DPG Media puts GPT-NL to the test in the newsroom

AI supporting journalistic expertise

Generative AI can support editorial teams in a wide range of tasks, giving journalists more time to focus on what matters most: producing high-quality journalism. But once AI goes beyond routine tasks and starts playing a role in safeguarding language and quality, nuance and context become essential. A model must not only understand what is being said, but also recognise how a particular publication writes and the journalistic choices that shape its style.

This is precisely where DPG Media is exploring the potential value of GPT-NL. The Dutch language model is being developed using Dutch and Flemish data and is designed for applications in which local language, context and values matter. For a media company, digital sovereignty is also becoming increasingly important. News organisations want to retain control over their technology, data and intellectual property, particularly now that generative AI is also affecting journalism’s traditional business model.

Jona Zuiver, AI Business Partner for the news media and magazines of DPG Media in the Netherlands and Belgium, is closely involved in the initiative. He helps ensure that new AI applications comply with legal and ethical guidelines and supports editorial teams in using the technology safely and effectively.

From article to compelling headline

The first phase of the feasibility study focuses on generating news headlines. It may sound like a straightforward task, but in practice there is a great deal involved. A good headline does more than summarise the facts. It needs to capture the essence of the story, reflect the publication’s tone of voice and encourage readers to continue reading, without resorting to exaggeration or clickbait.

‘A good headline is an essential part of an article,’ says Zuiver. ‘A copy editor needs to think broadly and test several alternatives. AI can both support and challenge that process.’ A model trained specifically on Dutch and Flemish language data may be better equipped to reflect this journalistic context. International models have improved significantly in recent years, but they can still sometimes produce phrasing that feels translated or is strongly influenced by American English conventions.

The initial tests show that creating a useful model is not simply a matter of switching it on. At the outset, GPT-NL still had to learn how to interpret the task correctly. It can now generate several headline options, which can then be assessed and refined. The evaluation looks at factors including the variety of the suggestions, the accurate representation of quotations and how well a headline fits the article. At the end of this phase, editors from two DPG Media publications will carry out a qualitative evaluation.

An extra pair of eyes: the AI copy editor

The second application goes a step further: an AI copy editor that provides support during the writing and editing process. The assistant could, for example, flag sentences that are too long or complex, suggest ways to break them up and check whether abbreviations have been written correctly. The journalist remains in control, while the model acts as an extra pair of eyes.
The main challenge lies in the differences between media brands. An article for de Volkskrant follows different language and style conventions from a Flemish news publication. The AI copy editor therefore needs to recognise common style issues and make suggestions that are appropriate for the individual publication. To do so, the model needs to understand not only the content, but also the publication’s editorial identity.

Independence and transparency

Interest in GPT-NL therefore goes beyond technical performance. DPG Media operates a range of news brands, each with its own identity. According to Zuiver, independence is part of the company’s DNA. ‘That also means being transparent with readers about the technology we use and the values underpinning it. This is particularly important at a time when trust in news and the media is under pressure.’

DPG Media has already contributed to GPT-NL through the industry association NDP Nieuwsmedia by making training data available. More than thirty national and regional news organisations opened up their archives to help train the model. A proportion of any future revenues will be returned to the copyright holders. According to Zuiver, the contribution was driven primarily by a shared conviction: the importance of developing a language model that takes account of Dutch and Flemish journalism and the position of content creators.

Targeted improvement and measurable results

The feasibility study will run until the end of January 2027. DPG Media and TNO are looking not only at what GPT-NL can do, but more importantly at how the criteria for strong news headlines and effective editorial support can be made concrete and measurable. These criteria will then be used to further fine-tune the model for the selected tasks.

The aim is not to develop a single model for every task. International models have a clear advantage in terms of scale, investment and breadth of application. GPT-NL’s strength may instead lie in clearly defined use cases where linguistic nuance, journalistic quality and control are decisive. If the model can demonstrably align more closely with day-to-day editorial practice, it could not only make editorial support more efficient but also strengthen the autonomy of news organisations.

Visual: DPG Media