Spain Just Made AI Provenance a Legal Duty. Owning the Stack Settles It

Spain has stopped asking nicely. On 26 May 2026 the Council of Ministers approved the draft Organic Law on the proper use and governance of artificial intelligence, and sent it to the Congress of Deputies, where it is now working through parliamentary processing. It is one of the most assertive national AI regimes in the European Union, and its centre of gravity is a single idea that every regulated leader should sit up for. If your systems generate or alter content, you have to say so, clearly, and you have to be able to stand behind it.

That is not a ban. Spain is not outlawing AI. It is building a labelling, transparency and enforcement regime, and it has given that regime teeth. For anyone running AI inside a regulated business, the interesting question is no longer whether disclosure is required. It is whether you can prove you complied.

What the law actually says

The headline duty is provenance. Content that is generated or meaningfully altered by AI, especially images, audio and video that resemble real people, places or objects, has to be labelled. Failing to label it is treated as a serious infringement, not a footnote.

The penalties are structured in tiers, deliberately aligned with the EU AI Act that Spain is transposing and reinforcing. Very serious infringements can reach 35 million euros or 7 percent of worldwide annual turnover, whichever is higher. Serious infringements can reach 15 million euros or 3 percent of turnover. Minor infringements can reach 500,000 euros. Those ceilings mirror the European thresholds, so a firm operating across the bloc sees the same order of exposure everywhere.

Enforcement sits with AESIA, the Agencia Española de Supervisión de Inteligencia Artificial, headquartered in A Coruña. AESIA is not a plan on paper. It has been operational since 2024, launched under founding director Ignasi Belda and now led by director general Alberto Gago, who took over in December 2025. Its market-surveillance mandate is live. Alongside it, the directly applicable transparency obligations of Article 50 of the EU AI Act land on 2 August 2026, covering labelling of AI-generated content, disclosure of deepfakes, and telling people when they are talking to a machine.

So the shape is clear. A live regulator, a live European baseline arriving in August, and a national law raising unlabelled AI content to a serious offence with very large fines behind it.

Why "we labelled it" is not the same as "we can prove we labelled it"

Here is where most AI stacks quietly fail. Labelling is easy to assert and hard to evidence. When a regulator, a court, or a counterparty asks you to demonstrate that a specific image, a specific audio clip, or a specific automated decision came from your system and carried the required disclosure, a screenshot is not proof. A policy document is not proof. An engineer's recollection is not proof.

The gap is architectural. If your content passes through third-party APIs, shared inference endpoints and log stores you do not fully control, your provenance record is a story you are asking everyone to trust. Under a regime where mislabelling is a serious infringement carrying fines up to 35 million euros, "trust us" is an expensive position to defend.

I think the honest answer is that provenance has to be built into the system, not bolted on after the fact.

Provenance by architecture, not by promise

I am Micky Irons, and I build Mickai. Mickai is a Sovereign Intelligence Operating System, a SIOS. Regulated organisations own it and run it inside their own walls, air-gapped where the workload demands it, on their own hardware. It is built and it is live.

The part that matters for Spain's law is this. Every action Mickai takes writes a cryptographically-signed audit record. Every generation, every alteration, every automated decision leaves a signed, tamper-evident entry that ties the output to the model, the moment and the account that produced it. Provenance is not a feature you remember to switch on. It is a property of the runtime.

That changes the compliance posture from defensive to demonstrable. When AESIA or an auditor asks you to show that a piece of content was AI-generated and carried its label, you do not reconstruct the answer. You produce the record. You can prove what your system generated, when, and that the required disclosure travelled with it. That is compliance by architecture, not by promise.

Owning the stack settles the harder questions too. Because the intelligence runs on infrastructure you control, the audit trail never leaves your custody, the labelling logic is yours to configure to the exact standard the law requires, and there is no external vendor whose logs you have to subpoena to reconstruct what happened. The provenance and the data stay on your side of the wall.

The honest version of the sovereignty case

I want to be precise, because overclaiming helps nobody. Spain's regime, like the EU AI Act, DORA, GDPR and the FCA and PRA frameworks, does not bar regulated firms from the cloud. Almost every one of these regimes permits cloud with the right controls. The genuine no-cloud bar is workload-specific, covering classified material, isolated operational technology, and cases where a data protection assessment comes back negative.

The real driver is not a universal legal prohibition. It is preference, and it is provability. When the penalty for a provenance failure is a serious infringement with an eight-figure ceiling, controllers increasingly prefer an architecture where the proof is native, where nothing about labelling or audit depends on a third party's goodwill. That is the case Mickai is built for.

The takeaway

Spain has turned AI provenance into a legal duty and put a live regulator and very large fines behind it. The August 2026 European baseline makes disclosure unavoidable for anyone operating in the market. The organisations that will move through this comfortably are the ones that can prove, not merely assert, what their systems produced and that the required labelling was attached.

You can meet a labelling law with a promise. Or you can meet it with a signed record on every action, generated by a system you own and run yourself. One of those survives an audit. I built for the one that does.

Frequently asked questions

Does Spain's new law ban AI?

No. It is a labelling, transparency and enforcement regime that transposes and reinforces the EU AI Act. It requires clear disclosure of AI-generated and AI-altered content and prohibits certain harmful practices. It does not prohibit the use of AI itself.

What are the fines for failing to label AI content in Spain?

Failing to label AI-generated content is treated as a serious infringement. The sanction tiers run up to 500,000 euros for minor infringements, up to 15 million euros or 3 percent of turnover for serious ones, and up to 35 million euros or 7 percent of worldwide turnover for very serious ones. Enforcement sits with AESIA, Spain's AI supervisory agency.

When does the labelling obligation take effect?

The national Organic Law was approved by the Council of Ministers on 26 May 2026 and is in parliamentary processing. Separately, the directly applicable transparency duties under Article 50 of the EU AI Act, including labelling of AI-generated content and disclosure of deepfakes, apply from 2 August 2026.

How does running AI on your own hardware help with provenance rules?

When your AI runs inside your own walls with a cryptographically-signed audit record on every action, the provenance of every output is captured at the moment it is produced and stays in your custody. You can demonstrate what your system generated and that it carried the required label, rather than asking a regulator to trust an assertion. That is what I mean by compliance by architecture. It builds on the sovereignty and cryptographic audit themes I cover in my writing on the Sovereign Intelligence Operating System and on cryptographically-signed audit records for regulated AI.

By Micky Irons