AI in Media and Publishing

Publishing’s exposure is not that a model writes badly. It is that a model writes something attributable — to your masthead, and sometimes to a named person who never said it.

A disclosure that carries information

“This article was produced with the assistance of AI” tells a reader nothing they can use. It covers a piece that was drafted wholesale by a model and a piece where a reporter used transcription, and those deserve different reactions. A label that does work distinguishes what was automated:

• Generated and reviewed. A model produced the article; an editor checked and published it. The reader is being told a machine chose the sentences.
• Assisted drafting. A journalist wrote it, using a model for structure, headlines or tightening. The claims and the reporting are the journalist’s.
• Automated from structured data. Results, filings, fixtures. Templated or generated from a feed. Long-standing practice and uncontroversial when labelled.
• Translated. The piece was reported and checked in another language and machine-translated. What a reader should know is which version is authoritative.
• Illustrated synthetically. The image is not a photograph of anything. This is the one readers care most about and the one most often unlabelled, because image pipelines and article metadata are separate systems.

The point is not any particular set of labels — house style will differ — but that a taxonomy is decided once, applied by a field on the content model, and rendered consistently, rather than left to a per-story judgement made at eleven at night.

The fabricated quote

Among everything a model can get wrong in a newsroom, one failure is categorically different: attributing words to a real, named person who did not say them. That is not a quality problem to be traded off against speed. It is the thing defamation actions are made of, and it is a failure mode that generation produces naturally, because a quote is a highly probable continuation of a paragraph that has just introduced a source.

It is also, unusually, fully eliminable by a check that is not a model. Any quoted string in generated output must appear verbatim in the source transcript or document; if it does not, the piece does not render. That is a substring comparison after normalising quotation marks and whitespace, it runs in milliseconds, and it converts an unbounded risk into a build failure. The same check applies to summarisation of press releases, court filings and earnings calls, and it is worth wiring into the CMS rather than the prompt — see fabricated citations, which is the same failure wearing a different hat, and output validation.

Paraphrase needs a different treatment, since it will not match verbatim. There the workable rule is that a paraphrase must carry a span pointer into the source, and the editor sees both. This is the same citation discipline as contract review, arrived at from a completely different direction.

Provenance survives until your CDN resizes it

Content provenance standards attach signed metadata to a file recording how it was made and what was done to it. They are a genuine advance and they have an unglamorous failure mode that publishers hit immediately: the metadata lives in the file, and a publishing pipeline is a sequence of programs that rewrite files.

An image ingested with provenance intact passes through a resize, a format conversion to a modern codec, a quality re-encode at the CDN edge, and a social card generator. Most of those strip or invalidate the signature by default, so an asset that arrived verifiable is served unverifiable, and nobody notices because nothing errors. If provenance is part of the plan, the test is an end-to-end one — fetch the delivered asset from the public URL, at each rendition, and verify the manifest — run in CI rather than assumed from the ingest side.

A correction cannot catch a syndicated copy

Publishing already has a correction workflow, and it is built on an assumption that quietly stops holding at volume: that the canonical copy is the copy people read. Syndication partners, aggregators, newsletters, scrapers and now model training sets take copies, and a correction applied to the original propagates to approximately none of them.

That is an argument about where to spend the checking effort. For a handful of pieces a day, post-publication correction is a reasonable safety net. For a volume of automatically generated output it is not a net at all, because the error is distributed before anyone reads the first copy. The honest conclusion is that automation should be applied where pre-publication verification is cheap and mechanical — the verbatim quote check, a figure that must match a filing, a fixture list — rather than to the volume of judgement-heavy copy where verification is the expensive part.

The uses that hold up

The applications that survive editorial scrutiny share a property: the output is checkable against something that exists, or its errors cost a reader nothing.

• Transcription with timecodes. Interviews and press conferences, searchable, with the audio one click away. This is the single largest time saving available to a newsroom and nobody disputes it.
• Archive enrichment. Tagging decades of back catalogue against a controlled vocabulary, so that the archive becomes usable. Errors are recoverable and the alternative is that the archive stays dark.
• Headline and standfirst variants for a piece a human wrote, tested the usual way. The claims are already checked.
• Alt text and audio versions, expanding who can read the work. Reviewed for images where the description carries the story.
• Translation of an already-verified piece, with review proportional to what the piece asserts.

One rights caveat runs under all of it. An archive’s photographs and wire copy are usually licensed for specific uses, and those licences were written before anyone contemplated derivative generation or bulk processing by a third-party service. Whether a given archive may be sent anywhere is a contracts question, asset by asset, and it is the question that stalls archive projects rather than the engineering.