The EU AI Act has two main goals: boosting Europe’s competitiveness and building trust in AI. They’re valid ambitions, and you can see them running through the EU’s broader AI strategy, from the AI Continent Action Plan to the push for AI Factories and trustworthy adoption across industry.
However, there is a persistent problem that keeps tripping up the European Commission – they’re still trying to build trust around systems that are opaque by design.
Europe must prioritise making AI trustworthy at the infrastructure level, and ensure businesses operating in high-risk settings are not using problematic systems that cannot explain themselves.
That is why the latest delays to this regulation matter. The European Commission has said that the delayed availability of standards puts the original timetable for high-risk rules at risk, while the European Parliament has backed postponing certain requirements – which will only make AI adoption more difficult for enterprises and high-risk industries.
Europe is still regulating the symptoms
The current policy conversation is framed as a choice between two imperfect options. The first, Europe presses ahead with rules that industry says are hard to implement. The second, it delays and simplifies in the hope of staying competitive.
Either way, the spotlight falls on oversight, rather than addressing risky models at an architecture level, or recognising that more trustworthy alternatives already exist. The trust problem doesn’t begin when a regulator asks for a disclosure, but rather much earlier, when an organisation tries to use AI in a setting where outputs need to be consistent, reviewable and defensible.
This is especially important in high-stakes sectors. If you apply hallucinations or inaccuracies to insurance or banking, the level of risk becomes entirely different. In those settings, it’s not enough for an answer to sound fluent – which large language models (LLMs) are very good at. Workers in these fields need to understand how an output was reached, whether it stayed within the right constraints, and what happens when certainty is not possible.
Why black-box AI is the wrong foundation for regulated industries
Most enterprise AI today is still built on LLMs. These models are powerful and versatile, and they have opened up new possibilities for search, summarisation, drafting and interaction. But they are also neural systems trained on enormous volumes of data, and their internal processing remains opaque by design.
So they are not useless, but a poor fit for decisions that need to be traceable after the fact. A system can be linguistically impressive and still be nonfunctional to use in a regulated workflow if the only way to manage risk is to place a human reviewer behind every output.
This is one reason so many organisations hit a ‘trust ceiling’. Pilots continue, spending continues, enthusiasm continues, but scaling stalls because confidence can’t truly be grasped. The more consequential the workflow, the harder it becomes to justify relying on an answer that cannot show its workings.
“Human in the loop” – a process where humans actively review and verify automated systems decisions – is often presented as the solution to this. But if the reviewer is sense-checking the output of a model that cannot explain how it reached an answer, then the organisation has not removed the trust problem. It has instead inserted manual labour into an opaque system – which not only is inefficient but begs the question of why the AI is being used in the first place.
What backing auditable AI means in practice
This is where neurosymbolic AI deserves more serious attention from policymakers. The European Data Protection Supervisor describes neurosymbolic AI as a family of approaches that combines machine learning with symbolic reasoning and structured knowledge. In practical terms, that means bringing together the pattern-recognition strengths of neural models with the explicit rules, logic and constraints that symbolic systems handle well.
That combination matters because real-world enterprise problems usually involve both kinds of work. Language is messy, but regulation is not. Some documents may be unstructured, but policies, thresholds and obligations are often explicit. A model may be useful for extracting meaning from text, but a rule-based component is essential for checking whether an answer is compliant, consistent and within the boundaries set by the user.
Neurosymbolic approaches will not replace every other form of AI, nor should they. But in regulated sectors they offer something the current debate often overlooks: systems that are easier to audit, easier to constrain and easier to explain. That does not eliminate the need for governance. In fact, it makes governance more meaningful, because there is a clearer chain between how a system works and how its output can be defended.
Europe’s real opportunity is trusted deployment
The EU’s own strategy already points in this direction. Brussels is not only trying to regulate AI; it is trying to make Europe a serious place to build and deploy it, with public investment in compute, infrastructure and adoption. The opportunity now is to connect that ambition to a clearer technological thesis. What kind of AI economy does Europe want to build? One defined by more rules than any other jurisdictions or one where regulated industries can actually trust, deploy and see returns from AI?
That would mean doing more than refining timelines and obligations. It would mean using procurement, standards, regulatory sandboxes and industrial policy to favour systems that can show how decisions are made, operate within explicit constraints, and escalate uncertainty rather than bluffing through it. The AI Act can help create guardrails, but guardrails alone will not create an AI economy that people are willing to rely on.
Europe has spent years asking how to regulate black-box AI responsibly. It should now ask a more important question: is black-box AI the right foundation for a regulated AI economy in the first place?
