Trusted AI Cannot Be One Model

Abstract editorial illustration of a connected European trusted AI ecosystem

In my last article, I explored whether public broadcasting can offer a useful analogy for trusted AI.

Not as a perfect blueprint.

But as a reminder that trust is rarely created by technology alone. It is created through mandates, institutions, standards, accountability, representation, and public purpose.

The more I thought about it, the more one question stayed with me:

If we take the public-interest analogy seriously, what would Europe actually need to build?

Because the answer cannot simply be:

One European AI model.

That would still be too narrow.

The Question Is Not Only Which Model Wins

Much of the AI debate still circles around model performance.

Which model is strongest?

Which one reasons better?

Which one is cheaper, faster, more open, more sovereign, more multilingual, or more compliant?

These questions matter. Europe needs serious model capability. It needs compute, talent, data infrastructure, cloud capacity, open and commercial providers, and a stronger AI industrial base.

But trusted AI in public-interest domains is not just a model selection problem.

A hospital does not need the same AI ecosystem as a school.

A school does not need the same AI ecosystem as a tax authority.

A public administration system does not need the same AI ecosystem as a public broadcaster.

They may all use language models.

But they do not have the same trust requirements.

That distinction matters.

Because if Europe only asks, “Which sovereign model should we build?”, it may miss the more important question:

Which public-interest AI ecosystems do we need, what should they share, and where do national, cultural, sectoral, and institutional differences need to be designed in from the start?

Healthcare AI Is Not Education AI

Take healthcare.

An AI system used in a clinical context has to operate inside a world of patient safety, medical responsibility, data protection, clinical validation, liability, regulated workflows, professional judgement, and highly sensitive information.

The relevant question is not only whether the model can produce a good answer.

It is whether the system can be trusted inside a healthcare environment.

That means different data boundaries, different validation methods, different documentation requirements, different escalation paths, and different expectations for human oversight.

It also means national and regional context still matters. Health systems are organised differently. Medical data infrastructures differ. Clinical pathways differ. Language, documentation practices, reimbursement models, and public trust can differ.

The European Health Data Space points in the direction of shared rules and infrastructure for health data across Europe. TEF-Health points in the direction of sector-specific testing, compliance support, validation, and regulatory guidance for healthcare AI.

Those are important signals.

They show that healthcare AI cannot be treated as a generic chatbot problem.

Now take education.

An AI system used in schools raises a very different set of questions.

The issue is not patient safety. It is child safety, learning quality, age-appropriate interaction, bias, manipulation, source quality, teacher agency, data protection, cognitive development, and the risk of turning classrooms into commercial attention environments.

In education, a trusted AI system should not optimise for engagement at any cost.

It should support learning.

It should help students think, not only produce answers.

It should protect children from hidden persuasion, commercial steering, misleading content, or dependency on systems they cannot understand.

The European Commission’s ethical guidelines on the use of AI and data in teaching and learning already recognise that teachers and education systems need support in using AI responsibly.

Again, the point is clear:

Education AI is not healthcare AI.

The shared word is “AI”.

The ecosystem requirements are different.

Europe Is Already Building Pieces Of This

Europe is not starting from zero.

There are already public, private, and collaborative initiatives that point toward different parts of a trusted AI ecosystem.

AI Factories are intended to bring together computing power, data, talent, and support for AI development around Europe’s supercomputing infrastructure. The European Commission also frames AI Factories as an innovation environment for startups, SMEs, researchers, and industry. Their role is not to solve trust alone, but to provide parts of the shared capability layer: compute access, technical support, and an environment where AI development can happen closer to European needs.

Testing and Experimentation Facilities, or TEFs, are closer to the sector-specific trust layer. They are designed to help AI and robotics solutions be tested in real-world conditions before broader uptake. TEF-Health is one example, focused on healthcare validation, compliance, and regulatory support. Other TEFs cover areas such as smart cities and communities, manufacturing, and agri-food.

Common European Data Spaces are another part of the shared foundation. They aim to make data available and exchangeable in a trustworthy and secure way, while giving businesses, public administrations, and individuals control over the data they generate.

The European AI Office supports implementation of the AI Act and the development of trustworthy AI across Europe. It is part of the governance layer, not the product layer.

Language technology initiatives such as ALT-EDIC, the Alliance for Language Technologies, and OpenEuroLLM point toward another necessary piece: European language coverage, openness, transparency, and model capacity that reflects Europe’s linguistic diversity.

Private actors also matter. Mistral AI is building European frontier AI systems and enterprise AI products. Aleph Alpha focuses on specialised language models and AI solutions for enterprises and public institutions in Europe. Hugging Face provides a major platform for open models, datasets, and collaboration across the AI community.

These examples do different things.

Some are publicly funded infrastructure.

Some are commercial providers.

Some are open ecosystem platforms.

Some are sector-specific testing environments.

Some are governance bodies.

Together, they show that Europe already has many of the ingredients.

But ingredients are not yet operating capacity.

What Is Still Missing

The missing piece is not another announcement.

It is the connective tissue between infrastructure, governance, and actual adoption.

Compute is not adoption.

Testing is not deployment.

Regulation is not implementation.

Guidelines are not operating capacity.

And a model, however capable, is not an ecosystem.

A hospital should not have to become an AI governance laboratory before it can adopt a tool responsibly.

A school should not have to evaluate every AI vendor from first principles.

A public agency should not have to design its own audit framework every time it experiments with AI in public services.

An SME should not have to choose between doing nothing and becoming dependent on systems it cannot explain, validate, or exit.

This is where the public-interest AI question becomes practical.

What should be shared, and what should be sector-specific?

Shared layers make sense where duplication would be wasteful:

  • compute access

  • data governance principles

  • model documentation standards

  • audit and evaluation methods

  • cybersecurity baselines

  • procurement templates

  • accountability mechanisms

  • public-interest governance principles

Sector-specific layers are needed where context changes the trust problem:

  • clinical validation in healthcare

  • pedagogical integrity in education

  • legal accountability in public administration

  • source quality in democratic information spaces

  • operational resilience in critical infrastructure

  • explainability and liability in regulated industries

That distinction may sound simple.

But it changes the whole policy and investment conversation.

It means Europe does not need to choose between fragmentation and centralisation.

It can build shared foundations while allowing different sectors and countries to adapt responsibly.

Trust Also Needs Participation Rules

There is one more uncomfortable question that should not be left until later.

If Europe builds shared public-interest AI infrastructure, who gets to participate, and under what conditions?

This is not about excluding countries or institutions because they are politically inconvenient.

It is about protecting the integrity of shared systems when rule-of-law standards, data protection, fundamental rights, institutional independence, or democratic safeguards are undermined.

This can sound abstract, so it is worth making it more concrete.

Imagine a political shift in which the rights of certain groups become contested: LGBTQ+ people, trans people, migrants, religious minorities, or other communities whose rights depend on strong legal and institutional safeguards.

The risk is not only what an AI system says at the end.

The risk can enter much earlier, through the data, categories, assumptions, and evaluation criteria that shape the ecosystem itself.

In healthcare, this could affect whether certain patient groups are represented accurately in training and validation data, whether symptoms are interpreted without stigma, or whether guidance around care is shaped by clinical evidence rather than political pressure.

In education, it could affect whether learning tools represent different families and identities fairly, whether students receive balanced information about rights and society, or whether systems quietly steer teachers and children away from topics that have become politically sensitive.

There is also a second layer: how AI output changes behaviour.

If public agencies, schools, hospitals, or companies begin to re-engineer workflows around AI-generated recommendations, then biased or politically filtered outputs do not remain isolated text. They can become new procedures, new defaults, new risk scores, new escalation paths, or new forms of exclusion.

That is why trusted AI cannot only mean “the model performs well”.

It also has to mean:

  • the input into the ecosystem is governed responsibly

  • the validation reflects the people affected by the system

  • the outputs are monitored for institutional effects

  • the organisation using the system remains accountable for how behaviour changes

A trusted AI ecosystem needs participation rules before a crisis happens.

Those rules could include baseline commitments:

  • respect for EU law and fundamental rights

  • transparent data-use rules

  • independent auditability

  • non-discrimination and due-process safeguards

  • protection against political misuse

  • oversight for high-risk public-sector use

There should also be graduated responses.

Not every breach should lead to exclusion.

But shared infrastructure needs tools such as enhanced monitoring, restricted access to sensitive shared resources, suspension of specific use cases, freezing of data-sharing privileges, independent review, and, if necessary, removal from certain trust layers.

That may sound procedural.

But it is central to trust.

A public-interest AI ecosystem cannot ask people to trust it while having no plan for what happens when trust is abused.

The Public-Interest Test

So before asking how trusted AI should be financed, I think we need one prior question:

What would make an AI ecosystem worthy of public trust?

For me, the answer would include at least five tests.

First, it should be safe enough for the context in which it is used.

Second, it should be accountable to people, not only to internal performance metrics.

Third, it should protect sensitive data and make data boundaries understandable.

Fourth, it should be resistant to commercial manipulation where public-interest use is at stake.

Fifth, it should be usable by real institutions, not only by AI experts.

That last point matters.

If a hospital, school, city, or public agency needs to become an AI governance laboratory before it can adopt a tool responsibly, adoption will remain slow and uneven.

Trusted AI needs shared capacity.

Not to remove responsibility from institutions.

But to make responsible adoption possible.

Not One Model. Public-Interest AI Ecosystems.

The next phase of the trusted AI debate should not be only about whether Europe can build a stronger model.

It should be about whether Europe can build the conditions under which AI becomes trustworthy enough to use in sensitive, public-interest, and regulated contexts.

That means sector-specific AI ecosystems supported by shared European foundations.

It means healthcare AI that understands healthcare risk.

Education AI that protects learning.

Public-sector AI that respects legality and accountability.

Information AI that values source quality over manipulation.

And enterprise AI that gives organisations practical adoption paths without forcing them into dependencies they cannot explain.

The financing question becomes more concrete when we look at it this way.

Who pays for the shared capability layer?

Who funds sector-specific validation?

Who maintains governance and audit capacity over time?

Who makes sure public-interest use cases do not remain underbuilt because they are not always the easiest commercial markets?

We have established, that the real question is no longer:

Where is Europe’s OpenAI?

So if the real question is:

Can Europe build the public-interest AI ecosystems that people, institutions, and societies can actually trust enough to use?

Then, the logical follow up will be:

How do we finance them?

Sources and Further Reading