AI for media and publishing

AI for media and publishing: the archive is the asset, the rights are the risk

Publishers keep asking how AI can write more. The better question is why decades of already-paid-for material are effectively unfindable.

Start with retrieval, not generation

A media business with thirty years of output is sitting on the most defensible thing it owns, and usually cannot search it properly. Headlines are indexed, bodies are half-indexed, images are indexed by filename, audio and video are not indexed at all. Journalists re-report what a colleague covered in 2017 because finding it would take longer than redoing it.

Semantic search across the full archive, including transcribed audio and video, changes what the newsroom can do in a day. Ask for previous coverage of a company, every interview where a person discussed a topic, or images from a location in a date range, and get results with the original item attached. This is unglamorous compared to generated copy and it is worth considerably more.

Rights and licensing metadata is the harder half

Finding an asset is only useful if you know whether you can use it. Most archives carry incomplete rights data: agency images with expired licences, freelance contributions with territory limits, music cleared for one platform and not another, contributor agreements that predate digital distribution entirely.

A model can help reconstruct this by reading contracts and licence documents and extracting the fields that matter, including rights holder, permitted use, territory, term and any credit obligation. What it must never do is guess. An asset with unresolved rights is surfaced as unresolved, with a link to the document, and is blocked from reuse workflows until a human clears it. Systems that guess here create legal exposure that dwarfs any productivity gain.

Where generation is legitimate

• Format conversion: turning an existing, reported piece into a newsletter version, a social cut or a summary, from material your own journalists produced.
• Metadata and SEO: headlines variants, structured data, alt text, tags and internal links across a back catalogue that nobody will ever hand-optimise.
• Translation drafts: for review by a speaker, not straight to publication.

What does not work is generated original reporting. It carries no source, cannot be stood behind, and one incident does more brand damage than a year of efficiency saves.

The disclosure line

Decide your policy before the first piece ships, publish it, and apply it consistently. A workable line: material that is substantially machine-generated is labelled, material where a model assisted a human-reported piece is not, and no byline of a real person ever appears on something that person did not write and approve. Readers forgive the use of tools. They do not forgive discovering an undisclosed one.

The build order

Archive retrieval first, because it is safe and immediately useful. Rights metadata second, because it makes the archive commercially usable. Generation last, narrowly scoped, on your own material, with disclosure settled in advance. Publishers that reverse this order tend to get a short traffic bump and a long correction.

An AI audit maps what is in the archive, what state the rights data is in, and which of the three stages you can realistically start with.

Frequently asked questions

Related