AI in the newsroom
Disclosure norms, the EU AI Act's Article 50 obligations, hallucination and fabrication risks, and the dividing line between safe assistance and dangerous automation.
Generative AI sits in the newsroom whether leadership plans for it or not — in transcription, research, translation, headline variants, and increasingly in drafting. The professional question is no longer whether to use it, but where the line is: what AI may assist, what it may never do unsupervised, and what must be disclosed to the audience.
The regulatory floor
EU AI Act, Article 50 — the transparency article — is the binding baseline for EU-facing publishers:
- Providers and deployers of systems that interact with humans must make clear that users are interacting with AI (unless obvious).
- AI-generated or manipulated content (text, image, audio, video) published with AI involvement must be disclosed as artificially generated or manipulated, subject to limited exceptions for genuinely editorial control.
- Deepfakes must be labelled; emotion-recognition and biometric systems have disclosure duties.
Self-regulatory codes align: IFCN principle 5 on transparency of method, SPJ’s accountability clause, and emerging newsroom AI policies (AP, Reuters, Bloomberg, and major broadcasters all publish theirs). Common floor: disclose AI involvement, keep a human accountable for every published claim, and never let an AI system make a publish/no-publish decision alone.
The risk map
| Use | Risk | Working rule |
|---|---|---|
| Transcription, translation | Low-medium — errors are verifiable | Human spot-checks, especially quotes and numbers |
| Research assistance, summarising | Medium — summarised errors propagate | Verify claims against the primary source before use |
| Headline/summary variants | Medium — drift toward clickbait | Editorial sign-off on every variant published |
| Drafting routine copy (weather, results) | Medium-high — confident fabrication, stale facts | Template + verified data feed only; disclose; human review |
| Drafting news copy from prompts | High — hallucinated sources and quotes, style-of-author plagiarism | Not without full human re-verification; disclosure required |
| Autonomous publish decisions | Critical — accountability failure | Never. A person must own every published claim. |
The technical failure modes
- Hallucination — LLMs produce fluent, confident fabrications: sources that don’t exist, dates that never happened, quotes no one said. The failure is invisible to the writer unless checked — every AI-assisted claim requires the same verification as a source claim.
- Composite “quotes” — models synthesise plausible-sounding quotations; treat every AI-produced quote as unverified until checked against a recording or transcript.
- Style plagiarism — models trained on journalism can reproduce an outlet’s distinctive phrases; originality checks apply to AI output too.
- Data leakage — pasting unpublished copy, sources, or documents into external tools can breach source protection and embargoes; know where the data goes.
- Drift — model versions change behaviour silently; an approved workflow needs periodic re-evaluation.
A working newsroom policy
- Define allowed uses in writing (and update it — capabilities change quarterly).
- Require disclosure of AI involvement to the audience for any AI-generated content, and to the editor for any AI-assisted draft.
- Assign named human accountability for every published claim — “the AI wrote it” is never a defence.
- Log AI involvement in the editorial record so it can be audited later.
- Never feed confidential sources, unpublished investigations, or embargoed material into external AI systems.
MediaMeter context: our specification encodes these rules — automated-news is a distinct content type requiring ai-disclosure; detection of undisclosed AI content triggers scoring impact and mandatory human review; the charter prohibits MediaMeter itself from making publish/reject decisions.
Last reviewed: October 20, 2026. Part of the MediaMeter knowledge base — editorial reference material, not product documentation.