o, GPT-4o mini, GPT-5, GPT-5 mini |
| Finding and cross-checking sources | OpenAI (United States) | GPT-4o Search, GPT-4o mini Search |
| Condensing, assisted writing, translation, proofreading | Anthropic (United States) | Claude Opus 4, 4.5 and 4.7, Claude Sonnet 4 and 4.6, Claude Haiku 4.5 |
| Illustrations | Midjourney | Midjourney |
| Animating illustrations into video | Seedance | Seedance 2.0 |
| Podcast voice | ElevenLabs | eleven_v3 |
This list changes as our tools change. It is updated whenever they do.
5. Report an error
A text condensed by a machine can contain an error that proofreading did not catch. If you spot one, write to us: redaction@meelton.fr. A human will read your message and reply.