AI Content Disclosure

• Our system continuously collects stories from news sources and feeds across each topic area.
• AI works out which reports from different outlets are about the same real-world story — even when they use very different headlines or are in another language — and groups them together, so each article draws on several outlets' coverage instead of repeating the same story many times.
• Each grouped story is scored on how many outlets cover it and how quickly coverage is growing.
• Stories already reported by multiple outlets in our sources can be approved automatically; a member of our team reviews and approves the others, and can reject any story before it's written. Nothing is written about a story until it's approved.
• An AI language model reads the outlets' reports — including the text of their articles, which we read but never republish — and drafts an article in English, in its own words: a short lead, the key details, background, and why it matters. It's told to use only what the sources say, and to leave a section out rather than pad it. It also drafts the short headline printed on the article's share image, which is fact-checked with the rest.
• The draft passes automated quality and content-policy checks (graphic violence, adult content, hate speech, or promotion of drugs, weapons, or gambling).
• A member of our team fact-checks the draft against the sources — headline, names, numbers, dates, and attribution — edits it if needed, checks the image, and only then publishes it.

• AI summaries distill several sources into one clear read, which saves time but can lose nuance.
• The summary tends to reflect whichever source articles were most prominent — not necessarily the most balanced view.
• Satire, irony, and context-dependent language can be misread by AI.
• AI occasionally introduces small factual errors not present in the original sources.

Our team decides which published articles to share on Ajako Taja's own social media accounts. The captions and headlines for those posts are then written by AI — in English or Nepali, depending on the channel — from the reviewed article, and posted without a further manual check. For data reports, our system checks every number in each caption against the data, and replaces a caption that fails with the article's headline. If a social post and the article ever differ, the article on https://www.ajakotaja.com is the authoritative version.

Some articles — for example football match results, previews of upcoming fixtures, or end-of-day stock prices — may be written from a data provider's figures, or from a data file our team uploads (such as end-of-day Nepal Stock Exchange prices), instead of news outlets' reports. These are labeled "Data report" and show the figures in a data panel with the provider's name and the time the data was taken ("Data as of"). A member of our team chooses which results to write up (or sets a followed competition to be written up automatically); the AI is given only the provider's figures and told not to add anything else, and our system checks that every number and every statement in the draft comes from the data — no background, causes, comparisons or outside news — and a draft that fails the check is held for an editor. As with every article, a person reviews and publishes it; nothing is published automatically. Market reports describe end-of-day price moves only: the AI is told not to explain why a price moved, not to make predictions, and not to give advice, and each carries a "Not investment advice" note. The data comes from the provider and may contain its errors or delays.

• Many images are free-license stock photos from Pexels, chosen by search terms describing the story's general setting (e.g. "cricket stadium"). The photographer is credited under the image.
• Our team can replace any image with a photo we have the rights to use. We don't add watermarks to photos.
• When no suitable photo is available, the image is rendered in code: a gradient in the topic's brand color, our logo, and the names of the outlets that reported the story.
• No AI image-generation service is used.

• Stock photos are illustrative: they don't show the actual people, place, or event in the story.
• Generated images are abstract brand graphics, not depictions of real events, people, or locations.
• Please don't treat any article image as photographic evidence of what happened.

• Hallucination — AI language models can produce plausible-sounding but factually incorrect information.
• Source skew — if the source articles contain an error, the AI summary may reproduce or amplify it.
• Nuance loss — complex geopolitical, scientific, or cultural topics can be oversimplified.
• Bias — AI models reflect biases in their training data; we work to mitigate this but can't eliminate it.

• We show how many outlets covered each story, and link to them on every article so you can verify for yourself.
• The AI is instructed to stick to what the sources report and not to add outside claims.
• A member of our team fact-checks every article against its sources before it's published.
• We invite users to report inaccuracies — see the section below.
• Confirmed errors are corrected, or the article is removed.

Our team reviews the stories our system finds each day and approves the ones worth covering. In some topic areas, stories already reported by multiple outlets in our sources are approved automatically, and the team can still reject them before anything is written. Rejected stories are never written.

Each AI draft is checked automatically before it reaches a reviewer: a second AI pass compares every claim with the outlets' reports and lists anything they don't support (especially ages, dates, figures, titles and records, which AI models often get wrong from out-of-date memory). The draft is revised once to remove those claims; anything still unsupported keeps it as a draft. Drafts are also checked for length, hype or clickbait wording, prohibited content, and whether the story is a current news report rather than opinion, satire, a press release or an old story.

A reviewer compares the draft with the source reports — headline, names, numbers, dates, and who said what — corrects anything that doesn't match, confirms the image is suitable, and publishes it. A person, not the system, makes every publishing decision.

Our reviewers check the article against the sources it's based on. They don't independently re-report the story, so an error in the original reporting can still reach us — which is why every article links to its sources.

When a factual error is confirmed — either by our team or via a reader report — we correct the article or take it down.