{"id":50864,"date":"2024-03-05T12:35:21","date_gmt":"2024-03-05T11:35:21","guid":{"rendered":"https:\/\/www.accompio.com\/blog\/testen-von-ki-risikofaktoren\/"},"modified":"2026-03-10T07:17:55","modified_gmt":"2026-03-10T06:17:55","slug":"testing-of-ai-risk-factors","status":"publish","type":"blog","link":"https:\/\/www.accompio.com\/en\/blog\/testen-von-ki-risikofaktoren\/","title":{"rendered":"Testing AI: risk factors"},"content":{"rendered":"

In today's connected world, AI translators such as DeepL and Google Translate are indispensable. In the previous chapter, I put these translators through their paces<\/a>. But as with all technologies, the use of AI translators also carries certain risks. In this chapter, I will conclude my blog series and take a closer look at these risk factors and how they can be minimised.\u00a0<\/p>\n\n\n\n

When language changes\u00a0<\/h3>\n\n\n\n

Language is constantly changing. Not only the style but also the rules of the German language have changed considerably in recent years. New words and phrases have been introduced (Youth Word of the Year 2023: \u201eGoofy\u201c), while others have become outdated and less common. The influences of technology, social media, and cultural exchange have led to a rapid evolution of language <\/strong>In addition, grammar and spelling have also evolved to reflect changes in society and communication.\u00a0<\/p>\n\n\n\n

so that these trends and developments can also be reflected in the translations, the AI translator always kept up to date<\/strong> werden, especially the training data. The system learns from the training data to handle unknown texts and interpret them in a way that takes into account both the context and the nuances of the original text. If this data is not up-to-date, the system could Providing translations that are no longer current<\/strong>.\u00a0\u00a0<\/p>\n\n\n\n

Dealing with user feedback correctly\u00a0<\/h3>\n\n\n\n

So-called feedback loops allow systems to learn from mistakes and improve their own performance over time. For instance, if an AI translator mistranslates a piece of text, a user of the system can flag that translation as \u201ebad\u201c and suggest an alternative \u201ecorrect\u201c translation. This can then allow the system to use the suggested \u201ecorrect\u201c translation to adapt and improve its internal models. This process of continuous improvement<\/strong> allows the AI translator to autonomously refine its capabilities over time and become more accurate. However, it's important to note that the quality of the feedback \u2013 that is, whether a translation is marked as good or bad \u2013 and the accuracy of the suggested \u201ecorrect\u201c translation have a significant impact on this process. Incorrect or misleading feedback, whether intentional or unintentional, could lead the system to Incorrect translations interpreted as correct<\/strong>, these errors are repeated in future translations or even offensive content is embedded in the system.\u00a0<\/p>\n\n\n\n

So that this doesn't happen, this feedback should be carefully reviewed and moderated. This could be a combination of automated systems and human reviewers. In this way, the risk can be minimised that incorrect or misleading feedback will affect the performance or accuracy of the AI translator.\u00a0<\/p>\n\n\n\n

Results must be checked\u00a0<\/h3>\n\n\n\n

In every AI system, so-called \u201eoverfitting\u201c can occur. Overfitting<\/strong> occurs when a model is so adapted to the training data that it loses flexibility and performs poorly on new, unknown data. As a result, the AI system does not learn to recognise the underlying patterns. This leads to inaccurate predictions and poor overall performance of the model.\u00a0<\/p>\n\n\n\n

To ensure this behaviour is detected in a system test, a separate, independent test dataset should be used that has not been used for training \u2013 or, more accurately, cannot be used. In previous chapter<\/a> I have already drawn attention to this issue.\u00a0<\/p>\n\n\n\n

Attack on personal data\u00a0<\/h3>\n\n\n\n

As previously mentioned, training data is required for an AI system to function at all. However, one should ensure that this data No personal or private information<\/strong> include, such as names, addresses, phone numbers or other identifying features. If this data is still used for the training of an AI system, there is a risk that the privacy of the individuals concerned will be violated. This is because an attacker could potentially gain access to this data<\/a> and they misuse for undesirable purposes<\/strong>.\u00a0<\/p>\n\n\n\n

To prevent targeted advertising, identity theft, or even blackmail through the use of an AI system, all those connected to the AI Data anonymised<\/strong> including, in particular, training data. Data cleansing should also be part of the feedback loop. Otherwise, there is also the risk that entered information will be used unfiltered.\u00a0<\/p>\n\n\n\n

Further risks\u00a0<\/h3>\n\n\n\n

The aspects mentioned so far represent only a small selection of the risks. Further risks include:\u00a0<\/p>\n\n\n\n