Our commitment to accuracy, transparency, and responsible AI-assisted journalism.
Last updated: October 28, 2026
Ajako Taja — what every outlet is covering today — exists to bring the stories that many outlets are reporting together in one place, so readers don't miss important news just because they can't visit every outlet, and to present each one as an accurate, relevant summary in a few minutes' reading. We use AI as an editorial accelerator — not a replacement for journalistic responsibility. Our mission is to give readers a factual, balanced summary of events without noise, bias amplification, or misinformation.
We are guided by the principles of the International Fact-Checking Network (IFCN) and aim to follow Google News Publisher policies. We disclose our AI usage, acknowledge limitations, and prioritize corrections over self-preservation.
Our content pipeline is fully transparent. Every article combines automated steps with human review — our team chooses which stories are covered (stories already reported by multiple outlets in our sources can be selected automatically, and the team can reject them before anything is written), and a person fact-checks each article before it's published:
Our system collects stories from news feeds every hour and groups reports of the same event from different outlets into one topic. Depending on the topic area, an AI model reads the latest reports and groups them by the real-world story they describe, or a keyword-matching method does. In every topic area, an AI same-story check then compares each new story with those already covered in the last 48 hours — using AI text embeddings plus a direct AI comparison — so the same event reported with different wording, or in another language, is recognised as one story. Each topic gets a confidence score (how many outlets cover it, how consistent their reports are, how recent they are, and whether they name the same people and places) and a velocity score (how quickly coverage is growing). To see how widely a story is reported beyond our own feeds, we also check public news search (Google News) for other outlets reporting the same event in the last 48 hours; well-established outlets count more. This only affects how a story is ranked for review — those outlets are not used as sources for the article, and our team discards results that turn out to be a different story. These scores help our team prioritise. In some topic areas, stories already reported by multiple outlets in our sources are approved automatically; a member of the team reviews and approves the rest, and can reject any story before it's written. Only approved topics are sent to the AI.
For each qualifying topic, our AI receives what each outlet reported, so it can see where they agree or differ. Most published stories are covered by two or more outlets; when only one outlet has reported a story, the article attributes its facts to that outlet. Every article lists the outlets it draws on under “How outlets covered it”, with links.
Occasionally our team adds a story that no outlet has reported, from information it has gathered itself. That article is drafted only from the team’s own notes, goes through the same automated checks against those notes and the same human fact-check before publishing, and has no “How outlets covered it” list, because no outlet is involved.
A large language model (currently Google Gemini) reads each outlet's report — the RSS summary and, where the outlet allows it, the article text itself — and drafts an article in its own words: a short lead, the key details and, where the reports give it, background. The article text is read only to write from: it is never stored or republished. We read an outlet's page only if its robots.txt allows our bot (AjakoTajaBot) and the outlet hasn't asked us to stop; otherwise the article is written from its RSS headline and summary. We never get around paywalls or bot protection. The model is instructed to:
Separately from article content, our AI generates SEO titles, meta descriptions, and relevant hashtags. These are editorial metadata only and do not alter factual content. The “How outlets covered it” list itself is attached by our system from the outlets grouped into the story — it is not written by the AI.
Some articles may be data reports — for example football results, previews of upcoming fixtures (kick-off times in Nepal time, which can change; never predictions), or end-of-day market prices — written from a data provider's figures, or from a data file our team uploads (such as end-of-day Nepal Stock Exchange prices), rather than from outlets' reports. Uploaded files are checked before they are used: rows with missing or inconsistent numbers are left out. Our team chooses the competitions or teams to follow and which results to write up (a followed source can be set to send every result to the writer). The AI receives only the provider's figures and is told not to add players, quotes or context from memory. Market reports describe end-of-day price moves only — never causes, predictions or buy/sell advice — and are marked "Not investment advice". Before the draft reaches an editor, our system checks that every number and every statement in it comes from the data (no background, causes, comparisons with other periods or outside news); if anything doesn't, the AI gets one chance to fix it, and otherwise the draft is held for an editor. The provider is credited in the article's data panel together with the time the data was taken, and a person publishes every data report, as with all our articles. When a news story about the same event already exists, our team can add the data panel to that story instead of publishing a second article. The provider's raw responses are kept privately to show where an article's figures came from.
Featured images are not AI-generated. By default, each story gets a free-license stock photo from Pexels, found with a generic search phrase describing the setting (never a person's name), and credited under the image. Our team reviews and replaces unsuitable images. When no photo fits, a brand graphic is rendered in code. Stock photos are illustrative and don't show the actual people or events in the story. Featured images never use news outlets' photos. The “How outlets covered it” list on an article shows each outlet’s name, logo and headline with a link to its article. It doesn’t show the outlets’ photos.
Our team chooses the feeds that supply each topic area by hand. When adding a source, we look for:
We remove feeds that stop meeting these standards. Removing a source does not change articles already published from it — those keep their source credit.
Before publication every article passes through automated checks:
Articles that fail a check are kept as drafts and flagged to the reviewer, and aren't published unless the reviewer fixes the problem. Passing these checks doesn't replace the human fact-check in Section 5 — every article gets both.
Large language models can generate plausible-sounding but false content ("hallucinations"). We mitigate this by constraining the model to summarize only content provided in the prompt context — it is instructed not to add facts, context, or analysis the sources don't support. This reduces the risk but can't eliminate it, which is why every article links to its sources.
We do not send reporters to scenes. We do not conduct original interviews. We cannot independently verify claims made exclusively by a single source that our pipeline cannot cross-reference. In such cases, we clearly attribute the claim to its origin rather than stating it as established fact.
Every article is reviewed by a member of our team before it's published. Nothing on https://www.ajakotaja.com is published automatically. Our review process:
Reviewers check articles against their sources; they don't independently re-report the story. If an outlet we cite gets something wrong, that error can still reach us — which is why every article links to its sources.
Our editorial team at Guided Labs Private Limited (Lalitpur, Nepal) is responsible for pipeline configuration, source curation, and policy enforcement. Escalation contact: ajakotaja@gmail.com.
Readers can report factual errors by emailing ajakotaja@gmail.com with the article URL, the specific inaccuracy, and, where available, a primary source link supporting the correction.
Corrected articles are updated in place, show an "Updated" time, and the article's dateModified timestamp reflects the change. For significant corrections we add a short note to the article explaining what changed. Full details: our Corrections Policy.
If an article is found to be wholly inaccurate (e.g., based on a fabricated source or a source that issued a full retraction), the article is unpublished.
We will not knowingly publish or amplify:
When authoritative guidance on a fast-moving topic (e.g., a new health advisory) is unclear or contested, we present the range of expert positions rather than asserting a single conclusion.
News articles are time-sensitive. Our content freshness policy is:
Our editorial pipeline operates independently of our advertising and monetization systems. Specifically:
Every article on Ajako Taja includes:
For full disclosure of how AI is used across our platform, see our AI Content Disclosure.
Guided Labs Private Limited is the publisher of record for Ajako Taja and bears full responsibility for editorial decisions made on this platform. Read more about our team and mission.
We welcome external audits of our editorial process and will cooperate with recognized press oversight bodies, fact-checking organizations, and regulatory authorities upon legitimate request.
For editorial inquiries, corrections, or source removal requests:
