AI & translation policy

The plain statement

baba News uses generative AI systems to write, translate, rate, illustrate and voice its content. No person reviews that content before it is published. Text marked as “by baba”, images marked as illustrations by baba, the voices of The Desk and of the Good Morning, Israel podcast, and the answers of Ask Duki are all generated by AI. The newsroom’s own article, in its own language, is the authority for what it reported. We link to it from every digest.

What AI writes

For each article we collect, a language model reads the article and writes:

• A headline in English and a headline in Hebrew. When a Hebrew article appears in the Hebrew edition, we keep the newsroom’s own headline.
• A digest in our own words. It is a few short paragraphs and never longer than the article. For a report that was published in English, we show key points and the link, and no digest of our own.
• A summary of a few sentences and up to seven key points.
• The people, organizations and places named, a topic, a story type and the main claims of the report.

The model is told to keep a neutral tone, to add no commentary, to add nothing that is not in the article and not to copy the article’s sentences. It still makes mistakes.

For a story that several newsrooms reported, AI systems also produce:

• The grouping of reports into one story.
• The summary at the top of the full coverage page.
• The comparison of the newsrooms’ headlines.
• “Who reports what”: which newsrooms’ reports include each claim.
• The briefs, on the site and by email.

In baba’s own voice, AI systems write the scripts of The Desk and of the podcast, our posts and videos on social media, and the answers of Ask Duki. These follow a house style, which our editorial standards set out. The house style does not apply to the digest of a newsroom’s article.

On the election pages, a language model reads poll figures out of published reports. Software then checks that every seat number it found appears in the report. The average we show is a simple average of recent polls, not a forecast.

What is rated, and by whom

• By people: the list of sources, and each newsroom’s group and lean. We assign these by hand. They are our editorial opinion.
• By an AI scoring system: the tone of each report (how positive or negative, and how heated) and its stance. The system reads our English digest without the newsroom’s name. It does not read the original article.
• By a formula: the mood index, which combines the tone scores of the day’s stories and gives more weight to stories that more newsrooms covered.
• By comparing our digests: “Who reports what” asks whether each newsroom’s digest contains a claim. It compares our digests and never the original articles, so a gap can come from our digest and not from the newsroom.
• By a rule: Not Everywhere, which counts which groups of the press carried a story. No language model decides it.
• By a statistical model: the Mamad forecast, which estimates the chance of sirens from past Home Front Command alerts and an automated reading of the news. It is not a warning system. When a siren sounds, follow the Home Front Command.

Our methodology describes each of these. They are opinions and statistical readings. They are not findings of fact about any newsroom or person.

Images and audio

• Illustrations. An AI image model draws the picture on a story from the headline and the summary. It is instructed to show no people and no faces. The picture is a symbol of the story. It is not a photograph and does not show the real event.
• Portraits. Portraits of public figures are watercolours made by an AI image model from freely licensed photographs. The photographers and licences are listed on our portrait credits page.
• Real photographs and clips. Photos released by the IDF Spokesperson and Magen David Adom appear with their credit. Newsrooms’ clips and posts appear in the platform’s own player. The one exception is an Instagram Reel attached to a story, which we copy and keep for up to 30 days because Instagram’s player does not load for some readers. None of these is made by AI.
• Audio. The Desk and the Good Morning, Israel podcast are written by AI and spoken by synthetic voices from AI speech systems. Nobody recorded them in a studio.

How AI content is labelled

On baba News, “by baba” means made by baba’s AI systems. These are the labels you will see:

• On an article page: “Translated & summarized from [newsroom] by baba”, or “Summarized from [newsroom] by baba” when the source is in the page’s own language.
• On a full coverage page: “Translated and summarized by baba from [newsroom], the latest report.”
• On a headline comparison: “Comparison written by baba from the newsrooms’ headlines. It describes wording, not which report is right.”
• On an illustration: “Illustration by baba”, or “Editorial illustration generated by baba News, not a photograph of the event.”

Some AI content carries no label of its own. Headlines and thumbnails in lists, tone and lean chips, the mood index, Not Everywhere, “Who reports what”, the briefs, The Desk’s player, the election figures and some of our social posts are not marked one by one. Treat them as machine-made. As a rule: anything on baba News that is not a newsroom’s own headline, a quotation, an embedded post or a credited official photo was produced by our automated systems.

What people do, and what they do not

No person reads a digest, a translation, a rating, an image, a bulletin, a social post or an answer before it is published. baba News follows the whole Israeli press as it publishes, and that volume is beyond what a small team can read.

People choose the sources, assign groups and lean, write the instructions the models follow, set the house style and the method, and read the reports that readers and newsrooms send. A person can correct a digest, have it written again, hide a story or stop collecting from a source.

Some checks are automated. A second system judges whether an article is one we should publish and scores the digest. A second language model from a different family reviews our posts on X before they go out. A vocabulary check rewrites text in baba’s own voice once and then refuses it. These checks reduce errors. They do not remove them.

Where it goes wrong

• A word, an idiom or a name is translated or spelled wrongly.
• A claim comes out softer or sharper than in the original, or its meaning is reversed.
• A quotation is credited to the wrong person. Quotations in a digest are translations and not the speaker’s exact words.
• A number, a date or a place is wrong.
• The digest leaves out something that mattered, or includes something the article did not say.
• Unrelated reports are joined into one story, or one story is split in two.
• The digest is out of date. We go back to articles from the last 24 hours in rotation. A change a newsroom makes later than that is not picked up on its own.
• A newsroom is shown as missing a story it did report, because we did not find or did not match its report.
• A tone or stance score reflects our digest and not the article.
• Duki answers from incomplete coverage, or answers with confidence where the coverage is thin.

For anything that matters, such as a legal, financial, medical or safety decision, or before you quote or republish, read the original article.

How to report an error

Write to news@itsbaba.com with a link to the story, what is wrong and what it should say. In the app, every article has a “Something wrong?” button. We correct the text or have it written again from the original article. Our corrections page describes the process, including what we do not do.

The AI systems we use

We do not build our own models. Digests, summaries, scripts and answers come from language models made by Google, Anthropic and OpenAI, which we reach through OpenRouter. Scoring and checking come from TypeSafe. Illustrations come from an image model hosted by FAL. Speech comes from ElevenLabs and Google. This list changes as we change providers. We send the article’s text to these providers so that they can produce the result. We do not train AI models of our own on newsrooms’ articles or on readers’ questions. Our Privacy Policy says what personal information reaches them.

Transparency under the law

We publish this page so that no reader takes machine-made content for the work of a human journalist or editor. For the purposes of Article 50 of the European Union’s Artificial Intelligence Act: the text baba News publishes about the news is generated by AI systems and is not reviewed by a person before publication, the illustrations and the audio are generated by AI systems, and Ask Duki is an AI system that you interact with directly.

Newsrooms and rights

baba News claims no ownership of a newsroom’s article. The copyright in the original belongs to the newsroom. A newsroom can ask us to stop including it, to remove a story or to correct a digest, at news@itsbaba.com. Our page for publishers and our terms explain how we handle these requests. The terms also set out the limits of our responsibility for AI content.