European public broadcasters are testing whether generative AI can become part of ordinary production workflows without pushing sensitive data, applications, and model dependencies almost entirely onto global technology platforms.

ORF has brought together the ARD Cloud Unit, EBU Technology & Innovation, and Schwarz Digits in a Joint Innovation Lab to create a digitally sovereign version of AiDitor, its AI-assisted media-production environment. The resulting proof of concept is being presented at IBC2026 in Amsterdam from 11 to 14 September.

The system connects AiDitor’s editorial and production workflows with StackIT, the European cloud platform operated by Schwarz Digits, alongside infrastructure and cloud expertise from ARD. Locally operated AI services include large language models, speech recognition, speech synthesis, and other tools intended to support media production.

The project gives Europe’s sovereignty debate a practical architecture problem to solve. Keeping information inside a European data centre does not by itself establish meaningful control if applications remain dependent on externally administered models, identity systems, APIs, or management services that are difficult to replace.

Public-service broadcasting makes those dependencies unusually visible because production environments can contain unpublished journalism, archive material, rights-controlled content, personal information, and sensitive editorial work. Once AI tools become embedded in transcription, research, search, translation, editing, or production processes, replacing the underlying supplier can become considerably harder than cancelling a conventional software subscription.

The consortium is therefore testing control across several layers rather than presenting one European model as an answer to every use case. StackIT provides the infrastructure environment, AiDitor remains the application through which media workers access tools, and locally operated AI services supply individual capabilities underneath it.

That modularity could make services easier to replace over time, although it also shifts more integration work onto the organisations operating the stack. Hyperscale cloud platforms are attractive partly because infrastructure, models, security, developer tools, identity, monitoring, and support arrive within one commercial ecosystem.

Sovereignty becomes an architecture decision

European cloud policy has increasingly moved beyond simple data residency towards questions of operational control, supplier dependence, legal jurisdiction, and portability. A workload can sit in a European facility while still relying heavily on technology controlled from elsewhere, which is why organisations with stricter sovereignty requirements increasingly scrutinise the full service chain rather than a server’s physical location.

ORF’s proof of concept addresses that problem by assembling separate European-controlled components around an existing production application. It does not remove technological dependencies altogether, but it is intended to give broadcasters more direct choices over where models run, where data moves, and which parts of the system can be replaced.

Media provides several credible AI workloads against which the architecture can be tested. Speech recognition can accelerate transcription, text-to-speech can support accessibility, language models can help organise archive material or generate metadata, and translation can make content usable across a multilingual European market.

Each application also carries editorial and governance requirements that differ from ordinary office productivity. Automated transcripts can contain errors, generated text needs human review, archive material can carry rights restrictions, and synthetic media introduces questions around provenance and disclosure that public-service organisations cannot resolve simply by choosing a technically capable model.

Shared European infrastructure could reduce some duplication if several broadcasters can use components that have already been evaluated for security, interoperability, or editorial controls. The EBU provides an existing collaborative structure across public-service media, while ARD brings experience from a large distributed broadcasting environment.

Yet a common stack has to remain technically flexible enough to keep pace with AI development. Models and services are changing quickly, so sovereignty would become counterproductive if gaining greater control required broadcasters to remain on older technology because individual components were too tightly coupled to replace.

Interoperability is therefore central to the proposition. Applications need a stable way to call different language, speech, and generative-media services without rebuilding the surrounding workflow whenever one model changes. The easier that substitution becomes, the more meaningful the claim of technological choice.

Cost provides another test. Large cloud and AI companies benefit from enormous economies of scale and mature developer ecosystems, while smaller sovereign platforms have to justify themselves through control, regulatory alignment, resilience, or specialised service. European ownership alone does not remove the need to compete on reliability, performance, and operating effort.

Public broadcasters may nevertheless place a higher value on those non-price factors because infrastructure choices intersect with editorial independence, archive preservation, public accountability, and long-term access to content. Dependence on a supplier that changes terms, withdraws a service, or limits access can become more consequential when systems support a public institution rather than a disposable digital product.

The IBC system remains a proof of concept, so it does not establish that the architecture can yet be operated economically across multiple broadcasters or that locally hosted models can match every external service the organisations might want. Nor does it demonstrate that an assembled European stack is automatically more secure.

What it does provide is a more useful definition of sovereign AI than ownership labels alone. Control has to survive contact with applications, models, cloud infrastructure, data movement, and day-to-day users — otherwise sovereignty remains a procurement claim rather than an operating characteristic.