Who is safe when the system is secure?

At seven in the morning, a patient opens a hospital portal and finds that her appointment has disappeared. She cannot tell whether the clinic cancelled it, an attacker disrupted the system, or a damaged record placed her in the wrong queue. The distinction will matter to the investigation. It matters less to the person whose treatment is now uncertain.

Across town, an applicant receives an automatic refusal for a job. The service worked exactly as designed. No outage occurred, no account was breached, and every database remained available. Yet the applicant suspects that a system has treated an old absence, an address, or a pattern in their employment history as a reason to close the door.

Europe has built two major laws for parts of this landscape. NIS2 concerns the security and resilience of network and information systems across critical sectors. The EU AI Act governs artificial intelligence according to the risks attached to its use. Both speak about safety. Both can protect people. They reach the person by very different routes.

That difference helps answer a blunt question: when Europe regulates digital systems, is it protecting human beings, or protecting something else?

What NIS2 sees

NIS2 begins with dependence. Hospitals, energy networks, water suppliers, public authorities, communications providers, cloud services, transport systems, and other critical organisations now rely on connected infrastructure. A failure in one place can travel through suppliers and networks until ordinary life stops somewhere else.

The directive requires covered organisations to manage cyber risk, secure supply chains, handle vulnerabilities, prepare for incidents, preserve continuity, and report serious failures. It places responsibility close to senior management rather than leaving cybersecurity as a technical problem for a distant department. Its legal purpose is a high common level of cybersecurity across the Union and a better-functioning internal market.

People benefit from this architecture. A patient benefits when a hospital can recover. A household benefits when water and electricity remain available. A customer benefits when a service warns them about a threat and explains what they can do. NIS2 recognises that an incident may cause material or non-material damage to people beyond the organisation where it began.

Still, the person usually appears as a recipient of a service whose continuity must be protected. The main object in view is the organisation and the systems on which it depends. NIS2 tells institutions to become harder to disrupt and quicker to report. It does not give every injured person a complete route to explanation, compensation, or repair.

That is not a defect in the directive’s design. It is a limit we should be able to see. Resilience protects the conditions under which people receive care, power, water, communication, and public service. It does not decide whether those services treat people fairly once they are running.

What the AI Act sees

The AI Act begins elsewhere. Its risk categories are organised around what an AI system does and where it is used. Some practices are prohibited because their power to manipulate, exploit, classify, or watch people is considered unacceptable. Systems used in areas such as employment, education, credit, public services, law enforcement, migration, and parts of healthcare may be treated as high-risk because they can shape safety or fundamental rights.

The Act asks providers and users of covered systems to document risks, govern data, keep records, enable human oversight, monitor performance, and report serious incidents. In certain high-risk settings, public bodies and providers of public services must examine the likely effect on fundamental rights before use. An affected person may have a right to complain. In a narrower set of consequential decisions, they may ask for a clear and meaningful explanation of the role the system played.

This is more recognisably a law about people. It names health, safety, dignity, equality, privacy, non-discrimination, and other fundamental rights. It tries to intervene before an automated judgement becomes an injury.

But it is also a law of the market. It regulates products, providers, deployers, conformity assessments, technical documentation, and supervision. The route to human protection runs through the classification and governance of a system. Much depends on whether a particular use falls within the law’s categories and whether an authority can enforce the duties attached to it.

The timing reveals another limit. Much of the Act became enforceable in August 2026, including important transparency and governance provisions. The central requirements for many high-risk systems have been postponed: until December 2027 for stand-alone systems and August 2028 for high-risk AI embedded in regulated products. Rights written into law can exist before the institutions, standards, and enforcement practices needed to make them dependable.

Protection by proxy

Both laws protect human beings through proxies.

NIS2 protects networks, organisations, reporting channels, and essential services. The AI Act protects rights by placing duties on the people and companies that build or use certain systems. This is how much public law works. A food-safety rule protects the diner by regulating kitchens, temperatures, records, and inspections. A building code protects the resident through requirements placed on materials and construction.

The danger begins when the proxy becomes the purpose.

An organisation can treat NIS2 as a search for policies, logs, training certificates, and incident reports that will satisfy an auditor. An AI provider can treat the AI Act as a classification exercise: find the category, prepare the document, place the required notice, and continue. Each form may be necessary. None can establish by itself that a person is safe.

Compliance is attractive because it produces evidence that can move upward through an organisation. Human experience often moves in the opposite direction. The missed treatment, inexplicable refusal, lost income, exposed secret, or hours spent appealing a false record arrive one person at a time. They may be too small to qualify as a major incident and too dispersed to appear as a system failure.

When the paperwork and the lived result disagree, an institution reveals what it understands protection to mean. It can ask whether it followed the required process. Or it can ask what happened to the person and what must change now.

The gap between secure and just

A system can be secure and still be cruel.

Imagine a benefits agency with excellent access controls, reliable backups, careful supplier management, and a tested incident plan. Its automated fraud system is available every day and has never suffered a breach. If that system repeatedly flags disabled people whose lives do not fit an expected pattern, cybersecurity has preserved the machinery of an injustice.

The reverse is also possible. A carefully tested medical AI may perform consistently across patient groups and present its uncertainty honestly. If a weak password, compromised supplier, or corrupted update allows someone to alter its recommendations, fairness inside the model will not protect the patient.

NIS2 asks whether the service can withstand disruption and recover. The AI Act asks whether certain uses of AI present unacceptable or high risks and what controls should follow. A person living with the outcome needs both questions answered at once: can I depend on this institution, and what will it do with its power over me?

The overlap creates practical responsibilities that neither label can carry alone. A hospital procuring an AI system has to examine its clinical purpose, training data, human oversight, security, supplier access, failure modes, and route of appeal as one arrangement. Dividing the work between compliance teams may be administratively sensible. It should not divide the account given to the patient.

Rights need a route

A protection becomes real when a person can use it from where they stand.

After an incident, can they learn that something happened before the rumour reaches them? Can they correct a damaged record wherever it has travelled? If an AI-assisted decision affects a job, loan, benefit, or treatment, can they discover that the system was involved? Is there a person with the knowledge and authority to examine the decision again? Will the explanation describe what mattered in this case, or repeat a general account of how the product works?

The AI Act’s complaint and explanation provisions create part of this route. NIS2’s duties to report significant incidents and, where appropriate, inform affected service recipients create another. Existing data-protection, consumer, labour, equality, liability, and sectoral laws remain essential because neither NIS2 nor the AI Act contains the whole answer.

That legal layering can be protective. It can also send an injured person from one authority to another, each responsible for a different slice of the event. A secure complaint portal is little comfort when no institution accepts ownership of the problem described inside it.

Good implementation therefore requires more than agencies exchanging information. Someone must be able to assemble the fragments around the person: the cyber incident, the automated judgement, the discriminatory effect, the contractual failure, and the practical repair. The law may distribute duties. The institution facing the public should not distribute the human being.

After the audit

NIS2 and the AI Act do protect people. They also protect markets, institutional continuity, trustworthy products, administrative order, and public confidence. Those aims often support human life. They can also compete with it.

A regulator may value harmonised rules because they make trade and supervision possible across borders. A company may value a certification because it reduces uncertainty and proves diligence. A public authority may value continuity because losing a service would produce chaos. None of these motives is shameful. Trouble comes when the stable market, the compliant organisation, or the secure system is accepted as proof that the person has been protected.

Return to the patient whose appointment disappeared. NIS2 can make the hospital less likely to fail, require preparation for the day it does, and bring a serious incident to the attention of authorities. The AI Act may matter if a covered system placed the patient in the wrong queue or shaped a consequential clinical decision. Other laws may govern the health record, discrimination, professional duty, or compensation.

The patient needs the institution to do something simpler and harder. Find her. Tell her what is known. Restore the treatment. Correct the record. Learn which safeguards failed. Give her somewhere to return if the correction does not hold.

Law cannot promise that no system will fail or that no automated decision will be unjust. It can decide who must prepare, who must explain, who may challenge, and who remains responsible when technical categories no longer describe the whole event.

Europe’s digital laws will be judged through those routes back to a person. A protected network matters. A lawful model matters. Their purpose becomes clear only when the human being at the end of the system can still be seen, heard, and helped.