At the touch of a button, the doors close and the cabin starts to move. Hardly anyone stops to wonder what standards the brakes and doors are tested against, or how the sensors, control system and display work together reliably. Trust is built almost as an afterthought.
When artificial intelligence makes a shortlist in the recruitment process, detects suspicious payments or analyses medical images, questions immediately arise: How reliable is the result? Does the system work equally well for different groups of people? And who bears responsibility if it acts unethically or makes a mistake?
The short answer is: good intentions are not enough. For trust to be more than just a promise, we need shared and verifiable rules.
Standards are rarely spectacular. That is precisely where their strength lies. They do not need to stand out to be effective. Working behind the scenes, they create the conditions necessary for technology to function reliably. For example, they specify how something should be described, measured, tested or interconnected. This means that not everyone involved has to renegotiate every technical detail time and time again.
An emergency sign should be understood regardless of the building. A sheet of paper should fit into different printers. Components from different manufacturers should be compatible. Measurements should be comparable. Technical standards often underpin such things we take for granted.
Standards reduce complexity and provide guidance. Thanks to them, users do not need to worry about how each individual component works. Businesses benefit from consistent terminology, standardised testing procedures and interoperable interfaces. Regulatory bodies, testing organisations and business partners can refer to a common framework. This creates a kind of technical shorthand: anyone who understands and applies it can rely on established knowledge.
Standards are not laws. Their application is, in principle, voluntary. In practice, however, they are of great importance, for example when contracts, procurement processes, industry requirements or legal regulations refer to them. In such cases, a standard applied on a voluntary basis can help to demonstrate that a product, process or service meets specific requirements.
Common technical rules are also indispensable in the digital world. They ensure that data formats, interfaces, cloud services and electronic signatures function across different systems. AI systems also require such rules. However, the requirements for them are less straightforward to define and verify than those for technical components.
Traditional technical properties can often be measured unambiguously: a component either withstands a specific load or it does not. AI systems, on the other hand, frequently produce probability-based results. Their performance depends on data, the area of application and the conditions. A model may perform well in tests but fail in everyday use. It may achieve a high hit rate yet still deliver poorer results for certain groups.
Furthermore, errors are not always visible. A faulty component can often be detected with the naked eye. A factually incorrect AI recommendation, on the other hand, may appear linguistically correct and convincing. Standards must therefore take into account not only the end result but the entire life cycle of an AI system, from the data foundation and development right through to monitoring during operation. This involves, amongst other things, robustness, transparency, security and reliability.
Terms such as ‘safe’, ‘fair’, ‘transparent’ or ‘trustworthy’ sound convincing, but are too vague for practical application. What does ‘reliable’ mean when a system is intended to detect diseases? Is a 95 per cent accuracy rate sufficient? Has the system also been tested against rare and unexpected scenarios?
This is precisely where standards can make a difference. They translate abstract expectations into concrete questions, procedures and evidence. Depending on the field of application, they can, for example, specify:
Standards do not answer every ethical or societal question, nor do they automatically make AI safe or fair. However, they create a common foundation on which such characteristics can be described, assessed and improved. Without a common yardstick, every company could define for itself what ‘trustworthy’ means. Comparisons would be difficult, assessments inconsistent and promises hardly credible. Standards turn well-meaning principles into verifiable practice.
Artificial intelligence is increasingly being regulated. With the EU AI Act, the European Union has created a risk-based regulatory framework. Depending on the use case and risk category, different levels of stringency apply; certain AI practices are outright prohibited. Switzerland, too, is working on a consultation draft for AI regulation.
A law may, for example, require transparency, risk management, data quality or human oversight. However, it cannot specify in detail for every technology and every use case how these requirements are to be implemented. Yet companies need specific answers to concrete questions: Which processes are appropriate? What needs to be documented? How can compliance be demonstrated?
This is where technical standards come into play. They translate general requirements into specific technical and organisational guidelines and can thus provide a recognised pathway from the legal objectives to practical implementation.
In the European system, so-called harmonised standards are developed for this purpose – that is, standards drawn up by European standardisation organisations on behalf of the European Commission and intended to specify the requirements of an EU legal act. In the case of artificial intelligence, these are being developed to support the requirements of the EU AI Act. Once the harmonised standards are published in the Official Journal of the European Union, they can establish a presumption of conformity for the requirements they cover. Their application remains voluntary, but can make it considerably easier for companies to demonstrate that they meet certain legal requirements.
In the context of AI and regulation, standards are also gaining economic significance. They can help shape access to markets.
For Swiss companies wishing to offer AI solutions in Europe, the applicable requirements are crucial. Those who are aware of the relevant standards at an early stage can align their products, processes and documentation accordingly. Those who only discover them shortly before market entry may have to make costly adjustments.
In this context, standards can fulfil a function similar to that of a technical passport. They do not automatically open up the market, but they do make it easier to demonstrate compliance with the requirements and create a common language between development, compliance, testing bodies and customers.
Small and medium-sized enterprises also benefit from this. As they often do not have their own legal, standardisation and compliance departments, they are particularly reliant on guidance. Common standards help them to address requirements systematically.
Standards are developed through an interdisciplinary process. Experts from industry, academia, government and other interested parties work together to draw them up. Those who define how risks are assessed, data documented or biases measured also influence which solutions will later be regarded as reliable and marketable.
Intensive work is currently underway on AI standards at both global and European levels. The Joint Technical Committee 21 Artificial Intelligence(JTC 21), supported by CEN and CENELEC, is developing European AI standards, including the harmonised standards for implementing the EU AI Act. Switzerland is participating in this work through the national subcommittee UK 42 Artificial Intelligence of the Swiss Standards Association (SNV). The Swiss Academy of Engineering Sciences (SATW) is actively involved in UK 42 and, as co-chair of UK 42, plays a leading role in coordinating experts from industry, academia and public authorities.
From 6 to 9 October 2026, JTC 21 will convene in Winterthur for its second plenary meeting of the year. Several key standards are due to be finalised by the end of 2026. The work is therefore at a crucial stage and is the focus of significant international attention.
The fact that experts from numerous countries are currently working in Switzerland on the foundations of future AI practice makes the work on AI standards tangible. At the same time, it demonstrates how important diverse experiences and perspectives are.
Artificial intelligence is developing rapidly. This makes it all the more tempting to view standards as something slow, technical or an afterthought. Yet without common terminology, testing procedures and evidence, it remains unclear how responsible AI can be identified at all.
Standards do not make decisions for companies, public authorities or society. Nor do they resolve every conflict of interest. But they create a common basis on which quality, safety and responsibility can be discussed and assessed.
When it comes to lifts, this trust has long been part of everyday life. With artificial intelligence, it is only just beginning to emerge. The crucial question is therefore not only what AI should be capable of in the future – but also what rules it must follow in doing so.
| Role | Title + Name |
|---|---|
| Text by | Sandro Brawand |
| Editorial staff | Sereina Schär |
