The AI model may change, but the knowledge must remain

Most publishers now have a fairly clear idea of what AI can do. It writes teasers and summaries, translates and drafts emails. Baier considers this debate to be over. He is far more interested in what happens to the data that a publishing house hands over to AI every time it uses the technology.

He believes that a publisher’s assets worthy of protection extend far beyond texts and images. They include rights and licences, the tone in which a publishing house communicates with the outside world, terminology databases and glossaries. For Baier, this whole constitutes the business model. And it is precisely this that is gradually being transferred into AI tools as soon as editorial teams start working with them in earnest.

The mechanism is straightforward. The more useful a result is intended to be, the more detailed the prompt becomes. The more detailed the prompt, the more in-house knowledge ends up with a provider. Added to this are the sets of rules that are fed into the AI. Baier’s example comes from everyday life. Anyone who instructs the AI never to use dashes, so that no one recognises the text as machine-generated, is already outsourcing a part of their house style.

There is even less control in situations where staff work without authorised tools and simply open an AI window in their browser. In such cases, the publisher has no way of knowing what information is being disclosed, nor whether the use of the tool is cost-effective.

Publishers know what technological dependence feels like. Baier recalls Google’s Panda update in 2011, which caused many niche websites to plummet in the search results overnight. Back then, the publishers were able to take countermeasures. They rewrote their content, redesigned their adverts and were back in the game after a few weeks.

With AI, the dependency runs deeper. According to Baier, anyone who incorporates prompts, agents, context and rule sets into a particular tool creates an additional layer of dependency, alongside the technical one. If you change the model, you generally get different results. These may feel better, but they are also different, and an editorial team that requires a consistent tone would have to monitor this constantly. If several AI agents are linked in series, this monitoring becomes more labour-intensive and expensive.

If the tool disappears completely – whether because it is taken over or shut down – the process chain breaks down. The AI market is volatile enough that no one can rule this out.

Carsten Schwab, owner of XPS AG in Zurich, describes the same mechanism from the perspective of data processing. With every query, the response from a language model is ‘generated anew’. New model versions alter the results even if nobody makes any changes to the system itself. From this, he concludes that a long-term solution must not be a ‘finished product’ that takes a year to programme and which nobody can then adapt afterwards.

Bayer’s alternative to this dependency is a controlled environment. The editorial team works in a protected space containing the content, its sources and metadata, the organisation’s rules, roles, approvals and statuses. The language model accesses this space – and this space alone. The publisher thus determines what constitutes the truth for the AI.

He identifies three use cases. When querying its own content – known as Retrieval Augmented Generation (RAG) – the model works exclusively with defined and approved sources, and each response cites its source. For teasers and social media texts, quality is ensured by fixed rules governing tone, length and terminology for each channel, which are defined outside the model. This means that not every editor has to become a ‘prompt artist’, as Baier puts it. Finally, when it comes to translations, the AI operates within the existing XML structure, whilst glossaries and technical terms are preserved.

At the end of each of these chains is a person who checks and approves the content. Responsibility for sources, rights, tone and liability for what is published remains with the editorial team. The secure space comprising content, rules and context remains; Baier regards the language model as an interchangeable engine. The boundary no longer lies between humans and machines, but between efficiency and responsibility, as Baier explains.

Fabasoft Xpublisher sets out the questions a publisher should clarify in its white paper ‘10 Steps to the Right Publishing System’ [PLACEHOLDER: Link to white paper]. In the chapter on integrated AI, these are presented as a checklist. What data is the AI permitted to use, and where is it processed? Is it integrated into roles, permissions and approvals? Do the results remain transparent, verifiable and under human control?

On the first day of the conference, Carsten Schwab places the data itself at the centre of this same analysis. Anyone setting up a RAG system for a publisher – whether for a chatbot or a search function that answers questions based on the publisher’s own content – first breaks this content down into small units of text, known as ‘chunks’. How these are structured determines the quality of every subsequent response.

Schwab advises that publishers should therefore know how their chunks are created and be able to examine them. Above all, a publisher must be able to ‘pack up’ its chunks at any time and move on. If they are regarded as the service provider’s property or as part of its software, any change means starting from scratch.

His advice centres on a neutral XML format from which chunks can be regenerated for each model and each database. “Whoever has this XML in their hands has control over the data,” says Schwab. XPS offers this kind of data preparation itself; that is part of the package. His call is in line with Baier’s core principle that the model may change but the knowledge must remain.

Manuel Montero, Managing Director of Data to Type GmbH, takes things a step further back on the second day of the conference. Before chunks even exist, the content must be processed, and, as he has observed, this processing is often carried out using low-cost conversion services provided by major tech conglomerates. Many publishers are concerned about the language model, to which only individual sections of text are ultimately fed, and overlook this stage.

“You’re uploading all your knowledge, neatly organised, to them,” warns Montero, who has been working on the automated processing of structured content for more than twenty years. Anyone who, as part of a selection process, has several providers test their own catalogue is replicating this approach. This may also include cited third-party content to which the publisher itself holds no rights.

Five questions can be derived from the three presentations, which a publisher should clarify before entrusting rules, prompts or content to an AI tool.

1. Where can we find our rules? Tonalities, terminology and writing conventions belong in a separate, version-controlled repository from which each model draws. Why this is important: only then can the organisation’s knowledge survive a change of model.

2. Will we receive our processed data? Before the project begins, clarify whether chunks and intermediate formats will be exported in an open format. Why this is important: without them, every change of provider means starting from scratch.

3. Who produces our content, and where? Ask which departments are involved in the conversion, even during the trial phase. Why this is important: it is often at this stage that the entire stock leaves the premises.

4. Do we notice when the quality changes? A fixed set of test questions with known answers shows where a new model performs better or worse. Why this is important: Answers are generated anew for every query, and providers update their models on an ongoing basis.

5. Who ticks the box? For every AI-supported process, it is specified who checks and approves the content. Why this is important: Liability for published content remains with the publisher, regardless of which model is involved in the writing process.

In June 2026, Anthropic temporarily suspended access to its Fable model without prior notice. Three weeks later, it was available again. With the next model, the next provider or the next takeover, things may not turn out so well. The three presentations all boil down to the same idea. Models come and go, and what a publisher has built up over the years must withstand this change. This includes its rules, its processed data and its responsibility for what it publishes.

For decision-makers, this means first clarifying the architecture for every AI project and then selecting the tool. Where are our rules, who prepares our data, and can we take it with us? If you answer these questions before getting started, you’ll be able to take the next model change in your stride.

The article is part of the Channels: Digital Publishing Technologies, which focuses on content strategies and processes. The channel is sponsored by Fabasoft Xpublisher.