Editorial Standards & AI Policy

Effective Date: April 24, 2026

1. What The Masthead Is

The Masthead is an AI-assisted news curator. We aggregate publicly-available RSS feeds and podcast feeds from trusted independent publishers worldwide, then use artificial intelligence to generate summaries, translations, categorizations, and cross-source perspectives.

We do not conduct original reporting. Every story on The Masthead originates with a licensed source publication, which is credited and linked on every article page.

2. Where AI Is Used

The following elements on our article cards and digest entries are generated by AI models (OpenAI and Anthropic):

  • Translated headline — the original-language headline is preserved beneath; the AI version above is for readers in your language.
  • WHAT HAPPENED — a single 25-word fact summarising the news, written in your language and grounded only in the source.
  • MASTHEAD VIEW — a 25–55-word analytical breakdown of the story (significance, pattern, what to watch — not opinion), marked with an “AI-generated analysis” disclaimer on every card.
  • Digest synthesis (5-section “In Depth”) — a 200–300-word synthesis written by Claude Haiku across the primary source, our cross-publisher perspectives, and 1–3 fresh web searches the model performs to surface predecessor events and alternative framings.
  • Translations — AI-translated card content in your language, cached per-language so repeat readers get instant results.
  • Topic tags, categories, and curation score — AI-classified for ranking and personalisation.
  • Cross-source perspectives — AI-matched articles from other publishers/regions that cover the same event from a different angle.
  • Story arcs — AI-clustered multi-day developments (a war, a trial, a regulatory process) with a short narrative covering what changed since the last update.
  • Audio briefings & podcast summaries — Claude-written script rendered to audio via text-to-speech and cached as MP3.

Original publisher content — the full article body — is not reproduced on The Masthead. We link users to the source publication to read the full story.

3. AI Transparency

Every card carries explicit AI markers: the publisher name and country are shown above the headline, the original-language headline is displayed beneath any translation, the analytical-breakdown section is labelled “MASTHEAD VIEW” with an “AI-generated analysis” disclaimer beneath the text, and a “See original” button is always one tap away. Digest entries open with a five-section synthesis labelled as such; the source publishers are credited individually.

AI output can be imperfect. The MASTHEAD VIEW is an analytical layer — it explains significance, pattern, and what to watch, but it is AI-generated and not a substitute for the original reporting. We encourage readers to click through to the source for the full story, nuance, and context that a short breakdown cannot capture.

4. Editorial Oversight

The Masthead is operated by Dmytro Bohdanov, sole proprietor registered in Poland (NIP: 8992981145). Editorial responsibility for the platform — model selection, prompt design, source inclusion, categorization rules, and correction handling — rests with the operator.

Editorial oversight is exercised through automated quality gates and periodic editorial review of representative samples. The automated gates include language-consistency checks, a taxonomy validator, a circuit-breaker that pauses AI calls after repeated failures, and per-source profile tracking. Outputs that fail validation are discarded rather than published. Sampled review feeds back into the prompts, the taxonomy, and the source-inclusion list.

5. Source Attribution

Every article on The Masthead shows the name of the original publisher and a direct link to the original article. We do not claim authorship of any reporting. All copyright remains with the original publishers.

If you are a publisher and you do not want your feed aggregated by The Masthead, contact us at editorial@themasthead.app and we will remove your feed within five business days.

6. Corrections & Accuracy

If you spot a factual error in an AI-generated summary on The Masthead, email editorial@themasthead.app with the article URL and a brief description of the issue. We aim to investigate within three business days.

Depending on the nature of the error, we may:

  • Regenerate the summary using a newer AI model;
  • Remove the article from our aggregation;
  • Apply a prompt or taxonomy fix that prevents the same class of error;
  • Publish a visible correction note, where the error was material.

We do not edit or correct the original publisher's content; errors there should be reported to the source.

7. What We Do Not Do

  • We do not publish AI-fabricated news stories. Every item links back to a real source publication.
  • We do not alter the original source article. Our cards reference it; readers tap through to the publisher's page to read the full reporting.
  • We do not generate or rehost photography on article cards. Featured images come from the original publisher&apos;s RSS or Open Graph metadata and are hot-linked directly from the publisher&apos;s CDN — we do not store image bytes on our servers. Story-arc cover illustrations <em>are</em> AI-generated (the only AI imagery we produce); they are clearly displayed as arc artwork, not as photographs of real events.

8. Content Policies

We exclude feeds that regularly publish hate speech, harassment, sexually explicit material, medical misinformation, or content that violates applicable law. If an article from an otherwise acceptable source triggers these concerns, we remove it from our aggregation on report.

Pipeline Transparency

These are the editorial flows that produce what you see in The Masthead. We document them here so readers understand exactly what each pipeline does, where AI is involved, and what gets surfaced.

Feeds (RSS Ingestion)

Every five minutes a cron fetches our RSS feed roster across many countries. We parse the standard RSS/Atom payload — title, summary, link, image — and store the article URL, the publisher-provided body excerpt, and any image URL the feed declares (media:thumbnail, enclosure, or og:image as fallback). We do not store the publisher's full article text on our servers. We do not rehost article images — the image URLs in the app point directly to the publisher's CDN. When the RSS body is too thin to support a useful card, the AI runs an additional search on the article's topic to gather context — predecessor events, related coverage, alternative framings — and folds it into the prompt below. The retrieved context is used only to inform the AI; it is never republished on the card or website.

Card Generation (WHAT HAPPENED + MASTHEAD VIEW)

Within five minutes of ingestion, an approval cron passes new articles to OpenAI (gpt-5.4-nano) in batches of 25. The model produces four things per article: a translated headline; a single 25-word "WHAT HAPPENED" fact in the user's language; a 25–55-word "MASTHEAD VIEW" analytical breakdown explaining significance, pattern, and what to watch (not an opinion or position); and tags + a curation score. The breakdown is restricted to general knowledge plus claims grounded in the bullet — the model is forbidden from fabricating numbers, dates, names, or quotes that aren't in the source. When a story has no broader structural angle, the breakdown is omitted and only the WHAT HAPPENED fact is shown. A separate validator checks each output is in the expected language; mismatches are discarded and the article is held for the next pass.

Translation

When you open the app in a different language than the article's source, we translate the headline, the WHAT HAPPENED fact, and the MASTHEAD VIEW breakdown via OpenAI on first request and cache the result per-language in our database. The next reader in that language gets the cached translation instantly. Translations preserve meaning, not surface form. The legacy single-paragraph summary that older clients read is no longer generated — only translated when present in old data.

Digest (Daily 5-Section Synthesis)

A digest entry is a synthesis piece, not a summary. The hourly digest cron picks up to 10 of the day's top-scoring stories — but only stories that have at least one perspective (another publisher covering the same event from a different angle, detected by Claude Haiku in our perspective-detection step). Stories without cross-coverage are not eligible. When you open a digest entry, Claude Haiku composes a five-section editorial — THE NEWS, WHAT PRECEDED, WHAT IT MEANS, WHAT'S UNCLEAR, IMPLICATION — combining the primary article, the in-app perspectives we already indexed, and 1–3 fresh web searches the model runs to surface predecessor events, expert reactions, and alternative framings. The result is 200–300 words. The model is forbidden from reproducing source lists, mirroring sentence structure, or writing "according to [source]" paraphrase — the digest entry must be analytical narrative, not a recap.

Morning Briefing

Every hour a cron checks which paid users are at their chosen briefing hour in their local timezone, and for each it composes a personalised audio briefing. Claude Haiku writes the script from that user's top digest stories plus their interest profile (built from likes, dislikes, reads, and click-throughs with a daily-decaying weight). The script is rendered to audio via Google Gemini TTS (with OpenAI TTS as fallback), the audio is cached in S3, and a push notification points the user at it. Every word of the briefing comes from the AI script — no human voice, no recorded segments.

Story Arcs

An "arc" tracks a developing story across days — a war, a trial, a regulatory process. The daily arc cron clusters tagged articles by topic-similarity, asks Claude to confirm or merge candidate clusters with rolling LLM-match calls, and writes an arc card with a short narrative covering what changed since the last update. Articles already published on cards continue to be visible individually; arcs are an additional surface that follows the trajectory rather than the latest headline.

Podcasts

Podcast feeds are fetched twice an hour (xx:15 and xx:45). The episode summary readers see is generated by Claude Haiku from the publisher's RSS show notes — the description, episode title, and any chapter list the feed publishes. We do not transcribe the episode audio. We do not host the original audio either; the player links straight to the publisher's feed URL. Trending-podcast discovery runs once a day via a separate cron; it surfaces newly active podcasts in your interest tags.

9. Contact

Editorial questions, correction requests, or source-removal requests: