Nine checks for AI content operations under UK copyright law, from text and data mining to moral rights

What to take away

The UK copyright framework that applies to AI content operations

UK copyright law protects original literary, dramatic, musical and artistic works, plus sound recordings, films and broadcasts. It arises automatically on creation, with no registration. That matters for AI content operations because both the material you train on and the material you publish may carry rights owned by someone else.

The Copyright, Designs and Patents Act 1988 is the core statute. It sets out the exclusive rights of the owner: copying, issuing copies to the public, renting or lending, performing, communicating by electronic transmission and adapting. Training an AI system on a protected work normally involves copying, so a licence or an exception is needed.

Ownership rules decide who can grant that licence. A work made by an employee in the course of employment is usually owned by the employer. Freelance contributors keep their rights unless a contract assigns them, and commissioned photography, music and illustration frequently sit with the creator, not the commissioner.

Duration is long. Literary, dramatic, musical and artistic works generally last for the life of the author plus 70 years, so material from the twentieth century can still be in copyright. That is why archive scraping carries risk, and why the age of a source is no proof that it is free to use.

Database rights sit alongside copyright. A database can attract copyright if the selection or arrangement is original, and a separate database right can apply to substantial investment in obtaining, verifying or presenting contents. Both are relevant when you build or buy a training corpus, and both belong in the rules and ethics review.

Text and data mining exceptions and what they do not cover

The UK text and data mining exception is section 29A of the 1988 Act. It permits a copy made with lawful access for computational analysis for non-commercial research. The research must be non-commercial and the access lawful.

Lawful access means a subscription, a purchase, an open licence or a permission. Scraping a website that sits behind a paywall, or breaching terms of use to obtain files, does not create lawful access. Contract terms cannot be used to restrict a copy that the exception permits, but the exception stays confined to non-commercial research.

The exception does not extend to commercial product development, training for a commercial service, or the creation of a searchable database for sale. If your organisation builds or fine-tunes an AI system to power a commercial content operation, you cannot rely on section 29A.

You need a licence, an assignment, material that is out of copyright, or material released under terms that permit the use. The temporary copies exception and the quotation exception were not designed for mass ingestion either, and the quotation exception needs a fair dealing assessment and sufficient acknowledgement.

Opt-outs and machine-readable reservations matter here. Rights holders increasingly publish signals such as robots.txt directives or licensing statements that restrict mining. Ignoring them weakens any argument that your copying was reasonable, and it is the kind of fact that appears early in a dispute.

Policy is moving. The Intellectual Property Office and two government departments ran a consultation on copyright and artificial intelligence from 17 December 2024 to 25 February 2025, and a report and impact assessment have since been published. Until the law changes, plan on the current position: commercial mining needs a lawful basis beyond the research exception.

Data protection law runs in parallel. If training data includes personal data, UK GDPR and the Data Protection Act 2018 apply, and the ICO expects a lawful basis, transparency and, where relevant, a data protection impact assessment. The data protection duties for content teams are covered separately.

Licensing third-party material for AI training and output

Licensing is the cleanest route to third-party material reuse. It converts an uncertain legal position into a documented permission with scope, territory, duration and price. For AI content operations, the licence must be drafted for machine use, not just for human reading.

Start with the collecting societies and licensing bodies that already represent large repertoires. They can grant blanket or transactional licences for text, images and music, and they understand the difference between display and ingestion. A licence that permits display on a website may not permit training, so read the grant clause carefully.

Publishers and platforms increasingly offer direct AI licensing deals. These usually specify permitted uses, output restrictions, attribution requirements and audit rights. Some prohibit content that competes with the source, and some require you to keep training data segregated.

Open licences are not a free pass. Creative Commons terms vary: some permit commercial use, some require share-alike, some prohibit derivatives. Attribution conditions still apply, and an output that reproduces a substantial part of a share-alike work may create obligations you did not intend.

When you commission material, get the rights in writing. The commercial contracts should cover the scope of use, including training and machine learning, plus territory, duration, sublicensing and moral rights waivers where permitted.

Where no licence is available, consider alternatives: public domain works, synthetic data, a curated dataset licensed from a vendor, or a system restricted to your own first-party content. Each option has its own evidence requirements.

Vendor selection is part of licensing. Ask what training data a tool uses, whether the vendor has licences for it, and whether indemnities are offered. The vendor due diligence checklist covers the questions to put in writing.

Nine checks before you publish AI-assisted content

Work through the nine checks in order for each significant output, and record the result. A content lead can run them with legal input, so the work does not wait for a lawyer to be free.

A worked example makes the point. A retailer uses a third-party tool to draft product descriptions. The tool's terms say outputs may be used commercially, but the vendor does not warrant its training data. The retailer compares descriptions against supplier copy, keeps the vendor contract on file and records the process in its register.

When a supplier queries a phrase, the retailer can show the contract, the comparison and the date. For a repeatable version of the sequence, adapt the quality checklist to your publishing cycle.

Attribution, moral rights and the evidence trail

Attribution is a licence condition more often than a legal requirement, but it is still enforceable. If a licence or publisher agreement requires credit, publishing without it is a breach. Build credit into the template and keep the disclosure policy consistent with it.

Moral rights are separate from economic rights. The right to be identified as author is not infringed unless it has been asserted, and the right to object to derogatory treatment applies to certain works. A contract can waive moral rights, but only if the wording does so clearly.

There is no moral rights issue with a machine, but there is with the human authors whose works were used. Distorting a photograph or adapting a passage in a way that prejudices the author's honour or reputation can trigger a claim even where a licence exists.

An evidence trail is what turns a policy into a defence. If a rights holder asks why their work appears in your output, answer with a document, not a recollection. Record the source, the permission, the output check and the takedown route as you go, not after a complaint.

Use a register with fields for source, basis, licence reference, territory, review date and owner, and make it searchable. A folder nobody can search is not an evidence trail. Review the register each quarter, because sources change, licences expire and vendors update their terms.

Retention should match your risk. Copyright claims can be brought years after publication, so keep licence and permission records for the life of the content plus a margin. Data protection records have their own retention rules under UK GDPR.

Common questions