In the sixth instalment of the series ‘Switzerland as a Hub for Innovation’, Johan Poccard and Dr Iason Kastanis from the Centre Suisse d’Electronique et de Microtechnique (CSEM) explain the conditions under which machine learning can achieve superhuman performance and why high accuracy alone does not constitute a reliable AI system.
Using practical examples from industrial image processing, robotics and laboratory automation, the authors demonstrate how companies can realistically evaluate AI systems. They explain the importance of suitable benchmarks, transparent model decisions and high-quality training data. The limitations of large language models and the need for human oversight are also discussed.
This article in KMU-Magazin forms part of the series ‘Switzerland as a Hub for Innovation’, which emerged as a follow-up to the SATW publication ‘AI Orientation: Challenges and Opportunities for Swiss SMEs’.
The content was developed in collaboration with stakeholders from the SAIROP (Swiss AI Research Overview Platform) network and the SATW’s network of experts. SAIROP promotes collaboration between research, industry and government and supports knowledge transfer within the Swiss AI ecosystem.
Yes, when it comes to clearly defined and measurable tasks, AI systems can outperform humans. For example, large volumes of images can be analysed more quickly and consistently than humans could. However, this does not imply general superiority. Performance depends on the specific task, the training data and the operating conditions.
Benchmarks are standardised testing procedures that enable the performance of different AI models to be compared. They specify which data, evaluation methods and metrics are to be used. It is crucial that the tests replicate real-world conditions as closely as possible, as good laboratory results do not automatically guarantee reliable performance in everyday use.
Explainability helps experts understand what information an AI model uses to make its predictions. This makes it possible to identify flawed assumptions, biases or unsuitable training data. This is particularly important in safety-critical applications, in order to assess reliability and strengthen confidence in the technology.
Large language models can assist with complex tasks, but they sometimes generate convincingly worded yet factually incorrect answers. For this reason, they should not be used as autonomous experts, particularly in safety-critical or high-stakes applications. Human oversight and clearly defined options for action help to mitigate risks.
Most AI systems used in industry do not continue to learn on an ongoing basis. Once training is complete, their model parameters generally remain unchanged. To make improvements, new data must be collected, processed and used for retraining. The updated model must then be thoroughly tested under realistic conditions before it is put into use.
| Role | Title + Name |
|---|---|
| Text by | Iason Kastanis, Johan Poccard |
