AI opinion pieces on your desk: how to check provenance and transparency per submission

Treat each submitted opinion piece as a verification task: record whether AI use was flagged, which editor checked it and whether this was disclosed to the reader. Do not rely on one detector, but on a fixed workflow with logging and a consistent, human-readable label that aligns with journalistic codes.

The occasion is an analysis by AIReport, the project of technology journalist Alexander Klöpping, which according to NL Times examined 252 submitted opinion pieces at NRC, Trouw, de Volkskrant, Het Parool and FD between 27 July and 27 August 2026. Of these, 49 were written entirely by AI and 57 partly. These pieces were published as opinion under human names, usually without the reader knowing. For opinion desks, the practical question is therefore not whether AI occurs in the submissions, but how you handle that in a verifiable way.

What exactly did AIReport find about AI-written opinion pieces in Dutch newspapers?

AIReport describes, in the account by De Dagelijkse Standaard, the method and the distribution per newspaper. The submissions examined concerned reader contributions on the opinion pages of five large Dutch titles.

• A total of 252 opinion pieces examined between 27 July and 27 August 2026.
• 49 pieces fully AI-generated, 57 partly.
• At de Volkskrant, according to AIReport, more than a quarter of the opinion contributions were AI-written.
• At Trouw it was 17 per cent, at Het Parool 15 per cent and at FD 11 per cent.
• At NRC not a single piece was written entirely by AI.

The striking difference between the titles suggests, in our assessment, that this is not only a matter of technology but also of editorial arrangement: we expect that where desks check more actively for AI provenance, less undetected AI text will appear, although the available research does not yet demonstrate this connection firmly.

What do journalistic codes and the NPO principles expect of AI transparency?

The standards are already established. In its AI guideline and transparency article, the Belgian Council for Journalism states that the editorial team remains fully responsible for AI-driven output and that editorial teams communicate transparently when content is produced wholly or partly through automated processes, including reference to underlying sources where possible.

The broadcaster-wide principles of the NPO likewise emphasise transparency and responsibility with generative AI, and mention practical options such as disclaimers or labels. The NPO itself notes that concrete implementation differs per organisation. AI opinion pieces that appear without any disclosure are in tension with these codes. For those who want to secure the legal side, this aligns with the broader obligation to mark AI content according to the AP's supervision under the EU AI Act. Anyone who also processes personal data in the submissions should additionally take account of the General Data Protection Regulation (GDPR), the European privacy legislation that requires processing to remain lawful, transparent and limited to what is necessary.

Why do labels alone not solve the trust problem?

Sticking on a label is not automatically a solution. Research from 2026, published in Digital Journalism (Taylor & Francis), shows a twofold picture. On the one hand, readers want clear, visible and detailed AI labels, including explicit statements such as "generated by AI" alongside author and source details. On the other hand, the related study "Feeling Iffy About Generative AI" finds that task-specific AI disclosures significantly lower perceived trustworthiness, with differences per audience segment.

That is the core tension: hiding AI undermines integrity, but a generic disclaimer such as "possibly made with AI" may actually erode trust without explaining anything. In our assessment, only a label linked to demonstrable human review works: what was done, by whom, and why the piece was published.

What does a verifiable workflow for opinion desks look like?

Instead of relying loosely on one detector, an opinion desk can arrange the handling as a fixed chain. Our editorial recommendation, based on the codes and the audience research above:

1. Intake with flagging. Scan submissions for possible AI provenance and treat a signal as grounds for further enquiry, not as proof.
2. Explicit publication policy. Set out when AI use is acceptable and under what form of openness, so authors know in advance what is asked of them.
3. Logging of responsibility. Record which editor edited, fact-checked and approved, per piece.
4. Consistent, explanatory label. Link the label to that review rather than to a general disclaimer.
5. Review when in doubt. Detectors do not provide certainty; treat outcomes as a signal and combine methods, as also applies when teams build detection as a layered stack.

The same discipline that applies to recognising and checking misleading AI citations is useful here: not trusting the first impression, but making provenance and source testable. More examples of testable workflows are in our topic hub on AI verification in practice.

What role can a verification layer such as Vera play in this?

Vera is not a chatbot and not its own language model, but a privacy-focused verification layer. For an opinion desk, such a layer can help to make visible, per piece, which verification steps were taken, which corrections or contradictions came up and which sources were consulted. That supports review and accountability, but does not guarantee that a piece is correct and does not remove the risk of errors.

For sensitive or as yet unpublished texts, it is relevant that the pre-processing and anonymisation take place on EU infrastructure, that the workflow is designed to send only anonymised content to the selected AI models, and that when a privacy check fails nothing is sent onward. The professional final judgement — publish, reject or label — remains with the editorial team. The verification layer mainly makes visible what has happened; the journalistic weighing remains human work.

Sources: The article draws on the AIReport analysis as reported by NL Times and De Dagelijkse Standaard, the AI guideline of the Belgian Council for Journalism, the NPO principles and audience research in Digital Journalism.

← All articles in this topic ← All articles