From guidelines to architecture: How newsrooms are rethinking AI governance

"We scream." That is what Ole Reissmann, Director of AI at Der Spiegel, said when I asked how he thinks about AI and the governance questions it raises. "We (now) break it up and give all the bits and pieces away to an AI system we don't control, that might have biases we don't know about, that assembles it back together, and then attaches our name and our brand to it,” he said. “How does that look? We scream."

Der Spiegel is not alone in facing these challenges. Many other publications are struggling to implement governance structures to navigate the challenges posed by the rise of AI. 

For this article I spoke to 20 newsroom leaders, experts and academics from 13 countries around the world. Whilst their concerns varied, a theme emerged amongst the leaders further along in their AI journeys. The more they experimented with AI, the more they realised the hardest governance questions provoked much bigger existential questions. 

Martin Schori, former Director of Editorial AI at Swedish newspaper Aftonbladet and now co-founder of Hint, wanted to know how journalism can remain relevant when a large language model can summarise almost any piece into three bullet points. Luke Bradley-Jones, the newly appointed CEO of The Economist Group, told me that understanding what journalism looks like in a post-search world is what keeps him awake at night.

These are questions that are not just about the technology or business models but are about power and accountability. They force news organisations to ask what journalism is for in an age of AI.  Who is actually controlling the systems shaping what audiences consume? Who carries responsibility when something goes wrong? And how do news organisations uphold trust when relying on tools they neither fully own nor entirely understand?

Many news organisations are still at an early stage of grappling with these questions. According to a recent study by FT Strategies and WAN-IFRA based on responses from 448 newsrooms, 57% have no AI representation in their organisations, while 64% still produce news for a single legacy channel. The most common measure of AI success was also efficiency, with 42% reporting time saved.

1. How newsrooms approach AI governance

Three broad AI governance approaches are emerging. 

At one end of the spectrum are organisations where AI use has emerged organically rather than through a formal governance process. Journalists are using tools like ChatGPT with no formal newsroom policies and oversight mechanisms, often due to lack of resources or bandwidth. 

The second approach starts with guidelines or principles, supported by cross departmental committees and ‘Human in the Loop’ oversight of AI generated content. Within this group, some organisations are more explicit than others about the role of their values in shaping how governance is designed.

Some larger organisations have gone for a third approach and built more formal structures. YLE, the BBC and Czech Television, for example, have dedicated Responsible AI leads and teams.

Czech Television appointed Veronika Macurova Krizova, as its first Responsible AI and Strategy Officer in January this year. She is an IP lawyer by training, with a remit extending across governance, AI literacy and ethical AI adoption. She’s also keen to explore practical mechanisms that could help journalists protect themselves against synthetic media, social engineering and other manipulation attempts, an area she feels is currently underreported.

At the BBC, James Fletcher is the Responsible AI lead for the organisation, not just for news. His team is built around three pillars: governance and risk, principles and policy, and evaluation and data science. The Editorial Policy unit owns the editorial principles. What Fletcher's team does is to ensure AI does not undermine them. 

"We think a lot about how we do that through culture as much as through compliance,” he told me. “We have a great example in our broader editorial culture: at the BBC thousands of people make editorial decisions every day about reasonably significant risks. By and large, they do it by themselves, and they by and large get it right, and if they need support, they can come to a central team. That’s the sort of model we want to extend to AI governance more broadly"

JP/Politikens in Denmark started their governance journey in 2019 in collaboration with four universities supported by Innovation Fund Denmark. The ambition was to develop responsible AI systems whilst reducing dependence on the major technology platforms. 

Their first output was a governance framework called what they called a Values Compass, about which our own Marina Adami wrote about in December 2024. The compass helps publishers define their role in society and their values before making technology decisions.

Kasper Lindskow, the Head of AI for JP/Politikens Group, describes a process that starts before any technology decision is made: "Developing an AI artifact starts with you defining your role in society as a publisher, independent of technology, to try to make sure that your efforts are anchored in that."

Franco Piccato is the Executive Director of Chequeado, an Argentine fact-checking organisation that has been working with AI since 2015. Piccato believes news organisations should aligned any AI initiatives with their mission: "We've been rushing a lot to implement AI because it's the new shiny thing. But if we could listen to our audiences first, and see where we can be more helpful, then we would be aligning our mission with our internal processes and the outputs that we need to serve them better."

Laura Zommer led Chequeado for a decade before co-founding Factchequeado in 2022, a collaboration with Spanish fact-checker Maldita serving Spanish speaking audiences in the United States. She carried the same philosophy into the new organisation and extended it to procurement. 

Factchequeado's ‘Chat Migrante’ is a closed chatbot drawing on curated official sources and the organisation's own journalism. It was built by Maldita's technology arm Botalite, partly because a Spain based vendor meant European data protection standards applied. Factchequeado also negotiated regular automated testing with the vendor, in addition to their own fortnightly internal evaluations.

2. Human oversight is under pressure

Regardless of how organisations approached AI governance, or where they chose to place the emphasis, one thing remained consistent across almost every newsroom I spoke to and that is that keeping human oversight is still seen as the primary safeguard when it comes to AI generated content.

But the strain on this approach is now beginning to show. As AI models become increasingly capable, the challenge is not just about reviewing AI generated outputs well, but about the impact this might have on the journalists themselves.

"AI doesn't reduce work. It intensifies it." This is how the Harvard Business Review recently described what is happening across organisations adopting AI. Across the news industry, this intensity appears to be falling on editors who were already overstretched.

Annika Ruoranen was the Responsible AI Lead at Finnish public broadcaster YLE before moving into consultancy. She said she was concerned about where the burden of governance landed in the newsroom.

"That middle manager level is super packed,” she said. “They are working with the daily things, working for the quality of stories, and then I have to go there and ask them to take care of AI governance as well."

Sannuta Raghu, Head of AI Product at Scroll in India and former fellow at the Reuters Institute, shares similar views: "Because there is this burden of AI on it, we have to now look at it with this fine-tooth comb, which we would [otherwise] be very comfortable putting out... and so that is causing fatigue.

This fatigue has caused Scroll to introduce limits on the level of AI-assisted output staff are expected to review.

From 2 August 2026, Article 50 of the European Union’s AI Act requires deployers operating in the EU “to disclose AI-generated text they publish to inform the public on matters of public interest…with exception for text which has undergone human review or editorial control which a natural or legal person holds editorial responsibility for.”

Macurova Krizova was a part of the working groups that played an active role in developing the operational "standard" for the European Union’s Code of Practice on Transparency of AI-Generated Content. However, she is sceptical about how these exceptions will work in practice and raises questions about what meaningful oversight looks like. 

“It is very easy not to open the citation snippets,” she told me. “It’s very easy to take one step further and end up with an almost ready final piece. This is why I believe public service media matter more than ever. They still stand for reliable journalism and the responsible work of professionals even in the age of AI.”

3. Cognitive surrender and sycophancy

In addition to the pressures of workload, fatigue and the new legal requirements, researchers are also raising concerns about the way many AI models are designed and the impact these designs are having on user behaviour. 

Mala Kumar is the Executive Director of Humane Intelligence, a nonprofit dedicated to breaking down barriers to AI deployment for social good. Kumar specialises in evaluating AI models and has spent years red teaming some of the AI systems people use every day.  

Red teaming is a practice that was originally used to test systems for the purposes of cybersecurity. Mala’s team uses it to run structured tests to expose harmful taxonomies that may or may not be present as well as capacity building and as a form of AI literacy. 

Through this work, she has noticed a recurring design choice in some of the most widely used models: "So OpenAI has made a design decision, essentially, to validate a lot of things that people say, even if they're categorically a bad idea. Claude and Gemini models tend to take a much more balanced approach when they phrase something. So they may say, this is not a good idea or here's why it could be a good idea or a bad idea. They'll do something to qualify the statement, which at least shocks the user out of this idea and this kind of AI psychosis that a lot of people have been spiraling into, where they're constantly validated and constantly given the information they want."

"Cognitive surrender" is a term introduced in a 2026 working paper by researchers Steven Shaw and Gideon Nave at Wharton University of Pennsylvania. 

They define cognitive surrender as "the behavioral and motivational tendency to defer judgment, effort, and responsibility to AI outputs, particularly when those outputs are delivered fluently, confidently, or with minimal friction." 

Across three studies involving 1,372 participants, they found that people adopted AI outputs even when those outputs were deliberately wrong and showed that confidence increased regardless.

Whilst this research is still emerging, some newsroom leaders are beginning to describe similar behaviours in reality. Ibrahim Shehu, Group Editorial Director and Editor-in-Chief of the Media Trust Group in Nigeria, is concerned about what he is seeing in his newsroom. "I feel like [AI] is killing human creativity,” he said. “A lot of young people in the newsroom are not thinking a lot, are not being creative. Everybody wants to use the tool to get things done quickly." 

Shehu’s concerns extend to audience trust. "[It’s about] authenticity…[if] your stories are being told by machines, there's that fear that we may end up losing audiences," he said. To that end, he is exploring ways to encourage original thinking and discourage overreliance on AI tools, including the possible use of AI detection.

4. AI as a tool to reduce biases

In this piece I wrote in 2025, I outlined a range of biases that can emerge across data, models and the machine learning lifecycle, with examples of how newsrooms were mitigating them. Those concerns and examples remain relevant today.

One year later, I was struck by how many newsrooms are using AI to help surface biases in their own journalism. David Caswell, who advises newsroom leaders across the industry, believes some biases may be harder to change than the ones found in AI systems. 

Reflecting on his time working for the BBC during the UK's EU referendum, he says: "We missed Brexit... half the country was in an absolute rage at the status quo. And in our entire newsroom we didn't see it, because we're all living inside the M25."

Schori,when he was still at Aftonbladet described the experiments looking at who is represented in coverage: 

"if you look at the numbers, you could see that women don't read a lot of the… our politics coverage. And… you can then come to the conclusion that women don't like politics. I don't think that's the answer. It's probably… that women maybe are not attract[ed to] the way we cover politics."

The same techniques that can surface bias in newsroom decision making can help reveal how bias may be emerging within AI systems themselves. As James Fletcher from the BBC, explained, "there are some things you can only spot at scale; certain types of bias are a clear example of that. If you're looking at one piece of text, you can't tell whether it's using male pronouns more than female pronouns in translation. But if you look at 10,000 pieces of text, you can."

It is a principle that will matter more, not less, as newsrooms move towards more agentic systems. "One thing is having human-in-the-loop on an isolated case that any sort of generation is,” said Kasper Lindskow. “Another thing is having a human-in-the-loop across the output where you can see patterns. This is a good thing to add, always, but it will also be increasingly necessary as GenAI becomes better and plays an increasingly central role in the newsroom."

5. What’s next for AI governance 

Not all of the big governance questions can be answered by individual newsrooms. Coalitions are beginning to form and governments and regulators are starting to pay closer attention. The ‘Spur’ coalition that now comprises more than 30 media organisations, is one example which is working to establish standards and shared infrastructure for a fairer relationship between publishers and AI platforms.

For newsrooms earlier in their AI journeys, governance still begins with mission, values and editorial principles. Human oversight remains an important safeguard, but it is equally important to recognise its limits and design systems that do not rely on humans catching every error. 

Some organisations are already moving in that direction. Sannuta Raghu, Head of AI at Scroll in India, has been implementing the so-called "News Atom", a framework she developed during her fellowship with the Reuters Institute. It labels the constituent parts of journalism at sentence level with metadata that preserves provenance, attribution and editorial context. 

For Raghu, these infrastructure decisions are governance: "That infrastructure-level restructuring is sort of a governance directive in itself, because that's how you're storing your data, that's how you're looking at it."

The IPTC, the Global Standards Body of the News Media is now working with Sannuta to explore if the news atom in whole or a newly-iterated version could become a data standard.

Others are seeking influence beyond their own systems. Ole Reissmann told me this is why Der Spiegel is moving beyond content licensing to work directly with companies including Google on how its journalism is surfaced, attributed and represented inside AI products. The Economist is also experimenting with versions of its journalism designed specifically for AI agents.

The organisations that have been "screaming" (in Ole's words) are increasingly doing more than raising the alarm. They are testing new infrastructure, new partnerships and new forms of influence over the systems that will shape journalism's future. If governance is to keep pace with increasingly autonomous AI, it will need to move beyond policies and principles, and consider how it becomes a part of the architecture itself.