Rules and ethics
Part of AI fact checking for UK publishers: five checks
AI fact checking for UK publishers: five checks
AI fact checking for UK publishers: five practical checks, official records to consult, and the editorial, defamation and data protection rules to apply.
What to take away
- AI fact checking starts with recognising predictable failures: invented sources, wrong dates, misquoted figures and claims that are true elsewhere but false in your context.
- UK publishers answer to editorial codes such as the IPSO Editors' Code and the Impress Standards Code, plus defamation and data protection law.
- Verification means checking each claim against a primary source, then recording who checked it and when.
- Treat AI output as an unvetted draft from an unidentified contributor.
Why AI drafts slip through review
A language model generates likely text, not verified text. A fabricated date reads as smoothly as a correct one. A plausible citation can point to a page that was never published. Fluency is not evidence.
Newsroom checking usually begins with a reporter who knows where a figure came from. AI output has no chain of custody. You verify from scratch, which is closer to sub-editing unsourced wire copy than polishing your own draft.
Fact-checking means testing each claim against a primary source rather than a secondary summary. That distinction carries most of the weight when the draft arrived in seconds.
In England and Wales, publishing a false allegation that damages someone's reputation can lead to a defamation claim. A tool drafted the words, but the publisher published them. Scotland's defamation rules come from separate legislation.
Five checks for every AI draft
Work through the checks in order. Some AI fact-checking tools can flag unsupported claims; Google Fact Check Explorer can surface published fact-checks, and Google Lens can help investigate images. Treat these as leads: a person must verify each claim against its source.
- Break the draft into checkable claims.Mark names, dates, figures, quotations, and claims about people, law or health. Fluent wording is not evidence.
- Trace every citation.Open it and confirm that the document exists and supports the exact claim. If the draft gives no source, search distinctive wording on the original publisher's or official site's pages; remove or qualify claims you cannot trace.
- Check official records.Use Companies House for company records, the Office for National Statistics (ONS) for official statistics, legislation.gov.uk for legislation, and Find Case Law from The National Archives for published judgments. Check that the record and its date support the wording in the draft.
- Check accuracy, context and legal risk.Verify quotations against their original source and claims about identifiable people against the evidence. Apply IPSO Editors' Code of Practice Clause 1 (Accuracy) or the Impress Standards Code as relevant; for advertising claims, check CAP Code rule 3.7 on substantiation. Consider the Defamation Act 2013 in England and Wales and the Defamation and Malicious Publication (Scotland) Act 2021 in Scotland.
- Check data handling and record the decision.Apply the UK GDPR and Data Protection Act 2018 to personal data. Note the claim, source, date checked, checker and any qualification, then have another person verify it before editorial sign-off.
Five checks for every AI draft
- Names and job titlesofficial site or filing
- Numbers and datesoriginal release or table
- Quotestrace to transcript or recording
- Links and citationsopen every URL
- Law and regulationcheck current text
How UK codes shape the check
Different UK regulators impose different duties, and several can apply to one article.
UK codes and what they require
IPSO Editors' Code
- Who it covers
- Most UK newspapers
- Duty for AI drafts
- Accuracy and privacy
- Complaints route
- IPSO process
Impress Standards Code
- Who it covers
- Impress publishers
- Duty for AI drafts
- Accuracy and transparency
- Complaints route
- Impress process
CAP Code
- Who it covers
- Advertising content
- Duty for AI drafts
- Evidence before publication
- Complaints route
- ASA rulings
UK GDPR / DPA 2018
- Who it covers
- Personal data
- Duty for AI drafts
- Lawful handling and accuracy
- Complaints route
- ICO enforcement
Who it covers
- IPSO Editors' Code of Practice
- Most UK newspapers, magazines and their websites
- Impress Standards Code
- Publishers regulated by Impress
- CAP Code
- Advertising and marketing content
- UK GDPR and the Data Protection Act 2018
- Personal data in prompts, drafts and notes
What it means for an AI draft
- IPSO Editors' Code of Practice
- Accuracy, privacy and correction duties apply to the published article.
- Impress Standards Code
- Similar accuracy and transparency rules, with a complaints process.
- CAP Code
- Claims need evidence held before publication.
- UK GDPR and the Data Protection Act 2018
- Lawful handling and accuracy duties cover the people you write about.
Marketing content sits under a separate regime. The CAP Code requires advertisers to hold evidence for claims before publication, so an AI-written product page needs its proof assembled first.
Codes tell you what is expected, not how to brief a writer. Pair them with an AI content style guide that names the claims needing a source: statistics, legal statements, quotes and anything about a named person.
Personal data and confidential material
Pasting interview notes, customer records or staff details into an AI tool is processing personal data. The Data Protection Act 2018 and UK GDPR set the rules for handling it, and the accuracy principle covers the people you write about.
The ICO's guidance on UK GDPR covers lawful bases, retention and individual rights. Check too whether your tool provider uses inputs for training, because unpublished drafts often contain material you cannot share further.
What to record for each claim
Keep the note short, next to the draft:
- The claim as it will appear, in one line.
- The primary source and the date you opened it.
- The person who checked it.
- Any qualification the source added.
Building the check into your workflow
Give the task an owner. On most teams the writer drafts, a second person verifies, and the editor signs off, so nobody marks their own work.
Record the outcome in an AI content governance framework, with the checker's name, the date and the source used. That record is what you produce if a complaint arrives months later. The split also protects the writer, who does not have to defend a claim they never checked.
A working rule: no claim reaches publication until someone has seen the primary source with their own eyes.
Common questions
Does editing AI text remove the need to fact-check it?
No. Editing fixes tone, structure and clarity. Verification is a separate job. A sub-editor who rewrites a sentence about a court case has not confirmed the outcome.
Can I trust the citations an AI tool gives me?
Treat them as leads, not sources. Open each one and confirm the page exists and says what the draft claims. Tools have produced convincing references to documents that were never published.
Who is responsible if AI-assisted content is wrong?
The publisher. Regulators, courts and readers deal with the outlet, not the software. Keep a record of who checked each claim and when.
