AI in journalism — assistant, threat, or both · AI, Ethics, and Society

A working journalist in Kantipur, Setopati, Kathmandu Post or Onlinekhabar in 2026 is already using AI, whether their editor says so or not. Transcription of an interview that used to consume two hours of a reporter’s afternoon now takes six minutes. A long English wire report becomes a clean Nepali draft in a single prompt. A tip from a Facebook page in Birgunj is summarised, translated, and routed to the right desk before the reporter has finished their tea. The newsroom has not announced any of this. It is happening anyway, in private tabs, on personal accounts, with no policy and no disclosure. That is the situation, and ignoring it is no longer an option.

The genuine opportunity

It is worth being honest about how useful this technology is in a small newsroom.

A regional paper with five reporters in a district headquarters cannot, on staff, compete with a Kathmandu daily on translation, monitoring, or production speed. With AI, that math changes. The same five reporters can transcribe every interview they record, translate wire copy in both directions between Nepali and English, monitor twenty local Facebook pages and a dozen Viber groups for tips, and produce double the number of clean drafts per day. None of this replaces reporting — the actual reporting, the going-to-the-place, the looking-at-the-document, the asking-the-uncomfortable-question, still has to be done by a human. But it removes a great deal of the friction that used to sit between reporting and publication.

For a country whose journalism is concentrated in a few Kathmandu outlets and a thin scatter of underfunded regional papers, this is genuinely good news. AI lowers the cost of being a competent newsroom. The story of the next decade of Nepali media may turn on whether smaller outlets seize that opportunity.

The genuine threat

It is also worth being honest about the other side.

The same technology that helps a five-person newsroom also helps the people producing low-quality, partisan, or outright fabricated content. The supply of “news” — text, video, screenshot, claim — is going to grow much faster than the supply of trustworthy news. Reader attention is fixed; trust is not. In an environment where it is cheap to produce something that looks like journalism, the price of looking trustworthy goes up. The newsrooms that survive will be the ones that earn trust faster than they spend it.

There is also an internal threat. A reporter who lets an AI write the lede of a story, signs their name to it, and never quite admits the AI was involved has started a small lie. Most of the time the lie is harmless. Occasionally the AI fabricates a fact, attributes a quote to the wrong person, or invents a statistic, and the reporter — having not really written it — does not catch it. When the error is discovered (and it will be discovered), the paper apologises. The reader notices. Repeat that twenty times across a year and the paper’s reputation has quietly eroded.

Disclosed AI use vs. silent AI use

The dividing line that matters most is disclosure.

There is nothing inherently wrong with a journalist using an AI to draft a translation, transcribe an interview, summarise a long PDF, or rewrite a clunky paragraph. These are tools — like spell-check, like a calculator, like a search engine. The question is whether the reader is being deceived about how the work was produced.

A useful rule of thumb is material involvement. If the AI shaped a piece of work in a way the reader would care about — wrote substantial original text that appears under a byline, made an editorial judgement about emphasis or framing, produced a translation that wasn’t checked sentence-by-sentence by the bylined journalist — the reader is entitled to know. If the AI helped with mechanical tasks that the journalist then verified (a transcription the journalist read against the audio, a translation the journalist edited line-by-line, a summary the journalist confirmed from the source) — disclosure is a courtesy, not a requirement.

A small disclosure line at the bottom of a story — “This article was drafted with assistance from an AI tool and reviewed by the reporter” — is enough. It is not a confession. It is a clarification. Readers, in the surveys, are not punishing it.

A workflow that preserves trust

If you are a working journalist, or an editor thinking about a policy, the workflow that holds up under scrutiny looks roughly like this.

AI is a first-draft tool, not a final-draft tool. Use it to get to a starting point faster — a rough translation, a structural outline, a transcription. Do not paste its output into the page unread.

Humans are accountable for facts, quotes, and judgement. Every name, every number, every quotation in the published piece is the reporter’s responsibility. If the AI hallucinated a quote and the reporter didn’t catch it, that is on the reporter, not on the tool.

Disclose when AI shaped the output materially. A translation reviewed line-by-line by a Nepali-fluent editor probably does not need disclosure. A translation that the editor skimmed for sense does. An opinion piece partly drafted by AI almost certainly does. When in doubt, disclose.

Never feed unpublished source material to a public AI without thought. A whistleblower’s documents, a draft investigation, an interview with a vulnerable source — these should not be pasted into a consumer chatbot whose terms of service may allow your text to be used for training. Use an on-device tool or an enterprise contract that explicitly protects the input.

Train the desk, not just the senior editors. The interns and the junior reporters are using these tools the most. A policy that lives in the editor’s head and not in a written guideline does not exist.

What this looks like at Kantipur, Setopati, Onlinekhabar

The Nepali outlets that are likely to come out of this period strongest are the ones that pick a clear, simple, public policy and stick to it. We use AI for these things. We do not use it for these things. When AI shaped a piece of work, we say so at the bottom. Readers do not need the policy to be sophisticated. They need it to exist and to be honoured.

The outlets that lose ground will be the ones that quietly let AI seep through every part of the workflow without acknowledgement — until the first time a hallucinated quote makes it into print, and the apology has to admit that nobody really wrote the paragraph.

Trust, once lost in a small media market, is very expensive to rebuild.

Check your understanding

Quick check

—

A regional Nepali paper uses an AI tool to translate a foreign wire report, transcribe a long interview, and draft the first version of an opinion column. Which workflow best preserves reader trust over the long term?

Use the AI for all three, publish without changes, and do not mention the AI involvement — the readers cannot tell and there is no need to confuse them.
Refuse to use AI for any of these tasks on principle, even when the result would be more accurate and faster, because all AI involvement is inherently misleading.
Use the AI as a first draft, have a Nepali-fluent editor verify the translation line-by-line, have the reporter read the transcript against the audio, have the columnist substantially rewrite and own the opinion piece — and include a brief disclosure where AI materially shaped the published text.
Use the AI but disclose it only when a reader specifically asks, since unprompted disclosure draws unnecessary attention to a routine tool.

What comes next

The journalism question is one half of the trust picture. The other half is the demand side — the campaigns and operatives who use these same tools to shape what voters believe. The next section looks at how AI changes the economics of political persuasion in Nepal, what microtargeting actually means in a country with cheap Facebook ads, and the small set of personal habits that actually work as a defence.